DEV Community: Shawn

FutureX · Physical AI Daily — Issue 90 (08/16)

Shawn — Sat, 15 Aug 2026 14:50:32 +0000

Today's Highlights

· Waymo secures operating approval in 18 California counties, tripling service area across two major metros

· Tesla applies for 5,000 Robotaxi permits in Nevada, regulator approves only 10

· Unitree's bipedal humanoid robots reach cumulative production of roughly 18,000 units, up from just 6,500 for all of last year

· Tsinghua × Berkeley ODEWorld: real-robot success rate rises from 55% to 80%

· China's embodied intelligence funding tops 93.5 billion yuan in H1, up 5x year-on-year

· Tactile sensing sector sees 8 funding rounds exceeding 2.5 billion yuan since July, producing two unicorns valued over 10 billion yuan

· Shangwei New Materials posts a 167 million yuan loss for the half, with 164 million yuan poured into consumer embodied intelligence

I. Research Progress

ODEWorld: giving "physical time" back to world models, so robots no longer have to guess the environment frame by frame · world-model

Mainstream video-based world models are built on discrete next-frame prediction, where errors accumulate frame by frame over long-horizon physical tasks and easily drift from physical laws. ODEWorld, proposed by Tsinghua University's Institute for AI Industry Research (AIR) and UC Berkeley's BAIR lab, asks a different question — instead of predicting "what the next frame is," it learns which direction the world is currently changing in and at what continuous speed. The approach uses DINOv2 to extract features, compresses dynamic information into a compact representation space, then learns a continuous velocity field via Physical Time Flow (PT-Flow); at inference time an ordinary differential equation (ODE) solver integrates this field, allowing intermediate states at arbitrary time resolution — even inferring backward into the past. Using the predicted continuous intermediate states to guide robots, average success rate across four real-robot tasks rose from 55% to 80%.

Liu Dongxiu (co-first author) et al. (Tsinghua University AIR × UC Berkeley BAIR) · arXiv 2607.27924 source · Analysis: DeepTech source

FlashDrive: pushing a 10B-parameter driving VLA from 1.4 Hz to 6.6 Hz · autonomy

The compute wall in end-to-end VLA driving isn't a single bottleneck but a four-stage cascade — vision encoding wastes work on overlapping frames, the language model's prefill recomputes context that already existed a moment ago, inference tokens are serially generated despite being low-entropy, and flow-matching denoising applies uniform compute to a non-uniform velocity field; fixing just one stage is effectively fixing none. FlashDrive addresses all four simultaneously: cross-frame reuse via streaming KV-cache, speculative decoding with a non-autoregressive diffusion drafter, adaptive step caching structured around the velocity field's "steep at both ends, flat in the middle" shape, plus CUDA Graph compilation and operator fusion. Applied to a W4A8-quantized Alpamayo 1.5-10B, end-to-end latency drops from 717ms to 151ms (4.7×), with accuracy essentially unchanged (minADE6@6.4s shifts by only 0.08m).

Zekai Li et al. · arXiv 2608.12932 source

SimWAM: use video generation during training, discard the video branch entirely at deployment · autonomy

World Action Models (WAM) typically must generate future frames at inference time, a cost too high for production vehicles. SimWAM instead treats video generation purely as a training signal — a pretrained video expert and a lightweight action expert are jointly trained via joint flow matching, with an isolated attention mask ensuring action prediction doesn't depend on future frames; after training, the video branch can simply be discarded. The remaining action model retains a prior over traffic dynamics, requiring neither an auxiliary motion module nor frame generation at deployment — exactly the missing piece for production vehicles. According to the write-up, the work is open-sourced jointly by Huazhong University of Science and Technology and Dongfeng.

Huazhong University of Science and Technology · Dongfeng · arXiv 2608.07468 source · Analysis: CSDN source

Spatial Memory Agent: evolving a VLM's spatial reasoning without updating a single parameter · perception

The two previous routes to improving VLM spatial intelligence — post-training fine-tuning, or bolting on depth estimation/3D reconstruction tools at inference — both carry costs. SMA takes a third path: it queries a frozen model within a verifiable spatial environment, uses predictions and rewards for verifier-guided reflection, distills the experience into reusable "lessons" each tagged with a Transfer Reliability Score (TRS), and at deployment only retrieves — never updates weights. Across 5 spatial benchmarks × 4 base models, SMA achieves the top macro-average within every base-model group, taking best accuracy on most of the 20 evaluations (HF↑30).

Haokai Zhang et al. · arXiv 2608.12743 source

Other papers today: Enhancing Virtual Agents through SLMs and Edge-Computing — an exploratory evaluation of moving the "thinking" and "memory" cognitive components of virtual embodied agents to edge-side small language models (arXiv 2608.13420 source).

II. Funding & Deals

NavVis (Munich, Germany) | Series D | $85 million · adjacent ⚠️ Company-reported

Led by The Jordan Company, with participation from Yttrium, KOZO KEIKAKU, and Cipio Partners. NavVis builds mobile reality-capture scanning hardware plus its IVION cloud platform, serving over 1,500 customers including BMW, Siemens, and ExxonMobil. The real story isn't customer count but rescan frequency: the company self-reports over 1 billion square meters scanned in 2025 and over 2 billion square meters cumulative by year-end, with its website now listing over 2.5 billion square meters cumulative (all company-reported usage figures, not audited revenue). This round bets on a business transformation — from selling instruments that generate point clouds, to owning a spatial-data layer repeatedly queried by robots and AI systems. For embodied systems, a digital twin that's out of sync with reality even six months later is worthless; whether "rescan–registration–change management" can become routine is the dividing line. The sector is already consolidating: CoStar completed its roughly $1.6 billion acquisition of Matterport in 2025, and AMETEK acquired FARO the same year. $85 million isn't acquisition-scale capital, but it's enough to keep NavVis expanding independently.Source: Geoawesome source

Neros Technologies (USA) | Series C | $250 million | Post-money valuation $2.5 billion · adjacent ⚠️ Plan-stage figures

Co-led by Sequoia Capital and the American Strategic Technology Fund, with participation from Interlagos, Valor Equity Partners, Allen & Company, Thiel Capital, and Spark Capital, tripling its valuation from the previous round. The funding goes toward two new product lines: Archer AI, an FPV platform adding terminal guidance and GPS-denied position-holding autonomy; and Bandit, a counter-drone aircraft for intercepting Class 2/3 drone threats. The company says both will enter battlefield deployment by year-end, with plans to reach 1 million units annually by 2028.Source: The Robot Report source

QianXing Innovation (Shenzhen) | Series C+ | Amount undisclosed · adjacent

Investor is Zhejiang University Qizhen, announced August 13. Founded in 2016, the company makes consumer-grade portable underwater filming and exploration robots for entertainment, fishing, and diving markets, while also serving professional use cases like underwater search and rescue, aquaculture, and underwater photography. This round's funds go toward core technology R&D, product iteration, and market expansion.Source: iyiou source

Vexev | New round | $6 million · adjacent ⚠️ Aggregator-reported

A medical robotics company; this round raised $6 million, with no disclosure of investors or round-stage details.Source: Medical Device Industry Daily source

III. Commercialization & Deployment

Waymo approved for 18 California counties, more than tripling service area in Los Angeles and the Bay Area · autonomy

The California Public Utilities Commission (CPUC) approved Waymo to operate in 18 counties, including San Diego and Sacramento, plus parts of Marin, Napa, Orange, Riverside, and other counties; the approval stems from the company's application filed in January this year. The new service area extends from Santa Clarita and Thousand Oaks all the way to San Diego in Southern California, and from Sacramento to San Jose in Northern California — more than tripling the service footprint across the Los Angeles and San Francisco metros. Waymo spokesperson Sandy Karp said the expansion will roll out gradually, guided by the company's safety framework, with the immediate priority being fully driverless rides in San Diego later this summer and continued technical validation in Sacramento. The expansion follows the launch of the new Ojai vehicle model, which is larger and cheaper to manufacture than the previously retrofitted Jaguar I-Pace. Waymo now operates in more than 10 cities, with cumulative fully autonomous mileage exceeding 220 million miles.Source: Los Angeles Times source

Tesla sought 5,000 Robotaxi permits; Nevada approved only 10 · autonomy

A permit issued by the Nevada Transportation Authority on July 27 shows Tesla applied to deploy up to 5,000 Robotaxis in Clark County, but was approved for only 10, with a string of restrictions attached: a top speed of 45 mph, operation confined to the Las Vegas Strip corridor, no pickups at Harry Reid International Airport, vehicles must be clearly labeled "robotaxi," and every ride must inform passengers it is a driverless service. For comparison, Amazon-owned Zoox already operates 65 vehicles in Las Vegas with room to expand to 100. The two-order-of-magnitude gap between the number requested and the number approved is itself a public price the regulator has set on the company's fleet-scale ambitions.Source: 8 News Now source

Unitree's bipedal humanoid robots reach roughly 18,000 units in cumulative production · humanoid

Unitree Robotics' official WeChat account disclosed on August 14 that its bipedal humanoid robots have reached roughly 18,000 units in cumulative production (excluding other humanoid or wheeled-chassis humanoid products). The ramp curve is steep: full-year 2025 production was 6,500 units, and as recently as June this year the flagship G1 model had just announced surpassing 11,000 units cumulative — two months later, the overall figure has reached 18,000, marking entry into a stable, scaled mass-production cycle. The shipment-basis rankings were reported earlier: in H1, Unitree shipped roughly 5,900 units, up 170% year-on-year, for a 31% share, trailing Zhiyuan Robotics (Chinese humanoid startup) at 8,400 units (44% share) — together the two account for nearly 75% of the global total.Source: 21st Century Business Herald source

China Post's Guangzhou sorting line: single-unit feed throughput rises from 300 to 1,200 packages/hour · industrial

On the mail sorting line at the Jianggao overland export center of the Guangzhou postal district hub, embodied intelligence robots handle waybill recognition, package feeding, and exception-item sorting. After more than 5 months of continuous optimization, a single robot's feed throughput rose from an initial 300 packages/hour to a peak of 1,200 packages/hour, with accuracy above 95% and 24-hour continuous operation. Wang Lei, head of the operations and maintenance center, said efficiency is expected to reach 1,400 units/hour per unit in Q4 this year and 1,600 in Q1 next year, with exploration underway to extend into loading, unloading, and warehouse distribution ⚠️ Plan-stage figures. A comparable case: at an auto plant in Beijing's Yizhuang district, robots follow the production line's tempo to practice component pick-and-place, flexible cover placement, and material box folding, achieving a 98% success rate at a single workstation.Source: CCTV News source

X Square Robot livestreams package unloading demo, Weiduwang records rate of 1,816 packages/hour · embodied ⚠️ Company-reported

Shenzhen-based embodied intelligence company X Square Robot livestreamed a logistics package-unloading demonstration simultaneously on X and YouTube on August 12: its proprietary embodied foundation model WALL-B, paired with a six-axis robotic arm, identifies, grasps, and sorts packages from a disorganized pile after truck unloading, feeding them one by one onto a conveyor for scanning and sorting — along the way also reorienting labels toward the scanner, smoothing out soft packaging, and righting tilted boxes. Package unloading is difficult because the pile's shape changes with every grasp — boxes and soft packages stack and overlap, labels face the wrong way, and the best grasp surface is often buried. Founder and CEO Wang Qian emphasized that the challenge isn't completing a single clean grasp, but continuing to make correct decisions as the pile's shape changes and recovering when errors occur. Worth noting: this was a live capability demonstration, not equivalent to a sustained production-line throughput metric.Source: Weiduwang source

Mine automation moves past the pilot stage: roughly 1,700 vehicles in operation, one remote operator can monitor 100 · autonomy ⚠️ Company-reported

According to a Reuters report on July 24, the global fleet associated with CiDi's autonomous driving system now exceeds 1,700 vehicles, deployed across roughly 30 coal mines and quarries, mostly in China; the figure disclosed by the company in March was over 1,500 as of the end of February. More notable is the human-to-machine ratio: the company says remote takeover can achieve up to 100 vehicles per operator, and states its autonomous driving equipment had cumulatively transported over 140 million tonnes as of February 2026. Reuters also reported that roughly 10% of mining trucks in China are now driverless. The company's previously reported cumulative shipment figure exceeds 1,900 units.Source: Click Oil and Gas, citing Reuters source

Brain Corp has deployed over 50,000 robots on the floors of Walmart, Sam's Club, and Target · industrial ⚠️ Company-reported

San Diego-based Brain Corp doesn't build complete robots — only the "brain" that drives them — with its AI software powering more than 50,000 deployed robots globally; the company says its sales grew 68% this year. The most recognizable form factor resembles a small ice-resurfacing machine used for floor scrubbing, but its latest software is taking on one of retail's most tedious tasks — inventory counting: a 6-foot-tall pole fitted with cameras and wheels replaces the store associate who once needed a full eight hours to scan every row and item. The company says its robots logged over 5.3 million operating hours in H1 2026. CTO John Black's assessment is worth quoting: adding a face, legs, and dozens of extra joints "only adds complexity, not value" — the best robot is one you don't notice — and he singled out robot-malfunction videos that go viral on social platforms as, from a safety-engineering standpoint, "highly suspicious."Source: Antelope Valley Press source

IV. Industry Developments

China's H1 embodied intelligence funding tops 93.5 billion yuan, up 5x year-on-year, as investment logic shifts toward delivery · adjacent ⚠️ Aggregator-reported

Total funding in China's embodied intelligence sector in H1 2026 topped 93.5 billion yuan, up 5x year-on-year. As the money has grown, so has the question investors ask: the metrics institutions now scrutinize have shifted from technology narratives to whether a company can deliver on engineering execution, achieve scaled application, and continue winning orders. Huang Jing, general manager of Topstar's robotics product line, breaks deployment into three stages — industrial scenarios first, then commercial scenarios, with home scenarios coming last. The obstacle in between is data: embodied models need to learn physical laws, but factory floor data is highly proprietary and doesn't circulate publicly. The industry's workaround is to send robots into real settings to accumulate data while working, then feed that data back to improve capability and unlock more scenarios — the so-called "data flywheel." Policy is also pushing this forward — in June this year the Ministry of Industry and Information Technology and the State-owned Assets Supervision and Administration Commission jointly launched a special initiative on real-world training for humanoid robots and embodied intelligence, requiring that by the end of 2026 over 100 high-value application scenarios be distilled and refined, driving deployment capacity at the 10,000-unit scale.Source: CCTV News source

World Labs launches Real-to-Sim-to-Real engine: generating thousands of simulation variants from a single real-robot task · world-model ⚠️ Company-reported

World Labs, founded by Fei-Fei Li, unveiled its simulation engine R2S2R (technology derived from SceniX, a startup it acquired in July), which captures robots, sensors, environments, and task demonstrations together and reconstructs them into an interactive virtual world with physically consistent behavior. The company's premise is that the bottleneck to robot deployment isn't model architecture but the sheer volume of experience needed for reliable operation — real-robot data is expensive and hard to control, while internet video can't cover the full spectrum of objects, physical conditions, and failure states. By varying parameters such as lighting, object placement, environment, and camera viewpoint, the engine generates thousands of controlled variants from a single real-robot task, trains a control model on them, and transfers it back to the real robot.Source: the-decoder source

Tactile sensing sector sees at least 8 funding rounds exceeding 2.5 billion yuan since July, producing two 10-billion-yuan unicorns · hardware ⚠️ Aggregator-reported

According to an incomplete count by Yicai reporters, the embodied intelligence tactile sensing sector has seen at least 8 funding events totaling over 2.5 billion yuan since July this year, producing two 10-billion-yuan-plus unicorns, Paxini and Yimu Technology, both of whose latest rounds raised 1 billion yuan each — now on par with humanoid-body companies. Ma Daolin, founder of Qianjue Robotics, says order volume in H1 this year is already several times all of last year's. Costs are also shifting: visuo-tactile sensors are priced at one to two thousand yuan each, with two-finger grippers requiring 4 units and dexterous hands often using 10; the flexible surface material has a limited lifespan, and full replacement further drives up costs, making modular replacement of just the surface layer the current cost-reduction path. But tactile sensing isn't plug-and-play — a paper co-authored by Fei-Fei Li this year found that the T-Rex team's approach of directly splicing tactile force signals into a pretrained VLA model actually lowered task success rate, attributed to multimodal alignment issues, frequency mismatches, and overly shallow representations. Ma Daolin's recommended sequence is to first train a strong standalone VLA capability, then systematically incorporate tactile information into training.Source: Yicai source

Shangwei New Materials posts H1 loss of 167 million yuan, pours 164 million yuan into consumer embodied intelligence · embodied

According to the half-year report disclosed on the evening of August 14, the company's H1 revenue was 803 million yuan, up 2.42% year-on-year, while net profit attributable to shareholders swung to a loss of 167 million yuan (versus a profit of 29.9004 million yuan in the same period last year); non-recurring net profit posted a loss of 169 million yuan, down 749.01% year-on-year. Revenue is up while profit turned to loss, driven by the new business: H1 R&D spending was 180 million yuan, up 738.66% year-on-year, raising R&D intensity from 2.74% of revenue in the same period last year to 22.46%, with the consumer-grade embodied intelligence robotics business accounting for 164 million yuan — over 90% of total R&D spending. The robotics business operates under the Qiyuan Robotics brand, which has launched two personal robot models this year, Qiyuan Q1 and Qiyuan T1, with experience stores opening simultaneously in Hangzhou and Wuhan on August 9. The 210 million yuan in advance payments generated by this business has not yet been recognized as revenue; the company says it remains in the R&D and application-scenario validation stage, with scaled production still some way off. As of the August 14 close, the company's share price was 167.33 yuan, with a market cap of 67.49 billion yuan.Source: China Fund News source

Zoox CEO publicly backs stricter regulation as NHTSA sends letter to autonomous vehicle developers · autonomy

Weeks after a Zoox robotaxi drove into an active fire scene being cleared in Las Vegas and the company subsequently issued a voluntary recall of its entire fleet, CEO Aicha Evans said plainly on Fox Business on Monday: "We need to be regulated." In that incident, the scene had not been cordoned off with cones, and the vehicle failed to recognize the heavy smoke present; the company subsequently pushed a software update to improve smoke recognition. This follows NHTSA Administrator Johnathan Morrison having already issued a call to action to autonomous vehicle developers, saying the agency had recorded multiple incidents of autonomous vehicles driving directly into active emergency scenes, blocking access for ambulances and firefighters, or failing to recognize basic safety indicators such as flashing lights, flares, smoke, fire, and cones. Evans said Zoox's cumulative U.S. mileage now exceeds 3 million miles, characterizing such incidents as "rare edge cases"; the company currently operates 65 vehicles in Las Vegas, with room to expand to 100, and would need Nevada Transportation Authority approval to go further.Source: Las Vegas Review-Journal source

LG partners with NVIDIA to build a bipedal humanoid, with Isaac GR00T as the "brain" and Jetson Thor as compute · humanoid ⚠️ Plan-stage figures

LG Group Chairman Koo Kwang-mo and NVIDIA CEO Jensen Huang signed a memorandum of understanding at NVIDIA's Santa Clara headquarters, expanding cooperation across three lines — humanoid robots, AI factories, and mobility — with the project timeline extending to 2028. LG will develop a bipedal humanoid reference platform based on NVIDIA's humanoid foundation model Isaac GR00T, using Jetson Thor as its "brain," with hardware supplied by LG affiliates, targeting an unveiling in Q1 next year; ahead of that, LG plans to deploy its wheeled humanoid LG CLOi on a washing machine production line at its Tennessee plant in the U.S. this year. Separately, an 80-megawatt LG AI factory in Cheonan, South Korea, is planned for completion in the first half of 2028, while on the mobility side, LG will pair NVIDIA Drive Hyperion with its in-vehicle infotainment and software capabilities, extending LG's automotive electronics business from cockpit systems into the autonomous driving domain.Source: Dong-A Ilbo source

Lumingrobot unveils Lumos MOS2: wheeled-arm heavy-payload embodied robot with 50kg dual-arm capacity · industrial ⚠️ Company-reported

On August 14, Lumingrobot unveiled its heavy-payload wheeled-arm embodied intelligence robot Lumos MOS2, featuring 50kg dual-arm payload capacity, with upgrades to hardware performance, omnidirectional mobility, multimodal perception, and overall control architecture. It's positioned as a "heavy-duty AI Worker" for industrial settings, handling high-intensity, sustained tasks. Within the company's system, MOS2 executes real-world industrial tasks while a separate system, Lumos NexCore, provides the foundation for continuous learning and skill evolution. The company's stated view of competition is that the outcome will hinge on who achieves higher real-world data-collection efficiency, lower hardware manufacturing costs, and faster industrial delivery.Source: Sina Finance source

igus's home-edition humanoid Iggy Rob starts at €55,000, deliberately skips bipedal walking · humanoid

Germany's igus unveiled the home edition of Iggy Rob, starting at €55,000, extending the platform for the first time from industrial production lines into service and home settings, with mechanics, electronics, and software all designed and manufactured in-house in Germany. It shares its platform with the industrial version launched in spring 2025, and like that version does not use bipedal walking, instead riding on the ReBelMove Pro autonomous mobile base with a maximum payload of 250 kg; the unit stands 1.30–1.70 meters tall, has 16 degrees of freedom, up to 8 hours of battery life, and navigates using lidar and 3D vision. It comes standard with two ReBel robotic arms (each with 2.5 kg payload capacity), swappable for third-party collaborative robots or humanoid hands. The platform is built on open ROS 2 control, is CE-certified, and complies with the VDA 5050 fleet operation standard. Alexander Mühlens, head of the low-cost automation business, put the reasoning plainly: wheeled mobile platforms are already well-proven in factories and buildings, easier to integrate safely, and easier to bring into CE compliance — bipedal locomotion is "neither necessary nor cost-effective" for most real-world scenarios.Source: Weiduwang source

Hardware & Supply Chain

· Horizon Starry chip: Horizon Robotics (Chinese autonomous-driving chip maker) ecosystem partner Pateo has won a design-win from a leading Chinese NEV automaker to use this cockpit-driving-integrated chip for a high-end AI cockpit, the industry's first high-end AI cockpit fully powered by a Chinese-made cockpit-driving fusion chip; the company says it can cut in-vehicle hardware footprint by roughly 50%, reduce overall vehicle hardware cost by 1,500–4,000 yuan, and compress smart-feature development-to-delivery cycles from 18 months to 8 months (⚠️ Company-reported) source

· SMIC: raised foundry pricing on AI-driven demand, pushing up upstream cost curves for robotics and automotive chips source

· MinXin Semiconductor: H1 revenue down 18.72% year-on-year, hedging cyclical swings with a full-range MEMS product lineup source

FutureX · Physical AI Daily — Issue 89 (08/15)

Shawn — Fri, 14 Aug 2026 14:51:59 +0000

Today's Highlights

· Former Qwen technical lead Lin Junyang starts a company, post-money valuation ~$2 billion

· Schaeffler completes validation of formed harmonic reducers, key-component cost down over 25%

· Tianjin's RD3 Ultra industrial humanoid rolls off the line, with existing production capacity in the hundreds

· 7 national standards for humanoid robot test methods launched for drafting

· Customs: industrial robot exports up 13.2% in the first 7 months, bionic robot exports up over 5x

· World model DreamX-Phi 1.0 takes first place in WorldArena 2.0 Track 1

I. Paper Progress

DreamX-Phi 1.0: action-conditioned video world model, tops WorldArena 2.0 Track 1 · world-model

Today's highest-trending paper in the community (HF↑77) targets an old problem with video world models: generated results look photorealistic but aren't faithful — a generation may look real while moving the wrong arm or losing track of the manipulated object. The approach injects each arm's SE(3) transform via PRoPE-style geometric encoding into attention to preserve arm identity and rigid-body motion structure, adds a lightweight depth branch to constrain scene geometry, and uses SAM3 masks paired with a frozen V-JEPA teacher to maintain object consistency throughout a grasp. The authors report the model ranked first on WorldArena 2.0 Challenge Track 1 and second on Track 2; the model and code will be open-sourced.

DreamX Team, Rui Chen et al. · arXiv 2608.13489 source

H2R-Bench: using human first-person videos to "manufacture" robot data — eleven models collectively fail · benchmark

"Replacing costly real-robot data collection with vast amounts of human first-person video" is a popular current approach to embodied data — this paper gives it its first systematic checkup: models rewrite human-hand demonstration videos into robot manipulation videos for a specified embodiment, scored across five dimensions — goal-state completion, action-event completion, functional-contact transfer, embodiment correctness, and image quality. Evaluating 11 leading video generation models across six manipulation task categories and two robot embodiments, the conclusion is that even top-performing models frequently fail on embodiment consistency and functional interaction — the cross-embodiment-transfer shortcut isn't viable yet.

Dingyi Rong et al. · arXiv 2608.13049 source · HF↑6

Temporal GRPO: fixing the "one wrong step, whole trajectory penalized" credit mismatch in VLA reinforcement learning · vla

Current GRPO-style VLA post-training spreads a trajectory's advantage value evenly across every action, so a rollout that gets every early stage right but fails at the end ends up penalizing the correct early actions too — what the authors call trajectory-level credit aliasing. The fix constructs detectable task stages, aligns each rollout to a stage interval, and only compares rollouts that enter the same stage against each other. On RoboTwin 2.0, both success rate and sample efficiency improve simultaneously; a controlled update on LIBERO-Long shows the improvement is concentrated at the stage where outcomes start to diverge.

Yao Zhou et al. · arXiv 2608.13026 source

VLAs hide a "progress bar" internally — a single linear probe can read it out · vla

VLAs are being deployed as general-purpose manipulation policies, yet the field has almost no tools to know what they represent internally or whether they're behaving normally at runtime. Using mechanistic-interpretability methods to probe the residual stream of π₀.₅, the authors find that "task progress" (normalized time remaining in the trajectory) is linearly readable, and that this signal already exists in the PaliGemma pretrained backbone before it ever touches robot data. A single linear probe generalizes to unseen tasks, though it isn't sufficient to steer the policy in return. Used as a label-free OOD detector to catch stuck tasks, it performs on par with SOTA methods — a lightweight path toward online monitoring.

Atiksh Bhardwaj et al. (Cornell) · arXiv 2608.13474 source

UniTexture: a single textured 3D object drops VLA success rate from 90% to 48% · vla

Prior adversarial attacks on robot policies were largely "one optimization per task"; this paper makes the attack generalize across tasks: gradients from the policy's action output are backpropagated through a differentiable renderer to the object surface's texture parameters, jointly optimizing a single shared texture across the distribution of tasks, instructions, states, and viewpoints. On OpenVLA and π₀.₅, average task success rate drops from a normal-condition 90.0% to 48.4%, and the texture transfers across tasks and models without re-optimization. As VLAs increasingly control physical bodies, this kind of general vulnerability — "place a single prop to induce unsafe actions" — deserves attention early.

Yukun Dai et al. · arXiv 2608.13453 source

GeoUP: porting a reconstruction model's geometric backbone into autonomous driving — SOTA across detection, occupancy, and depth · perception

Vision-only autonomous-driving perception commonly uses a backbone pretrained for semantic recognition, then patches in 3D geometry via downstream modules, making it hard for the shared representation to preserve metric scale and consistent scene structure. GeoUP adapts VGGT's reconstruction-oriented latent space to calibrated, streaming multi-camera scenes, splits cross-image interaction into self-attention, temporal, and view-wise pathways, and injects calibration-aware raymap encoding to supply metric scale; the same latent representation is decoded separately into depth, 3D detection, and semantic occupancy. It achieves SOTA on all three tasks across five datasets — nuScenes, Argoverse 2, Waymo, KITTI, and DDAD.

Longfei Xu et al. · arXiv 2608.13147 source

S2-HWM: 98.7% success rate on long-horizon surgical robot manipulation · world-model

Long-horizon surgical manipulation has extremely sparse rewards, and meaningful interaction changes happen at irregular points in time, yet existing world-model agents mostly "imagine" at primitive-step resolution, implicitly flattening out task progress. The authors have the model learn sparse event evidence from primitive-level latent trajectories, using an event-level manager to coordinate a primitive-level executor. This reaches a 98.7% success rate, 22.7 percentage points above a flat GAS DreamerV3 baseline.

Shuzhe Zhang et al. · arXiv 2608.13103 source

HumanoidVLN: filling the physics-simulation benchmark gap for humanoid vision-language navigation · benchmark

Existing VLN benchmarks mostly assume wheeled agents, sidestepping the physical constraints of bipedal walking, morphological differences across humanoid embodiments, and the first-person camera shake that walking itself causes. This physics-simulated benchmark covers multiple humanoid embodiments and is already compatible with leading methods including NaVILA, DualVLN, StreamVLN, and JanusVLN, offering a unified way to test "does it still walk with a different embodiment."

Quan-Dung Pham et al. · arXiv 2608.12860 source

Other papers today: FIRE-VLA converts unsolved failure cases from autonomous-driving VLA into privileged supervision, easing the no-signal problem GRPO faces when an entire group performs poorly (arXiv 2608.13395 source); Deliberate Practice computes a provably optimal allocation for skill learning under a limited practice budget (arXiv 2608.13415 source); Seeker learns region-of-interest attention purely from action supervision, with no gaze-point or object-category labels needed (arXiv 2608.13422 source); ContactGuard uses an action-conditioned latent world model to predict failure and abort before contact occurs (arXiv 2608.13438 source); EgoPHI is the first to jointly estimate dense contact maps and 3D force distribution on hand and object meshes from monocular RGB alone (arXiv 2608.13014 source); HumanTracker trains a human-preference-aligned humanoid motion-tracking metric from 12K motion pairs, exposing foot-slip and contact-timing errors that kinematic metrics miss (arXiv 2608.13555 source); WMRL replaces real-environment execution with a world model, speeding up reinforcement-learning training for autonomous research agents by 3-4x (arXiv 2608.12564 source); RoboSynChallenge is a competition-format unified evaluation of the generalization of manipulation policies trained on synthetic data (arXiv 2608.12416 source); FUSE lets agents actively decide "where to look" before judging affordance when functional cues are occluded (arXiv 2608.12683 source).

Open Source · Tools · Benchmarks

· Tier IV: the Japanese autonomous-driving company plans to develop its own automotive AI chip in-house and open-source the chip design, continuing its Autoware open-source approach source

· Liquid AI LFM2.5-VL-3B: a 3B-parameter on-device vision-language model built for reading screens, object localization, and on-device tool calling source

II. Funding & Deals

Pragmatik Labs (语用科技) | Angel round | ~$220 million | ~$2 billion post-money · world-model ⚠️ Reported, unconfirmed

Lin Junyang, former overall technical lead of Qwen (Alibaba's large-model line), announced his startup on X on August 12. The publicly reported investor lineup is co-led by Banyan Capital (Chinese VC firm) and HongShan (formerly Sequoia China), with backing from Tencent and the Shanghai Future Industry Fund; the funding size and valuation are drawn from public sources such as Tianyancha and people familiar with the matter, not company disclosure. The company positions itself around "world models and embodied intelligence," with two product lines — digital agents and physical agents — aimed at providing the underlying capability for next-generation agents that span both digital and physical worlds. Lin took over Qwen in late 2022, was promoted to Alibaba's P10 level last year, and left in March this year; as of the announcement the company has no commercial product or revenue, so the roughly 13.5 billion RMB-scale valuation rests almost entirely on the team and technical judgment. This is another instance — following ByteDance's Kong Tao and several core members of Tencent Robotics X — of top large-model talent migrating toward "agent + embodied."Source: Robot Insight (机器人大讲堂) source

INFIFORCE (原力无限) | Series A + A+ | ~RMB 1 billion · world-model

Led by Dunhong Capital (Chinese asset manager) and a top-tier state-owned investment platform, with follow-on from industrial investors including Zhejiang University's tech-transfer investment arm, Yandu State-owned Capital Holding, and Lishui municipal state capital, plus continued investment from existing shareholder CCV (Chinese venture firm). Funds go toward three areas: the embodied "brain" AtomBrain and causal world-model R&D, upgrading the full-stack AI infrastructure "DataGrid," and scaling delivery validation of multi-form robot bodies in real industrial settings. The company has bet on an ego-centric data approach since 2025 — using real human first-person operation to synchronously record vision, motion trajectories, spatial relationships, and environmental feedback, replacing costly real-robot data collection, to address the problem of models becoming unstable outside demo environments — and says it will launch its first embodied model built on high-quality ego data.Source: Gasgoo source

Light Origins (亮源新创) | Pre-A round | Hundreds of millions RMB · embodied

Led by Guoke Investment, with follow-on from China Merchants Capital and Xianghe Capital. This Beijing-based company focuses on foundation models for embodied intelligence, with funds going toward large-scale embodied model training, multimodal data infrastructure, full-stack software/hardware platform R&D, and team expansion. According to reports from ijiwei and iThome (Chinese tech-industry outlets), its founder is a former OpenAI algorithm specialist.Source: ijiwei source

Unitree Robotics | STAR Market IPO offline placement | Insurers allocated nearly RMB 1.03 billion · humanoid

Following the IPO price being set at RMB 150.80/share and online subscription oversubscribed by 8,288x, preliminary offline placement results show insurers as the dominant force in this subscription: 37 insurance companies and insurance asset managers were allocated about 6.8297 million shares, worth nearly RMB 1.03 billion, or 30.15% of the offline placement quantity; among the 2,617 valid quoting placement accounts under their management, various pension/annuity plan portfolios accounted for 1,635. The single largest recipient was Taikang Asset Management, whose 642 placement accounts together received 1.82 million shares, worth about RMB 274 million. The company's revenue for 2023–2025 was RMB 159 million, RMB 393 million, and RMB 1.699 billion respectively, with non-GAAP net profit attributable to parent of -RMB 18 million, RMB 78 million, and RMB 591 million. The listing date is still pending announcement from the Shanghai Stock Exchange.Source: 21jingji source

CablebotX (线控机器人, Nanjing Xiankong Technology) | Series A | Tens of millions RMB · industrial

Uses flexible cable-driven actuation to open up a new "spatial embodiment" category, targeting large-span, large-volume heavy-industry automation scenarios — areas long left unaddressed by traditional robotic arms, which are limited by reach and rigid structure.Source: DoNews source

Hengxuan Technology (恒璇科技) | Pre-A round | Tens of millions RMB · hardware

Led by Ninebot Capital and Daoyi Capital. Ninebot, itself a maker of full vehicles and balance/self-balancing scooters, moving into components investment is another example this round of industrial capital positioning upstream along the embodied-AI supply chain.Source: Sohu source

III. Commercialization & Deployment

Shiyun Technology's RD3 Ultra rolls off the line in Tianjin, industrial full-size humanoid enters local mass production · humanoid

The RD3 Ultra industrial-grade, full-function, full-size humanoid robot officially rolled off the line at the Tianjin Aerospace Digital Industrial Park, standing 174 cm tall with a standard battery life of 8 hours and a 10-second rapid battery swap, aimed at hazardous, repetitive, high-intensity specialized work. Alongside the production line sits a full-chain test center covering whole-machine calibration, walking-endurance aging, pre-shipment testing, and voice calibration: the walking-aging step requires the machine to walk continuously for over 1 hour, the line simulates composite terrain with a 5-degree slope (to be upgraded to 15 degrees), and replicates conditions like steps and walking backward up/down slopes. The company says it covers six application categories — station duty, industrial maintenance, specialized inspection, R&D training, environmental surveying, and emergency coordination — and is co-developing a health/elder-care humanoid solution with JD Robotics, with Shiyun providing the hardware body and motion control and JD providing the large model. Current production is in the hundreds of units, with the company planning to scale to a thousand-unit capacity next year. Compared with the recent wave of showroom and competition exposure, this kind of rollout — one that builds out a full testing line alongside it — comes closer to the threshold of industrial delivery.Source: Tianjin Daily source

SANY ships first batch of pure-electric unmanned mining trucks to South America, marking China's first entry into the Latin American unmanned mining-transport market · industrial

On August 12, SANY Heavy Equipment held a shipping ceremony at its Shenyang industrial park, with the first batch of SKT110Ei pure-electric unmanned mining trucks rolling off the line awaiting shipment; the delivery covers unmanned mining trucks, a supporting dispatch system, and maintenance services. The customer is moving toward automation under pressure from a long-term shortage of skilled mining-truck drivers and rising labor costs, and SANY says this is its first unmanned mining-truck project deployed in South America. As of July 2026, SANY Smart Mining says it has deployed over 300 unmanned mining trucks, with scaled commercial operations covering major open-pit mines in Xinjiang, Inner Mongolia, and Qinghai. China's unmanned mining-transport market is currently still dominated by third-party solution providers such as Trunk Technology (踏歌智行), Wisdom Motion (慧拓), and Easymile Robotics (易控智驾, unrelated to the French company EasyMile); OEMs going overseas with their own full-stack solutions is a separate path.Source: International Mining source

Kodiak receives California's first heavy-truck autonomous-driving testing permit · autonomy

The California DMV issued the permit on August 13, the first issued since the state passed its heavy-duty autonomous vehicle testing regulations on April 28, allowing Kodiak to test autonomous trucks with a safety operator on board. California's rules follow a phased path: testing with a safety operator → driverless testing → driverless deployment, and this is only the first phase. Founder and CEO Don Burnette said this lets the company "begin the first phase of expanding autonomous freight coast to coast." California is a major U.S. freight hub that has long kept heavy trucks outside its autonomous-vehicle regulations, so this permit's symbolic significance outweighs the current fleet size.Source: Truck News source

China Post's Guangzhou sorting center lifts embodied-robot parcel-feeding rate to 1,200 pieces/hour · embodied

The Guangzhou postal district center's Jianggao processing workshop handles an average of 6.5 million pieces of mail daily, with peak-day volume exceeding 10 million; in the past, peak periods relied mainly on manual sorting. A batch of embodied robots began installation this March, and after more than five months of operation and upgrades, parcel-feeding efficiency rose from 300 to a peak of 1,200 pieces per hour, reaching over 60% of manual processing efficiency. The robots use 3D vision to identify and locate parcels, paired with flexible dexterous hands adapted to common packaging materials. "60% of manual" may not sound high, but in a logistics-sorting role that runs 24 hours a day with fluctuating labor availability, what it measures is whether the robots can hold the position on a sustained basis, not peak-moment performance.Source: China Youth Online source

Cadillac's new XT5 PHEV becomes the first mass-production model equipped with the Momenta R7 world model · autonomy

SAIC-GM Cadillac announced on August 12 that the new XT5 PHEV is the first mass-production luxury hybrid SUV equipped with the Momenta R7 world model, with its driver-assistance capability usable from delivery; showroom units have started arriving at dealerships. The vehicle is equipped with a 1.5T engine (max power 115kW) and dual front/rear motors (160kW/110kW), paired with a 35.5kWh Ultium ternary lithium battery, giving a WLTC pure-electric range of 155 km. This production milestone traces back to July 8 — the day Momenta listed on the Hong Kong Stock Exchange main board, when Cadillac announced it would debut the industry's first mass-production application of the R7 world model. World models moving from papers and demos into a mass-production car you can actually buy is one of the few places this has landed so far.Source: Gasgoo source

AutoStore signs global strategic supply agreement with Amazon, but with no purchase commitment · industrial

Norwegian warehouse-automation company AutoStore announced on August 13 a global strategic supply agreement with Amazon, establishing framework terms for Amazon's worldwide procurement of its automation systems. The company emphasized that the agreement currently includes no purchase commitment, with amount, potential order volume, deployment locations, and timeline all undisclosed. The announcement was classified as inside information under EU market abuse regulations. AutoStore's system has robots walking atop dense grids of aluminum storage bins, retrieving bins and delivering them to workstations; the company says it now has roughly 2,000 systems installed across 68 countries. The agreement itself doesn't confirm a large order, but by settling the underlying terms once instead of negotiating deal by deal, it clears a channel for potentially larger-scale deployment later.Source: Robotics & Automation News source

IV. Industry Developments

Pony.ai expands partnership with Uber, plans to deploy over 2,000 robotaxis across five European cities · autonomy ⚠️ Plan stage

Pony.ai and Uber announced an expanded partnership, planning to deploy over 2,000 robotaxis across five European cities and extend into the Middle East, with specific timelines and cities not yet announced. The two companies first partnered in May 2025 focused on the Middle East, then shifted to Europe, launching commercial service in Zagreb, Croatia with local partner Verne; Uber has now allied with over 30 autonomous-driving companies.Source: TechCrunch source

LG signs MOU with Nvidia, first bipedal humanoid set to debut Q1 next year · humanoid ⚠️ Plan stage

LG Group Chairman Koo Kwang-mo and Nvidia CEO Jensen Huang signed a strategic-cooperation memorandum on August 13, covering three areas — humanoid robots, AI factories, and mobility — and showed a model of the first bipedal humanoid robot, planned to debut in the first quarter of next year. The unit is being developed on Nvidia's Isaac GR00T platform with a Jetson Thor compute module; LG is integrating components from its own subsidiaries including LG Electronics, LG Energy Solution, LG Innotek, and LG Display.Source: Chosunbiz source

Schaeffler completes validation of formed-process harmonic reducers, key-component cost down over 25%, mass production in 2027 · hardware

Harmonic reducers are core to humanoid robot joints, and actuators account for roughly half of a robot's total manufacturing cost. The current mainstream manufacturing method, precision machining, is time-consuming and capital-intensive, and is seen as the main bottleneck to scaling humanoid production over the next several years. Schaeffler has transplanted a forming process refined over years in the automotive sector, replacing material removal with high-pressure single-step forming; manufacturing time for key components drops from minutes to seconds, manufacturing cost falls by over 25%, and material consumption drops by over 75%. The company says it has completed validation testing of the formed harmonic reducer, laying the groundwork for mass production in 2027, starting in Germany before expanding to other regions; over the past decade it has already supplied more than 2 million formed harmonic reducers to major automotive markets worldwide. The cost-reduction narrative for humanoid robots has so far mostly stayed at the level of headline unit prices — once the reducer, the single largest cost item, has its manufacturing paradigm replaced, that's the sustainable part of the story.Source: Robotics & Automation News source

Customs: industrial robot exports up 13.2% in first 7 months, smart bionic robot exports up over 5x in half a year · adjacent

Wang Jun, deputy head of China's General Administration of Customs, disclosed that China's industrial robot exports grew 13.2% year-on-year in the first 7 months of this year; since a dedicated customs tariff code was introduced for smart bionic robots this January, the category's export value has grown over 5x within half a year. Over the same period, China's total goods trade reached RMB 30.13 trillion, up 17.3% year-on-year, with electromechanical product exports at RMB 11.12 trillion, up 21.2%. Data from the Ministry of Industry and Information Technology shows Chinese-developed quadruped robots hold nearly 70% of global sales share, and Chinese humanoid robot products number over 400 models, more than half the global total. At the local level, Beijing's service-robot output grew 2.3x and industrial-robot output grew 75.5% in the first half of 2026; key humanoid-robot manufacturers in the city produced nearly 6,000 units, up nearly 10x year-on-year. The dedicated tariff code itself is worth noting as a signal — only once a statistical category exists does this segment get trackable export data for the first time.Source: Sina Finance source

7 national standards for "Humanoid Robot Testing Methods" launched for drafting · adjacent

On August 13, the launch meeting for the "Humanoid Robot Testing Methods" series of national standards was held at Wuhan's Optics Valley, with 7 standards — including general principles, environmental perception, decision-making and planning, and motion control — launched for drafting simultaneously. Over 200 representatives from companies including Unitree Robotics, Xiaomi Robotics, MagicLab (Chinese humanoid-robot maker), ZTE, DEEP Robotics, Horizon Robotics, Linker Bionics (Chinese dexterous-hand maker), and Galbot (Chinese embodied-AI startup), along with universities including Wuhan University and Huazhong University of Science and Technology, attended. The Hubei Provincial Electronic Information Standardization Technical Committee's humanoid robot and embodied intelligence standards working group was also inaugurated the same day, with an initial roster of 29 member organizations. Liu Xin, secretary-general of the Humanoid Robot Sub-Technical Committee under the National Robotics Standardization Technical Committee, said testing methods specify "how to test and how to determine whether it truly meets the standard" — equivalent to a nationally-recognized exam syllabus; currently, companies' technical standards and test methods differ from one another, making scaled replication difficult. At a time when whole-machine specifications are inconsistent and success rates and battery life are each reported differently, unifying test methods is closer to the industry's real problem than yet another new machine.Source: Optical Communication Pro (光通信Pro) source

Hyundai Motor Group signs agreement with South Korea's Army and Ministry of Trade, Industry and Energy to advance defense physical AI · adjacent ⚠️ Plan stage

The three parties signed a memorandum on August 14 at the Army headquarters in Gyeryongdae, South Chungcheong Province, to evaluate applying robotics and physical AI to defense logistics. The first scenarios slated for evaluation include logistics automation, supply transport, and perimeter/security patrol; selected projects will be validated for reliability and usability at a testbed built within the Army, and joint civil-military research will take place at the "AX Hub" in Pangyo, set to be fully operational this year. The cooperation also covers hydrogen mobility and infrastructure, along with ecosystem links to the "Saemangeum AI Valley" robotics manufacturing cluster and AI data centers. The driving factor is a shrinking pool of military personnel from demographic change — also the shared backdrop behind several Japanese and South Korean automakers' recent collective pivot toward robotics.Source: Chosunbiz source

Engine AI (众擎机器人) co-founds new company, business scope includes smart unmanned aerial vehicle manufacturing · humanoid

Shenzhen-based Engine AI (humanoid robot maker) co-founded a technology company with partners, with a business scope including smart unmanned aerial vehicle manufacturing. Humanoid-robot makers branching sideways into low-altitude aerial vehicles is a recent side trend — what's shared is the underlying capability of motors, joint modules, and motion control, rather than building a new product category from scratch.Source: EEPW (电子工程专辑) source

Hardware · Supply Chain

· Micro planetary roller screws: Nuoshi Robotics (Chinese precision-actuator maker) has mass-produced planetary roller screws just 1.5mm in diameter with a 5.5mm nut diameter, C5-grade precision, ±0.01mm repeat positioning, and 50N max load; the integrated micro linear actuator outputs 15kg of thrust within a USB-drive-sized footprint. Its Suzhou plant, running 24 hours at full capacity for an annual output of 3 million units, still can't keep up with demand; the company aims to raise annual output to 10 million units by year-end. The founder says demand has extended from dexterous hands to wrists, forearms, and even neck and face, with neck degrees-of-freedom requirements rising from one or two to four to six source

· Bionic e-skin: a Chinese research team has developed bionic electronic skin that mimics human tactile perception; according to the researchers, current application directions include tactile sensing modules for humanoid dexterous hands, tactile data-collection gloves, and smart cockpits source

FutureX · Physical AI Daily — Issue 88 (08/14)

Shawn — Thu, 13 Aug 2026 14:47:41 +0000

Today's Highlights

· Waymo's Mesa factory inventory shows 953 Robotaxis, with 684 already converted and ready for deployment

· Uber partners with Hinomaru Kotsu to pilot Robotaxis in Tokyo by year-end

· Deep Sea Smart Human (Chinese deep-sea robotics startup) raises over RMB 500 million in Series A, with TotalEnergies' fund joining alongside state-owned capital

· UOTECH (Chinese servo joint maker) raises over RMB 200 million in Series B, with 1.5 million joint units in the order backlog

· NEURA acquires Bosch Rexroth's ACTIVE Shuttle autonomous transport business

· BYD's 31-DoF humanoid "Xiaodi" starts working as a greeter at dealerships

· G0.5 unifies reasoning and action in a single autoregressive flow, achieving 76.7% on real robots, surpassing π0.5

I. Research Progress

G0.5: A Single Autoregressive Flow Outputs Reasoning and Action Together · vla

The mainstream VLA recipe today is "pretrained VLM + separately trained flow-matching action expert," which degrades the VLM into a mere context encoder rather than a decision-maker. G0.5 instead has a single transformer decoder output both reasoning tokens and action tokens under one objective, scaling this approach to foundation-model size via a learnable cross-embodiment action tokenizer, a native chain-of-thought flow, and a multi-second visual memory module. Because reasoning and action share the same weights, the VLM's instruction-following ability transfers directly to physical behavior, and prompts can adjust action granularity and task duration without retraining. It beats SOTA on 7 independent benchmarks: 76.7% on real-robot fine-tuning with R1lite/R1pro (vs. 53.3% for π0.5 and 24.4% for GR00T-N1.7), 31.4% on the 50 long-horizon household tasks of the 2025 BEHAVIOR Challenge (vs. the challenge-winning 26.1%), 98.9% on LIBERO, and 93.3% on RoboTwin 2.0.

Yicheng Liu et al. · arXiv 2608.11739 source

StellaVLA: Give the Policy a "Structured Demonstration" and It Learns New Scenes at Test Time · vla

VLAs tend to break down when the scene, viewpoint, or objects change, usually requiring re-collected data and fine-tuning. StellaVLA instead retrieves just one demonstration at test time and injects it as context — and this demonstration is automatically converted by an offline pipeline into a task plan, sub-goal descriptions, and verbalized 3D motion, conveying "why to do it this way" rather than pixel trajectories, which makes it reusable across embodiments (real robots, human hands, XR demonstrations) while adding no inference-time latency since only the action expert runs at inference. It topped the August 1 VLA-Arena leaderboard with a total score of 0.63, versus 0.44 for π0.5 and 0.22 for LingBot-VLA, with a 98.8% average success rate on LIBERO.

Siyu Xu et al. · arXiv 2608.11671 source

RIFT: World-Action Models Can "Imagine" the Future Without Rollout · world-model

World-action models (WAMs) decide actions by predicting future frames, but generating video frame-by-frame at deployment drives up latency. The authors ran closed-loop interventions on 4 WAMs across 40 LIBERO tasks and found that replaying a fixed future K/V cache for Joint and Cosmos-2 barely affects execution, showing that "consuming the cache" and "producing the cache" can be decoupled. RIFT therefore uses a learnable anticipation token to construct the complete future cache in a single forward pass, matching the rollout approach's 98.8% LIBERO success rate while cutting action-chunk latency by 68.2%–89.1%.

Chushan Zhang et al. · arXiv 2608.11521 source

ForeWAM: Feeding Dynamics to the Action Head Without Decoding Future Video · world-model

Another rollout-free approach appearing the same day as RIFT: explicit future-frame WAMs require iterative denoising of video, which is expensive, while direct-policy WAMs are fast but expose no interface for predicted dynamics. ForeWAM performs only a single Video DiT prefill via Future-KV, and supervises a "dynamics register" with a frozen latent-action teacher so that the implicit future state captures object motion, contact changes, and task progress; real future frames are only needed during training, and deployment never generates future frames at all. Without pretraining on embodied robot data, it achieves 96.7%/96.9% average success on LIBERO.

Jiakai Huang et al. · arXiv 2608.11605 source

MiDAS: One Demonstration Plus Autonomous Interaction Teaches a General Policy New Tasks · vla

Fully autonomous learning is stuck on sparse rewards and weak zero-shot exploration — robots can hardly stumble into successful behaviors from scratch. MiDAS instead anchors a pretrained VLA to a target task with just one (or very few) demonstrations, then runs value-based online RL on a residual policy. On LIBERO and RoboCasa, strong performance is recovered from a single demonstration; on a dual-arm YAM real robot, a fragile policy learned from one demonstration developed new successful behaviors after roughly 6 hours of online interaction — the authors describe this as the first reliable adaptation of a robot policy from a single task demonstration.

Shreyas Kowshik, Max Simchowitz et al. · arXiv 2608.11363 source

SHAPER: Self-Upgrading "Skills + Framework" Through Evolution, Without Touching a Single Parameter · vla

HF trending +8. When fine-tuning is too costly or the robot API isn't programmable, how else can an embodied agent get stronger? SHAPER keeps model parameters frozen, using the same model as both planner and optimizer, repeatedly evolving reusable skills and a context-code harness through rollouts in the target environment. Comparisons against pure execution, supervised fine-tuning, and test-time-scaling baselines on VLABench and ESI-Bench show that "skill and framework optimization" is a viable path when training isn't available.

Peidong Wang et al. · arXiv 2608.11350 source

Handedness Bias in Bimanual Humanoids: Initial Pose Quietly Decides Success · manipulation

Aggregate success rates can mask pose-related failures. The authors quantify VLA-induced "handedness priors" across multiple policies and 17 initial configurations, finding that success rates for the same pose vary widely across policies, that certain initial arm configurations induce or suppress asymmetric hand preferences, and that the presence or absence of a wrist camera also affects the ranking of policies and their scores; expanding initial-pose coverage in the training set and applying targeted augmentation to low-scoring configurations both markedly improve robustness.

Chaeyeon Jung, Juyoun Park · arXiv 2608.11769 source

RoboSaGA: Saliency-Guided Augmentation to Withstand Shadows and Background Clutter · manipulation

Already cited 11 times. Standard random-crop and color-jitter augmentations used in visual behavior cloning fail under large visual domain shifts; RoboSaGA uses saliency-guided augmentation to maintain in-domain performance while significantly improving robustness to distractors, background changes, and lighting/shadow variation.

Zheyu Zhuang et al. · arXiv 2608.11870 source

Other papers today: SMPC demonstrations plus sparse offline-to-online RL for loco-manipulation, avoiding dense reward shaping (arXiv 2608.12063 source); HandEdit, the first unified benchmark for egocentric "human hand → robot hand" image editing (arXiv 2608.12122 source); DaViNCi, the first outdoor vision-language navigation dataset with both continuous action and dynamic elements (arXiv 2608.11901 source); Map-Det3D, single-view multi-view detection using feed-forward metric 3D reconstruction priors (arXiv 2608.12179 source); RoadWeaver, generating large-scale lane-level HD maps from scratch for autonomous-driving simulation (arXiv 2608.11580 source); TrafficDiffuser, traffic scene generation via joint diffusion of initial and goal states followed by trajectory completion (arXiv 2608.11407 source); Video2Track, converting real interaction videos into adversarial tests with tunable risk in closed venues (arXiv 2608.11592 source); linearly decoding early-warning signals of imminent failure from OpenVLA's internal activations (arXiv 2606.29699 source); HUGIN, joint vision-language planning across multiple scenarios for logistics sorting (arXiv 2608.11692 source); Thea, an embodied harness that packages robot capabilities as tools orchestrated by an agentic loop (arXiv 2608.11246 source).

II. Funding & Deals

Deep Sea Smart Human (Chinese deep-sea robotics startup) | Series A | Over RMB 500 million · industrial

New investors include Guanghe Venture Capital, Fortune Capital, GGV Capital, Yida Capital, China Insurance Investment, and CETC Investment, along with an energy technology fund managed by Cathay Capital with TotalEnergies as the funding entity; existing shareholders including Zhengxuan Investment, Yunze Capital, Yunshi Capital, and Gaojie Capital made large follow-on investments. Founded in 2023, the company builds work-class ROVs operating at depths of several thousand meters — heavy equipment essential to deep-sea oil and gas, subsea communications, offshore wind, and deep-sea mining, a market long dominated by European and American manufacturers. According to data obtained by ChinaVenture, as of this July the company had signed several hundred million RMB in new deep-sea robot equipment sales contracts, with a similarly-sized order backlog locked in for the next two months and another RMB 1 billion in potential orders under negotiation. This marks the first time four market-driven top-tier VCs have invested in deep-sea robotics simultaneously, signaling embodied AI's spillover from land to sea. Source: ChinaVenture source

UOTECH (Chinese servo joint maker) | Series B | Over RMB 200 million · hardware

12 institutions participated including Samsung Venture Investment, Guofang Venture Capital, and Jiaqiao Capital, with existing shareholders Jiaqiao Capital, Fudar Holding, and Xishan State-owned Investment continuing to add on; the company has completed three funding rounds this year alone, raising over RMB 500 million in total. This integrated servo-joint manufacturer says its joint order backlog for 2026 has already exceeded 1.5 million units, 15 times its total 2025 shipment volume; its fully automated production line built in Shanghai's Zhangjiang area earlier this year produces one harmonic joint every 60 seconds with a 98% first-pass yield, and its Wuxi base has purchased an additional 34 mu (about 5.6 acres) of land for a new automated facility. Its client list includes AgiBot (Chinese humanoid robotics company), iFlytek, FAW, Midea, and Siemens. Joints are a key bottleneck for humanoid whole-machine cost and consistency, and the jump in orders from "prototype supply" to million-unit scale is a rare hard data point of certainty in the supply chain. ⚠️ Manufacturer's figuresSource: China.com source

NOUSBOT | Series A+ | Nearly RMB 100 million · hardware

Led by Shunwei Capital with participation from Wangqian Capital, the funding will mainly go toward expanding production of miniature planetary roller screws, product iteration, and overseas expansion. Founded in July 2023, the company grew out of a family business with over 30 years of screw-manufacturing experience, and its shareholder history includes SAIC Capital, Shanghai Semiconductor Industry Investment, and Legend Capital. Founder Xu Yang believes screws will be the ultimate solution for humanoid linear joints — rotary joints are easy to control but struggle to combine high force with a compact size, and stall-induced heating remains unresolved, whereas a screw's self-locking force lets the motor operate intermittently. He also noted a shift in requirements: dexterous hands used to be "nearly unusable," frequently failing after about 200 hours, but transmission demand has now extended from hands to wrists, forearms, and even the neck and face, with neck degrees of freedom rising from one or two to four to six.

Source: 36Kr source

NEURA Robotics Acquires Bosch Rexroth's ACTIVE Shuttle Autonomous Transport Business · industrial

The deal takes effect October 1, with NEURA Mobile Robots taking over ACTIVE Shuttle's hardware, software, and ongoing customer service. This autonomous transport system is already used by numerous industrial companies for assembly-station replenishment and small-load transfers, a proven, field-tested product line, and NEURA plans to fold it into its own Physical AI ecosystem while adding AI capabilities and interfaces. Founder David Reger's stated goal is a shared technology base across mobile robots, industrial robots, and humanoids, rather than separate intelligence stacks for each robot category. The German company just closed a Series C of up to $1.4 billion in June, and this acquisition of a mature business from Bosch is one example of European robotics assets consolidating around new market entrants. Source: Tech.eu source

Anthropic | In Talks to Acquire Decart AI | ~$6 billion · world-model

Reuters, citing sources, reports that Anthropic is in talks to acquire Israeli AI startup Decart at a valuation of around $6 billion, targeting its video generation and chip optimization capabilities. Decart has been known for real-time video generation and world models; if the deal closes, it would signal that foundation-model companies are starting to treat "real-time generative worlds" as a must-have capability rather than something to source externally. ⚠️ Rumored termsSource: Reuters source

Hai Robotics | Hong Kong IPO Application Lapses · industrial

The Hong Kong listing application status for this box-based automated storage (ACR) leader has changed to lapsed; the company had previously disclosed a global ACR solutions market share of over 30%. Amid a wave of logistics robotics companies filing for listing, this is a rare process interruption. Source: FX168 Finance source

Unitree Robotics | STAR Market IPO | Share Payment Completed · humanoid

Following an online subscription that was oversubscribed 8,288 times, payment for this offering has been completed; online investors forfeited subscriptions for 8,734 shares, which will be underwritten by the lead underwriter. On the institutional placement side, 89 mutual funds collectively received allocations worth nearly RMB 1.8 billion, and 37 insurers took roughly 30% of the institutional tranche. Reports of scalpers' resale prices reaching RMB 410 trended on social media, though some experts said this reflects sentiment rather than fundamentals. Source: Sina Finance source

Wanji Yizu | New Funding Round | JD.com Enters Robot Leasing · adjacent

Following an earlier strategic round valuing it at RMB 1 billion, this robot leasing platform has secured a new round of funding, with JD.com simultaneously entering the sector. Leasing is a common way to sidestep the barrier of one-off customer purchases, and platform-level funding often runs ahead of the actual scaled delivery of the equipment itself. Source: Sina Finance source

III. Commercialization & Deployment

AgiBot's Dual-Arm Robot Live-Sorts 1,816 Items/Hour, Beating the Figure 03 Benchmark by 45% · embodied

During a continuous one-hour livestream, a dual-arm robot equipped with only a standard industrial gripper completed 1,816 valid sorts, exceeding the 1,248 items/hour benchmark set by Figure 03's earlier 200-hour continuous run by 45% in short-window throughput terms; according to public information, this solution cuts costs by 70% with over 98% sorting accuracy. Behind it is AgiBot's WALL-B, released in April, which uses a "world unified model" architecture that jointly trains vision, language, action, and physical-change prediction in a single network. It should be noted that livestream throughput is a capability demonstration rather than a mass-production-line metric. ⚠️ Manufacturer's figuresSource: iFanr source

Waymo's Mesa Factory Inventory Shows 953 Robotaxis, 684 Conversions Completed · autonomy

According to a monthly aerial survey cited by Electrek, Waymo's factory in Mesa, Arizona currently holds 953 dedicated Robotaxis, of which 684 — about 72% — have completed conversion and are deployment-ready, with 269 still awaiting conversion; the same count stood at roughly 500 in early June. Also confirmed for the first time were 12 complete vehicles based on the Hyundai Ioniq 5 platform, marking the first appearance of Waymo's second dedicated vehicle model on the completed list. The Ojai model's chassis is produced by Zeekr (Chinese EV brand under Geely) before being shipped to Mesa, where Magna installs the sixth-generation autonomous-driving system — 13 cameras, 4 lidars, and 6 radars, a 42% reduction in sensor count from the previous generation, with Waymo saying hardware cost can be kept under $20,000 per vehicle. Its weekly paid rides already exceed 500,000, with a target of 1 million within the year, while its operating fleet stood at roughly 2,500–3,000 vehicles at the start of the year — meaning the converted, completed vehicles already represent a substantial pool of pending capacity, and the next variable shifts from "can it be built" to "how fast can it be deployed." ⚠️ Estimated figuresSource: Digital Today (citing Electrek and aerial survey data) source

Uber Partners with Hinomaru Kotsu to Pilot Robotaxis in Tokyo by Year-End · autonomy

Uber announced a partnership with veteran Tokyo taxi operator Hinomaru Kotsu to deploy a Robotaxi pilot fleet, expected to launch by the end of 2026. Uber's autonomous-driving footprint in Japan had previously centered on Tokyo partnerships with Wayve and Nissan; this time a local fleet operator takes on the operational side — a typical division of labor for overseas Robotaxi deployment, with the technology provider supplying the system and the licensed fleet operator supplying capacity and compliance. Source: 36Kr source

BYD's Humanoid "Xiaodi" Enters Dealerships, 31 Degrees of Freedom · humanoid

"Xiaodi" made its debut at BYD Space experience centers in Zhengzhou, Shenzhen, and Shanghai — 1.61 meters tall, weighing 58 kg, with 31 degrees of freedom (7 in the hands alone), supporting six Chinese dialects and six foreign languages, with facial recognition, gesture recognition, and lip-reading capabilities, handling greeting, vehicle introduction, and Q&A duties. BYD says it plans to deploy two to three units per dealership. This puts a humanoid into an actual retail setting rather than a trade-show booth, though the role remains a sales-greeter type of task rather than production-line work; around the same time, Xpeng plans to bring its Iron humanoid into Chinese dealerships by Q1 2027. One practical reason automakers are choosing retail over factories for this rollout is that they already have an extremely wide dealership network they can use as a large-scale real-world test bed. Source: Interesting Engineering source

EACON Begins Unmanned Night-Shift Operations at Australia's Havana Mine · autonomy

EACON (Chinese mining autonomy company) has launched night-shift autonomous haulage operations at Australia's Havana mine. Night shifts are one of the truest tests of mine-site autonomy — poor lighting, high labor costs, and elevated accident rates mean that consistently running a night shift usually signals the system has moved past pure daytime pilot operation. Source: International Mining source

WeRide Considers Bringing Robotaxis to Australia and More Asian Markets · autonomy

The company is evaluating expansion into Australia and Asian markets, having already seen its overseas revenue growth clearly outpace its Chinese revenue growth. This remains at the market-evaluation stage, with no licensing or launch timeline yet confirmed. ⚠️ Planning stageSource: Capital Brief source

IV. Industry Developments

2nd World Humanoid Robot Games Open August 22, Adding 14 New Real-World Scenario Events Across 1,301 Matches · humanoid

This edition features 51 events across 1,301 matches, with 666 teams from 16 countries participating — a substantial expansion over the first Games — and includes night sessions for the first time. The most notable change is in the format: 14 new scenario-based events shift competition from gymnasium running-and-jumping toward "last-mile" real-world settings — pick-and-place, delivery, and human-robot collaboration in settings like hotels, retail stores, and production lines. A 5-on-5 soccer match has already kicked off. The Games are primarily a showcase, but the addition of scenario events adds a further function: comparing different manufacturers' actual task completion under a unified rule set on similar tasks — something ordinary manufacturer demos, each on its own terms, can't provide. Source: The Paper source

JD Logistics Half-Year Report: AI and Automation Scaled Across End-to-End Supply Chain, R&D Spending Up 53% · industrial

JD Logistics disclosed that AI and automation technologies have been deployed at scale across end-to-end supply-chain scenarios in the first half of the year, with group-level R&D spending up 53% year over year, positioning JoyAI externally as its "physical-world intelligence foundation." E-commerce logistics is one of the earliest sectors in China to create stable, paid deployment slots for robots, and its capital expenditure direction typically leads general embodied AI deployment by one to two steps. Source: Futu Bull source

Tokyo Electron Expands Partnership with Nvidia Toward Agentic AI and Robotics · world-model

Semiconductor equipment giant Tokyo Electron is expanding its partnership with Nvidia to cover agentic AI and robotics. Nvidia has been signing a string of deals with Japanese industrial players recently, following the same playbook as in automotive and healthcare — first embed simulation and foundation models into a partner's existing production lines, then move on to the robot itself. Source: Evertiq source

Daedong Partners with Rainbow Robotics to Develop Agricultural Humanoid "AGROID" · humanoid

South Korean farm-machinery giant Daedong and Samsung-affiliated robotics company Rainbow Robotics announced joint development of an agricultural humanoid robot called AGROID. Agriculture is one of the few settings where "labor shortage" and "unstructured work" both hold simultaneously, which is why it's often seen as a way for humanoids to sidestep the strict cycle-time demands of factories.

Source: Asia Economy source

Honda Unveils New Humanoid "Avatar" · humanoid

Honda has unveiled a new humanoid robot, Avatar, widely seen as a continuation of the ASIMO lineage. Japanese OEMs have kept a low profile through this humanoid boom, so Honda's re-emergence is itself a signal. Source: Robotics & Automation News source

UBTECH Partners with Maoming to Build Embodied AI Application Innovation Center · humanoid

The project has cleared filing, with planned investment of RMB 100 million. Local governments trading scenarios and funding for OEM deployments has become one of the most common partnership formats in humanoid robotics this year, though actual delivery volumes usually aren't clear until acceptance data comes in one to two quarters later. Source: Sina Finance source

Geek+ Launches RoboShuttle Hyper Series Autonomous Mobile Robots · industrial

A next-generation AMR product line for warehousing, emphasizing higher-throughput box handling. Source: GlobeNewswire source

Hardware & Supply Chain

· Zhenghe Industrial (Chinese chain-drive maker): This chain-drive leader's roughly RMB 700 million private placement was approved, with proceeds directed toward high-end farm-machinery alloy cutting tools and robot dexterous-hand components — another example of a traditional transmission-parts manufacturer moving into the embodied AI supply chain source

· Texas Instruments CAN XL Transceiver: A new-generation transceiver for higher-bandwidth automotive and industrial buses has launched; the bandwidth demands of multi-joint, high-frequency force control in robots are moving up the same curve source

· Wieson Technologies: First-half revenue of NT$1.567 billion, with second-half focus shifting toward AI thermal management and robotics-related business source

V. This Week's Watch

China Issues Mandatory National Standard for Autonomous Driving System Safety, First Unified Access Threshold for L3/L4 · autonomy

The "Intelligent and Connected Vehicles — Autonomous Driving System Safety Requirements" (GB 44721—2026), developed under China's Ministry of Industry and Information Technology, has been approved for release by the State Administration for Market Regulation and the Standardization Administration, and will take effect on July 1, 2027. It applies to M- and N-category vehicles equipped with L3 or L4 autonomous driving systems, but excludes automated parking systems. The key change is legal force: its predecessor, GB/T 44721—2024, was a recommended standard that companies "could" follow; as a mandatory national standard, compliance is now required — effectively setting a unified access floor for L3/L4 in China and placing safety responsibility on automakers across the full vehicle lifecycle. For the industry, this amounts to roughly an 11-month compliance countdown, since vehicle validation cycles for clauses covering failure response and minimum-risk strategies are typically measured in years. Source: Securities Times, Sina Finance

US Teamsters Sue California DMV to Overturn New Driverless Heavy Truck Rules · autonomy

California Teamsters filed suit against the California DMV in Alameda County Superior Court on August 5, with the core allegation being that the agency permitted autonomous freight operations without completing the legally required "Standardized Regulatory Impact Assessment" (SRIA) — the DMV lifted its ban on autonomous driving for vehicles over 10,001 pounds gross weight on April 28. The union says the DMV underestimated the employment impact on California's more than 200,000 commercial truck drivers, and questions whether heavy driverless trucks have undergone sufficient public-road testing, seeking to have the rule invalidated and enforcement blocked. This is the first institutional obstacle driverless trucking has hit in its US commercialization push, and it takes a procedural-legality route rather than a safety-evidence one. Source: Transport Topics, The Hill

South Korea Eases Robot-Fencing and "One Site, One License" Restrictions, Unlocking KRW 4.2 Trillion in Investment · industrial

South Korea's Minister of Economy and Finance Koo Yun-cheol announced a package of deregulation measures the government expects to unlock KRW 4.2 trillion (about $2.96 billion) in investment, focused on freeing up already-ready projects stuck in the approval process, serving semiconductors, physical AI, and AI data centers as the three priority areas. The two measures directly tied to robotics are removing the "one site, one license" restriction and lifting the industrial robot fencing requirement — the latter determines whether collaborative and mobile robots can share a workspace with people, a direct constraint on production-line deployment density. Source: The Korea Herald, Asia Economy

Goldman Sachs: AI Investment Theme Shifting to Humanoids, 76,000 Units Projected for 2027, Scale-Up Not Until 2027–2029 · humanoid

Goldman Sachs' latest research report judges that as compute-infrastructure investment continues to expand, the AI investment theme is shifting from the compute race toward end applications, with humanoid robots as the next-stage growth engine; it projects global humanoid shipments reaching 76,000 units in 2027 and 502,000 units by 2032, with large-scale deployment only gradually unfolding from 2027–2029. The report also notes current applications remain largely proof-of-concept (POC), with scenarios extending into industrial handling, logistics workflows, and structured commercial settings. This timeline judgment diverges notably from figures reported in China: China's Ministry of Industry and Information Technology's figure this month of roughly 23,000 units shipped in the first half had over 70% still in showrooms and labs — the two figures actually point to the same underlying reality, that the shipment curve has taken off while the deployed-workstation curve has not. Source: Goldman Sachs research report, via Sina Finance

Counterpoint: Commercial Cleaning Robot Revenue Exceeds $400 Million in 2025, Pudu Robotics Leads in Both Shipments and Revenue · industrial

A Counterpoint Research report shows global commercial cleaning robot shipments exceeded 50,000 units and revenue exceeded $400 million in 2025, with Pudu Robotics ranking first globally in both shipments and revenue. The report also gives two directional figures: global commercial cleaning robot penetration remains below 10%, while annual shipments could exceed 400,000 units by 2030 at a compound growth rate above 50%. While humanoids are still debating their first stable deployment slot, cleaning robots — a single-task, structured-scenario category — have already built a countable revenue base, offering a ready reference point for observing when robotics starts becoming a real business.

Source: Counterpoint Research

This Week's Supply Chain: Money Flows to Both Transmission and Tactile Sensing, Dexterous Hands Still Stuck on Standardization · hardware

Stringing together this week's scattered signals, funding is concentrated almost entirely on two end-effector segments of the humanoid. On the transmission side: UOTECH's over RMB 200 million Series B with a 1.5 million-unit joint order backlog, NOUSBOT's near-RMB 100 million Series A+ expanding miniature planetary roller screw production with a stated 10-million-unit target, and Zhenghe Industrial's roughly RMB 700 million private placement redirected toward dexterous-hand components — the scale of capacity planning has jumped from "tens of thousands of units" to "millions of units," spanning harmonic joints to lead screws. On the tactile side, Sino Silicon Century placed an order this week for 200,000 units of tactile sensors with Flysource New Materials, sufficient by its own count to equip 40,000 industrial-grade dexterous hands — flexible tactile sensing is moving from samples to bulk orders. The sharpest point of tension isn't capacity but standards: industry statistics show dexterous-hand-sector funding in the first half already exceeded RMB 25 billion, more than all of 2025, yet interfaces, degree-of-freedom configurations, and lifespan specs vary widely across manufacturers — NOUSBOT's founder noted that most dexterous hands on the market in 2025 failed frequently after about 200 hours, and by 2026 there's still a gap to industrial-grade reliability. Component makers are building capacity for a whole-machine form factor that hasn't yet converged — the most real risk exposure in this week's supply chain. ⚠️ Compiled figuresSource: 36Kr, Eastmoney, 10jqka

FutureX · Physical AI Daily — Issue 87 (08/13)

Shawn — Wed, 12 Aug 2026 14:50:09 +0000

Today's Highlights

· Nature sub-journal: a single grasping model drives 7 different robotic hands zero-shot, with 94.3%–98.0% success rates

· HKU-led RoboDojo unified benchmark: the strongest model achieves only 12.8% real-robot success, versus 100% for human experts

· WeRide's H1 revenue reached RMB 350 million, up 73% YoY, with overseas/international revenue up 154%

· Waymo has imported over 3,200 Zeekr-chassis Robotaxis, with per-vehicle cost estimated at roughly half the previous generation

· A3: North American Q2 robot orders reached 8,940 units, while automotive OEM orders fell 25% in H1

· FCC revokes drone certification for a "U.S.-branded" front company for the first time — Odyssey Robot loses approval

· Zhongke Guiji (Chinese tactile-sensor startup) places a single order for 200,000 tactile sensors, enough for 40,000 dexterous hands

· Ukraine has ordered over 3,000 ground robots cumulatively, with 640 units in July alone

I. Research Progress

One grasping model works across seven robotic hands — SpaHybGen published in Nature Machine Intelligence · manipulation

Robotic grasping has long been stuck in a dilemma: deep-learning policies are sample-efficient but locked to specific hardware, while analytical planners transfer across hands but struggle with sensing noise. SpaHybGen bridges the two — first training a neural network to infer a hardware-agnostic "spatial contact feature" representation directly from noisy depth observations, then handing it to a differentiable optimizer that solves for stable grasps on the specific hand geometry. Results from a single training run zero-shot drove seven different hands, from two-finger to five-finger grippers, with no hardware-specific retraining at all, achieving 94.3%–98.0% grasp success in semi-cluttered scenes and supporting dynamic 20 Hz grasping and multi-hand coordination in dense, cluttered environments. The authors decouple perception from action around a shared contact interface, and have open-sourced the code, model, and the generated contact-feature dataset.

Nature Machine Intelligence (2026) · Code and model open-sourced on GitHub (SpaHybGen) · Coverage: Nature source

GeniWorld: helping world models "understand" robot actions — π0 real-robot success rate rises from 40.8% to 69.0% · world-model

Expensive real-robot data and hard-to-build scenes remain long-standing bottlenecks for generalizing embodied policies. GeniWorld, jointly proposed by the Tsinghua Shenzhen International Graduate School, Tencent Robotics X, HKUST, and Shenzhen Technology University, uses URDF rendering to convert numerical action sequences into a "visualized action representation" aligned with real camera viewpoints, then concatenates it with noisy video latents and feeds them into a causal diffusion Transformer to predict the future — letting the world model be precisely controlled by actions rather than merely generating passively. On the real-robot side, using a dual-arm Xtrainer platform with only 25 demonstrations per task to train π0, adding GeniWorld synthetic data raised the overall success rate across four tasks (moving a bowl, folding a towel, etc.) from 40.8% to 69.0%, with spatial-rearrangement scenarios rising from 37.5% to 70.0%.

Tsinghua Shenzhen International Graduate School × Tencent Robotics X × HKUST × Shenzhen Technology University · Coverage: Sina Finance source

Flex-π: a frozen video VAE turns out to encode 3D point maps too, for free · world-model

Nearly all world-action models (WAMs) predict only RGB latents and train purely on pixel-level reconstruction, with no explicit supervision on the 3D geometry and object semantics that manipulation actually needs. The authors found a "free lunch": the frozen video-generation VAE used to encode RGB also encodes 3D point maps almost losslessly, with no point-map-specific training required. Building on this, the 6B-parameter model projects geometry, DINO semantics, and RGB into the same latent space within a Mixture-of-Transformers backbone, denoising them jointly with actions, while per-stream dropout lets a single checkpoint run on any subset of streams. On real dual-arm fine manipulation tasks, it outperforms the strongest baseline by 2–7x at best, both in- and out-of-distribution, and infers faster than π0.5.

Ge Yan et al. · arXiv 2608.10860 source

FACT: making failed actions usable as training targets too · world-model

arXiv has seen a cluster of world-action model papers today, and FACT targets a gap most of them skip: these models are almost always trained only on successful demonstrations, so they have no reason to learn what "bad actions lead to," making them prone to hallucinating successful futures during inference. FACT instead predicts future video and task progress conditioned on the actions actually executed, turning failed rollouts into legitimate supervision signals rather than discarded data; the progress predictor sees both successful and failed outcomes, and can score candidate actions at inference time. The authors report continued improvement as failure data is added in both simulation and real dual-arm experiments, with a significant reduction in success-biased hallucination under bad actions.

Quanquan Peng et al. (including Nicklas Hansen, Xiaolong Wang) · arXiv 2608.10232 source

SALT: action tokenizers optimize only for reconstruction, losing the semantics of verbs · vla

Action words describe not just outcomes but "how" something is done, yet VLA action representations are typically optimized in raw action space using L1/L2 reconstruction loss — numerical closeness doesn't equal linguistic meaning. The authors first show on BridgeV2 that action trajectories carry verb-grounding information beyond visual state changes, and that reconstruction-only discrete tokenization systematically erases it; SALT then adds an auxiliary objective to VQ-VAE-style tokenizers, requiring a frozen vision-language model to recover the episode's instruction from the quantized action latent. Policies trained with SALT achieve an average success rate of 71.9% on SimplerEnv, versus 42.7% for a reconstruction-only VQ-VAE tokenizer and 31.2% for FAST.

Li Wenjie et al. · arXiv 2608.10484 source

4D-WAM: driving world models shouldn't learn only 2D projections · autonomy

Existing driving world-action models are mostly trained on video, but video is only a 2D projection of the underlying 4D scene. 4D-WAM uses geometric foundation models to provide training-time supervision, modeling scene evolution in a 4D-consistent space, achieving state-of-the-art results on both the NAVSIM-v1 and NAVSIM-v2 benchmarks.

Jiacheng Fu et al. · arXiv 2608.10107 source

Hidden in Plain Sight: "invisible" attacks on VLAs using diffusion models · vla

Adversarial robustness for VLAs has so far received little systematic study, yet when these models are misled the consequences land directly in the physical world. The authors use diffusion models to construct attack samples unconstrained by perturbation norms — looking like natural scenes rather than noisy images — to probe VLA failure boundaries on real manipulation tasks. Worth watching closely as embodied models move faster into production lines.

Jiahui Han et al. · arXiv 2608.10393 source

Other papers today: JEPA-WAM does stage-level joint embedding prediction using a frozen V-JEPA2 encoder (arXiv 2608.10780 source); Surgical WAM does data-efficient surgical robot learning based on Cosmos Policy (arXiv 2608.11204 source); full-body planning for humanoid robots in confined spaces, training a residual RL policy with self-collision avoidance as reference (arXiv 2608.10220 source); single-view dual-arm coordinated dexterous grasping of large objects (arXiv 2608.10383 source); VIScore diagnoses the "planability" of latent-space world models (arXiv 2608.11174 source); HUI360, currently the largest 360° first-person dataset for human-robot interaction anticipation (arXiv 2608.11051 source); Gated VLA-Cache uses neural introspection to decide whether to reuse the KV cache (arXiv 2608.10824 source); Dreamer-SAC puts off-policy SAC directly inside a latent-space world model for autonomous driving (arXiv 2608.10386 source); DriveVLA-M0 builds a failure-case memory bank for retrieval augmentation (arXiv 2608.10413 source); AECNav does training-free zero-shot open-vocabulary object navigation via evidence gating (arXiv 2608.10817 source).

Open Source · Tools · Benchmarks

· RoboDojo: a unified manipulation benchmark led by the University of Hong Kong's MMLab together with nearly 20 universities (including UC Berkeley and Tsinghua), bringing simulation evaluation, standardized real-robot testing, and policy benchmarks into a single framework covering 42 simulation tasks, 18 real-robot tasks, and 30 representative policies, assessing generalization, memory, precision, and long-horizon execution. The first results deliver a sobering number: the best-performing model reaches 8.80% success in simulation and 12.8% on real robots, versus 76.03% and 100% for human experts respectively. In its first week, the project drew over 100,000 views on X, and its open-source resources have been downloaded over 100,000 times on Hugging Face. source

· NVIDIA JetPack 7.2.1: adds Agentic Video Skills and T3000 simulation support. source

· LeRobot Studio / ROSView: an open-source offline visualization tool for embodied data from IO-AI Tech (Chinese embodied-data infrastructure startup); the same release also included several new real-world datasets. source

II. Funding and Deals

Westlake Robotics | Series A | over RMB 100 million (RMB 500 million across four rounds in six months) · embodied

Investors include Sino-French Fund (Saif), Xiaomiao Chengcheng Capital, and Haiyuan Fund. Founded in 2024 and headquartered in Hangzhou, the company is Westlake University's first commercialization project in AI and robotics, pursuing a full-stack approach of "general-purpose brain + humanoid full-body cerebellum + proprietary humanoid platform," with plans to launch its fully self-developed humanoid, Westlake o1, in 2026. Founder Wang Donglin is a tenured professor at Westlake University and deputy director of its AI department; co-founder Zhang Yue is likewise a tenured professor there. The company says it holds nearly RMB 100 million in orders, spanning STEM education training, humanoid data collection, and power-grid inspection, with new funds focused on developing a unified full-body humanoid foundation model. Source: Shanghai Securities News · China Securities Journal (via Eastmoney) source

Demak Precision | Series C | RMB 420 million · hardware

Led by Furong Investment, with participation from Changxing Industrial Group, Nanhu Equity Fund, Yuanjing Capital, and Deqing Industrial Investment, plus a follow-on from existing investor Yunqi Capital. Founded in 2016 in Changxing, Huzhou, Zhejiang, the company's core business is precision machining of semiconductor-equipment chambers and large structural components for solar and wind power, with nearly a thousand imported high-end machine tools and over 20,000 square meters of temperature-controlled factory space. It is now applying that batch-consistency capability to humanoid robot torsos, joint housings, and limb connectors, having already established partnerships with Unitree Robotics, Zhiyuan Robotics (Chinese humanoid startup), and UBTech. As production scales from hundreds to tens of thousands of units, the yield bottleneck shifts to accumulated tolerances in structural components rather than algorithms — this round, with heavy participation from local state capital, directs funds primarily toward physical manufacturing capacity. Source: Robot Lecture Hall source

IO-AI Tech | new round | several hundred million RMB · adjacent

Jointly invested by Shunwei Capital, Matrix Partners China (Songhe Capital), and Shenzhen Capital Group, with strategic investment from a leading robot-platform company. Founded in 2023 in Shenzhen, the company builds embodied-data infrastructure rather than pure data-collection outsourcing, with a product line covering the general-purpose teleoperation collection system TeleXperience, the real-world human data collection system SenseXperience, and the data management platform EmbodiFlow. The company says it has served over 150 robot-platform and model companies worldwide, adapting to over 80 robot models cumulatively. Source: PEdaily source

Qiaojie Digital Matter | Pre-A+++ round | over RMB 100 million · embodied

Led by China Mobile Chain-Leader Fund, with follow-on investment from existing investors Fosun Chuangfu and Shenzhen Capital Group. The company builds a general-purpose robot "cerebellum," focused on bipedal and quadrupedal locomotion control, with its solution already covering over 20 robot models; it says it can compress the traditional 3–4 month locomotion-control development cycle down to weeks, with algorithms iterating monthly. The funds will support the strategic rollout of its "general-purpose robot operating system." Source: Guandian.cn source

Alloy Robotics (San Francisco, USA) | seed round | $8 million | valuation approx. $80 million · adjacent

Led by Square Peg. The product is an AI agent specialized in root-causing robot failures — CEO Joe Harris's observation is that the answer usually already sits in the robot's own logged data, but engineers must dig through it one run at a time across manually shared files, cloud storage buckets, SQL queries, Python scripts, and replay tools, taking hours to days. The company says its customer count has grown over 50% month-over-month since commercialization, and it has processed data from over 10,000 robot runs. Source: Forbes source

Cambridge Aerospace | new round | $300 million · adjacent

A defense drone developer; this round is among the largest in the current defense autonomous systems sector. Source: The Robot Report source

Turing (Japan) | planning U.S. office | targeting $10 billion IPO valuation · autonomy ⚠️ Company projection

The Japanese autonomous-driving startup plans to establish a U.S. office and has set a target IPO valuation of $10 billion. Both the target valuation and listing timeline are the company's own plans, with no offering arrangements in place yet. Source: Guandian.cn source

III. Commercialization and Deployment

WeRide's H1 revenue reaches RMB 350 million, up 73% YoY, with overseas/international revenue up 154% · autonomy

WeRide (NASDAQ: WRD, HKEX: 0800) reported Q2 and H1 2026 results: total H1 revenue of RMB 350 million, up 73% year-over-year; Q2 revenue of RMB 230 million, up 82% year-over-year and up 103% quarter-over-quarter, with gross margin rising to 37.5%, 9.4 percentage points higher than the same period in 2025. Overseas/international expansion is the main driver of growth, with H1 overseas/international revenue up 154% year-over-year and Q2 up 164% year-over-year. As of July 31, 2026, the company's global L4 fleet numbered approximately 3,400 vehicles, including over 1,800 Robotaxis; Q2 L4 business revenue rose 131% quarter-over-quarter, and Robotaxi registered users rose 35% quarter-over-quarter. The other growth leg is mass-production ADAS: L2++/L3 business revenue rose 219% quarter-over-quarter, with roughly 30,000 units of its self-developed end-to-end system WRD 3.0 shipped, production designations secured for over 30 vehicle models, and a joint L3 feature proof-of-concept underway with a German luxury automaker. The business now covers 13 countries and over 60 cities. Source: SmartHey source, Bloomberg source

Waymo has imported over 3,200 Zeekr-chassis Robotaxis, with per-vehicle hardware cost estimated roughly 48.5% below the previous generation · autonomy ⚠️ Estimate

According to Forbes reporting, corroborated by customs records, Waymo has cumulatively imported over 3,200 Ojai Robotaxis built on Zeekr's platform, with more than 2,600 recorded in 2026 alone; over 500 were present simultaneously at its Mesa, Arizona vehicle-integration facility in August. These vehicles are manufactured in Ningbo by Geely-owned Zeekr and imported without autonomous-driving systems installed — the Chinese-made portion covers the body, battery, and drivetrain, with Waymo's sixth-generation Driver (4 lidars, 6 radars, 13 cameras) fitted in the U.S. On cost, U.S. customs filings put the CM1e chassis factory price at approximately $38,000–$38,500; applying the upper bound plus a 102.5% tariff puts the landed cost at approximately $78,000. Adding Waymo's disclosed sixth-generation hardware cost of approximately $25,000 brings the total vehicle-plus-autonomous-hardware cost to roughly $103,000 — about 48.5% below the previous-generation Jaguar I-Pace setup, estimated at approximately $200,000 (vehicle $75,000 plus autonomous hardware $125,000). It should be noted that final integration cost is not publicly disclosed, so the $103,000 figure does not equal total landed vehicle cost; this comparison is based solely on the vehicle and hardware figures that have been disclosed. Source: CarNewsChina source

Ukrainian military has ordered over 3,000 ground robots cumulatively, with 640 units in July alone · adjacent

Via the military-technology marketplace operated by the state-backed innovation platform Brave1, the Ukrainian military has ordered over 3,000 ground robots since the marketplace launched last year, with over 640 units purchased in July alone — a 680% increase over the 2025 monthly average. Brave1 Market product manager Yuliia Myrna says ground robots and interceptor drones are the two fastest-growing categories in 2026; a single ground robot can cost tens of thousands of dollars, more than most drones, but offers diverse functionality and longer service life. Ukraine is deliberately accelerating their use for high-risk tasks such as logistics and casualty evacuation, incentivizing troop adoption through a points-based system. Brave1 CEO Andrii Hrytseniuk says the number of companies engaged in ground-robot production has grown from a handful before the war to about 280 today. Source: SSBCrack source

Unitree's cumulative humanoid robot production reaches approximately 18,000 units · humanoid

Following this week's IPO subscription being oversubscribed 8,288 times, Unitree Robotics disclosed that its cumulative humanoid robot production has reached approximately 18,000 units. This figure refers to cumulative units off the production line, not deliveries or shipments, and cannot be directly compared with recent industry shipment-ranking figures. Source: Jiemian News source

AgiBot livestreams sorting at 1,816 items/hour, with simpler hardware · embodied ⚠️ Company claim

AgiBot Robotics (Chinese embodied-AI startup, formerly known as Zhiyuan Robotics) held a one-hour fully autonomous logistics-sorting livestream, continuously sorting real packages of random size, shape, material, and weight at a speed exceeding 1,816 items/hour with over 98% accuracy, which the company says is about 45% faster than the 1,248 items/hour previously demonstrated by a U.S. company. Notably, the hardware direction is the opposite of what one might expect: AgiBot uses dual robotic arms with grippers rather than a "humanoid plus five-fingered dexterous hand" setup — partner and algorithm lead Gan Ruyi says this cuts hardware cost by 70%, while placing higher demands on the embodied foundation model, backed by their unified world model WALL-B, which fuses vision, language, touch, action, and physical prediction. A one-hour livestream demonstrates capability but is still a separate question from stability under continuous production-line operation. Source: Sohu (reposted) source

IV. Industry Developments

A3: North American Q2 robot orders reach 8,940 units, while automotive OEM orders fall 25% in H1 · industrial

According to data from the Association for Advancing Automation (A3), North American companies ordered 8,940 robots in Q2 2026, worth $622 million — up 4.3% year-over-year in units and 21.3% in value; H1 totals reached 17,995 units worth over $1.16 billion, up 2.0% and 6.6% respectively versus the same period in 2025. The real story is structural: H1 automotive OEM orders fell 25%, while semiconductor and electronics orders grew 35% and life sciences/pharmaceuticals grew 32%, exactly offsetting the automotive weakness. In Q2, semiconductor and electronics orders were up 38% year-over-year, automotive parts up 20%, and food/consumer goods and metals each up 18%. Non-automotive customers accounted for 56% of Q2 order volume, continuing the trend of robots spreading into more industries. Collaborative robots accounted for 2,774 units and $114 million in H1 orders — 15.4% of units and 9.8% of value — with especially strong penetration in life sciences and semiconductors, at 43.7% and 36.5% of each sector's H1 orders respectively. A3 executive vice president Alex Shikany says automotive remains an important source of demand, but the breadth of growth outside automotive OEMs is a trend worth watching closely. Source: DC Velocity source

FCC revokes equipment certification for a "U.S.-branded front" company for the first time · adjacent

The U.S. Federal Communications Commission issued a revocation order (DA 26-839) on August 11, permanently withdrawing drone and controller equipment certification from Delaware-registered Odyssey Robot LLC. Under rules effective December 2025, drones and key components manufactured in covered foreign countries are placed on a "Covered List," barring them from the equipment certification required for import, marketing, and sale. The FCC's investigation found that Odyssey claimed its products were designed in California and assembled by eTak Worldwide in Grand Prairie, Texas, but found no evidence of any domestic-manufacturing records — constituting a false statement. The investigation was triggered by research published in June 2026 by independent security researcher Konrad Iturbe; a prior order to show cause gave Odyssey a 10-calendar-day window to respond, which it did not use. This marks the FCC's first use of its expanded national-security authority to retroactively strip equipment approval from a company alleged to be a "U.S. front" for foreign-made hardware, with chairman Brendan Carr previously stating he is monitoring other white-label import brands using similar tactics. The FCC is also weighing whether to broaden a wider "military-grade" definition to include lidar, thermal imaging, and spraying payloads, expanding the scope of restrictions. Source: The Drone Girl source

Honor unveils Robot Phone, packing dexterous-hand-grade joint motors into a smartphone · hardware ⚠️ Company claim

Honor held a global launch event in Guangzhou, unveiling the Robot Phone, priced at RMB 9,999 (12GB+512GB) and RMB 12,999 (16GB+1TB), going on sale August 18. What's more interesting for Physical AI readers than the phone itself is its four-degree-of-freedom titanium-alloy dexterous gimbal: the gimbal motor weighs just 2.6 g, which Honor says is 34% smaller than dexterous-hand joint motors used in embodied AI, with machining precision reaching ±0.005 mm; the mechanical system integrates over 100 precision components, with a maximum tri-axis control speed of 360°/s. On the software side, the system is powered by an Agentic OS kernel, debuting the YOYO Pro mode, co-developed with Alibaba's Qwen for on-device capabilities; the gimbal can automatically frame shots following hand gestures, music rhythm, and call participants. Claims such as "world's first" and "pioneering" are the company's own framing, but once consumer-electronics production scale genuinely drives down the cost of this kind of miniature high-precision joint module, the benefits could extend well beyond smartphones. Source: Phoenix.com Reviews source

Dyna Robotics releases world-action model Dyna-2, claiming the first validation of a cross-embodiment scaling law · world-model ⚠️ Company claim

Dyna Robotics has released Dyna-2, a single generative model capable of denoising future video and actions either jointly or separately, pretrained on a dataset of over 1 million hours of first-person human manipulation video, 43.8 million clips, 97,160 task instructions, and 9,917 object types. The company says this is the first time a scaling law has been validated across platforms with differing embodiment — as pretraining data scales from 1,000 hours up to 1 million hours, the model's predictive performance on held-out human data rises monotonically, with an inflection point between 10,000 and 100,000 hours. Compared under matched conditions with the company's previous-generation VLA architecture, Dyna-1, the world-action model achieves a success rate 1.55x that of the VLA; more telling are the deployment numbers: both generations achieve 100% success in internal testing, but in actual customer environments Dyna-1's pass rate was 46% versus 87% for Dyna-2. The company was founded in 2024, is headquartered in the U.S. with a hardware R&D center in Shanghai, and closed a $120 million funding round last September with participation from Nvidia, Samsung, LG, and Amazon. All figures above are the company's own claims, not yet independently reproduced. Source: iFeng/Zhidongxi source

Shanghai Electric and Shanghai FANUC Robotics form joint venture Shangfa Technology, registered capital RMB 200 million · industrial

Shangfa Technology (Shanghai) Co., Ltd. was recently established, with legal representative Qian Hui and registered capital of RMB 200 million; its business scope includes intelligent robotics R&D, industrial robot manufacturing and installation/maintenance, and AI infrastructure software development. Equity records show the company is jointly held by Shanghai Electric (601727.SH)'s wholly owned subsidiary Shanghai Electric Automation Group and Shanghai FANUC Robotics. Source: Sina News (citing Qichacha) source

Einride partners with DAF Trucks to integrate the Einride Driver into production trucks · autonomy

Dutch truck manufacturer DAF will integrate Einride's autonomous driving system, Einride Driver, to advance autonomous electric freight at scale. Having the OEM itself, rather than a retrofitter, serve as the integrator is the typical path for driverless trucking moving from demonstration fleets to mass-production installation. Source: The Robot Report source

Pudu unveils ET1, bringing commercial floor-scrubbing robots into 100–800 square meter small stores · adjacent

Pudu Robotics has launched the PUDU ET1, the first product in its new E-series, targeting small commercial spaces with cleaning areas of 100–800 square meters, such as convenience stores, pharmacies, chain restaurants, budget hotels, and offices. The product combines floor scrubbing, sweeping, vacuuming, and mopping into a single pass — sweeping in front, scrubbing behind, in one go; the roller brush reaches up to 800 rpm with 20 kPa suction, and it supports 85°C hot-water scrubbing to handle grease in food-service settings, paired with an eight-in-one docking station for charging, water refilling, wastewater discharge, detergent refilling, and roller-brush cleaning/drying. Cleaning automation has mainly been deployed in large venues so far; the barrier to moving into small stores has always been that "the robot can get in, but still needs frequent human intervention."Source: AiThority source

Wulun Technology partners with Lens Technology's robotics unit on mass production of bionic humanoid heads · humanoid

The two companies signed a strategic cooperation framework agreement in Changsha on August 10, covering bionic robot busts and highly articulated head/face assemblies, jointly advancing mass production of bionic humanoid heads. Lens Robotics is a subsidiary of Lens Technology responsible for robot contract manufacturing and core smart-equipment production; the company says it has built annual production capacity for hundreds of thousands of robots, and this marks its zero-to-one expansion into the bionic-robotics segment. Source: China.com source

Hardware · Supply Chain

· Tactile sensors: Zhongke Guiji (Chinese tactile-sensor startup) places 200,000-unit order with Filna New Materials: the two companies signed a strategic cooperation agreement on August 12, under which Zhongke Guiji will procure 200,000 high-performance tactile sensors from Filna New Materials (605488.SH), which it says will support mass production of 40,000 industrial-grade dexterous hands. Zhongke Guiji was incubated from the State Key Laboratory of Multimodal Artificial Intelligence Systems at the Chinese Academy of Sciences' Institute of Automation; its Casia Hand series of dexterous hands has achieved 100% domestic sourcing of core components, with cumulative orders exceeding RMB 100 million in H1 2026. The company plans to reach stable annual production capacity of nearly 30,000 dexterous hands by the end of 2026 and exceed 100,000 units of annual capacity by the end of 2027 (⚠️ Company projection). Filna New Materials reported approximately RMB 2.8 billion in revenue in 2025, with its sensor business still in an early-stage phase. source

· The durability curse of dexterous hands: high-DOF tactile dexterous hands priced from the low hundred-thousands to several hundred thousand RMB often last only a few weeks to two or three months under continuous industrial operation — tendon-driven designs offer strong bio-mimicry and high DOF but lose force easily and are hard to maintain. Zhongke Huisi (Chinese dexterous-hand startup) chairman Zhang Zhengtao told National Business Daily that his solution is a "3-minute tendon swap," built into the company's pre-launch F-series tendon-driven dexterous hand, making tendon-replacement cost nearly negligible. GGII data: China's dexterous-hand market sold approximately 19,200 units in 2025, up 236.84% year-over-year, projected to reach 70,200 units in 2026 and potentially exceed 430,000 units by 2030; from 2023 through May 2026, China's dexterous-hand industry saw approximately 57 funding rounds totaling roughly RMB 12.8 billion. The company's CEO Yang Ping also noted that open-source fine-manipulation datasets for dexterous hands number only in the tens of thousands, orders of magnitude behind the millions of trajectories available for whole-robot manipulation data. source

FutureX · Physical AI Daily — Issue 86 (08/12)

Shawn — Tue, 11 Aug 2026 14:50:18 +0000

Today's Highlights

· Daimon Robotics (Chinese tactile-sensing robotics startup) raises hundreds of millions of yuan led by Ant Group, Ant's first bet on robotic touch

· Global humanoid shipments hit 19,100 units in H1, with Chinese makers taking 97%

· Zoox launches in Las Vegas, the first paid steering-wheel-free robotaxi in the US

· Hesai Autonomous Driving's H1 revenue reaches 804 million yuan, with over 100 unmanned mining trucks now running routinely at 7 mine sites

· Linghou Robotics raises over 100 million yuan Series C, led by CATL-affiliated Chendao Capital

· Guao Technology raises nearly 100 million yuan Series A+, its micron-precision industrial robots now sold into Bosch's production lines

· RynnValue switches to time-to-goal as its supervision signal, real-robot success rate rises from 52.5% to 72.5%

· Tianjin sets a 2028 target of 20 billion yuan core output value for intelligent robotics

I. Research Progress

RynnValue: using "how much longer to the goal" as the supervision signal, turning the reward model into a foundation model · manipulation

General-purpose reward models are becoming the bottleneck that keeps robot learning from scaling — both preference labels and normalized-progress anchors transfer poorly across embodiments and data sources. RynnValue instead supervises on temporal distance (the directed cost-to-go from the current observation to a language-specified goal), with labels derivable directly from timestamps, letting it scale to over 7,000 hours and roughly 3 million instruction-conditioned segments without any preference or progress annotation. It achieves a Kendall's tau_a of 0.675 on RBM-EVAL-OOD, beating the SOTA under full preference supervision (0.655) and more than doubling the pure-progress baseline (0.292); converted into a dense reward, it lifts real-robot policy success rate online from 52.5% to 72.5%, and offline from 63.8% to 82.5%. The model is open-sourced and is today's top-trending paper on HF (HF↑6).

Dongchi Huang et al. · arXiv 2608.09853 source

SpeedTuning: imitation-learning policies run slow because they're locked to the demonstrator's hand speed · manipulation

Chelsea Finn's team flags a long-avoided deployment issue: imitation-learning policies are inherently capped by the speed of the hardware and the human who collected the data, and almost no one has measured the empirical curve of "how much success rate you lose per unit of speed-up." SpeedTuning uses a lightweight reinforcement-learning framework to predict the optimal execution speed for each action, requiring no extra data collection and layering purely on top of a base policy, achieving over 2.4x speed-up on dynamic and high-precision tasks like pouring, throwing, and grasping — while still outperforming naive acceleration methods such as fixed-rate linear interpolation.

David D. Yuan et al. (Stanford) · arXiv 2608.09138 source

HarnessWAM: World Action Models can only "predict the next segment" — bolting on a planning-capable outer brain · world-model

arXiv saw a burst of eight papers today around World Action Models (WAM, models that jointly model future frames and actions), spanning 4D trajectory-field injection, video-diffusion prior distillation, joint-embedding modeling, and speculative-decoding acceleration — this paradigm's density is now approaching last year's level for VLA. HarnessWAM targets one of its most concrete shortcomings: WAMs can only predict over a limited time horizon and lack global planning, cross-stage state persistence, execution verification, and failure recovery. It bolts on a VLM task manager to maintain scene beliefs and a structured task graph, using an event-driven dual-timescale loop to let the robot recover state after a subtask fails without discarding existing scene knowledge, achieving 59.6% full-task and 69.9% subtask success on RoboMemArena.

Zhaopeng Gu et al. · arXiv 2608.09516 source

SLIM-0.5B: continuous manipulation doesn't actually need open-domain semantics — 0.5B parameters suffice · vla

VLA policies spend most of their parameters supporting open-domain semantics, when continuous manipulation really only needs a compact representation of observation, action, and the state change the action causes. SLIM uses self-supervised masked trajectory prediction to learn action-grounded predictive latents, paired with a Mixture-of-Transformers backbone and flow-matching action generation, matching or beating large-scale VLA and world-action baselines on simulation benchmarks and real-robot evaluations with just 0.5B parameters — with no additional embodiment-specific pretraining and significantly lower inference latency and memory footprint.

Jingkai Wang et al. · arXiv 2608.09771 source

MetaSpace: "metamorphic testing" for embodied agents catches over 90,000 spatial-cognition errors in one pass · benchmark

Existing embodied benchmarks either rely on manually labeled VQA or only look at task completion rate, and the latter lets agents "game" the metric via suboptimal or even unsafe paths, masking real flaws. MetaSpace borrows metamorphic testing from software engineering, automatically generating test cases from the spatiotemporal multimodal state of real execution trajectories, encoding metamorphic relations grounded in logical rules and physical laws as executable Prolog rules — any violation is flagged as a spatial-cognition failure. It detects 90,422 spatial-cognition errors across three classes of embodied scenarios; every SOTA multimodal large-model agent tested scores only 0.44–0.52 on spatial cognition, versus a human baseline of 0.96.

Gengyang Xu et al. · arXiv 2608.07533 source

Training a VLM into a VLA causes depth perception to collapse in the final layers · vla

A counterintuitive negative result: the authors take a weight-aligned open-source VLM/VLA pair (Molmo2-ER and MolmoAct2-LIBERO) and probe layer-by-layer depth decodability, finding the VLA is worse than the original VLM at every layer (which they call the floor), and the degradation is uneven — the base VLM's depth decodability keeps rising in the final layers, while the VLA collapses outright (which they call the cliff). They causally trace this cliff to interference from the final-layer MLP writes: ablating the final-layer MLP write recovers most of the decodability, while equivalent attention ablation, or the same intervention on the base VLM, has no such effect. This implies action post-training is quietly eating away at the vision-language backbone's spatial-geometric capability.

Alexander Hackett et al. · arXiv 2608.08904 source

LUCID: long-horizon humanoid loco-manipulation planned at the high level with "imagined dynamics" · locomotion

Long-horizon humanoid mobile manipulation has mostly relied on scripted planners or finite-state machines to sequence pretrained skills, which breaks down as task sequences grow complex. LUCID first trains a low-level whole-body policy conditioned on structured latent variables via adversarial imitation and freezes it, then jointly learns a high-level policy and a macro-dynamics world model — the latter predicts the temporally extended state transitions induced by latent-variable decisions, letting the high-level policy optimize over imagined rollouts. On multi-object rearrangement tasks, it outperforms existing baselines on both full-task success rate and partial-completion rate.

Cheng Guo et al. · arXiv 2608.07746 source

Other papers today: 4D-WAM injects spatiotemporal awareness into world action models via 3D trajectory fields, a model-agnostic training strategy (arXiv 2608.08023 source); Vid2WAM offline-distills diffusion priors from large video foundation models into a compact WAM, discarding the teacher at inference (arXiv 2608.08558 source); JEPA-WAM learns VLA policies via joint-embedding world modeling, avoiding the overhead of explicit future generation (arXiv 2608.09381 source); SG-WAM adds text-grounding and spatial-awareness semantic guidance to WAMs (arXiv 2608.08839 source); WA-SpecDec applies world-aware speculative decoding to VLA, compressing closed-loop latency (arXiv 2608.08725 source); WorldSimProbe diagnoses "simulation fidelity" in action-conditioned world models, distinguishing accurate transitions from merely plausible ones (arXiv 2608.09298 source); Auditing Instruction-Trajectory Mismatches performs training-free audits of demonstration data, catching samples where the trajectory is correct but the paired language instruction is wrong (arXiv 2608.07895 source); AdaDexGrasp fuses 3D visual and tactile representations for post-contact adaptive grasping (arXiv 2608.07600 source); an ultra-low-impedance gripper drops the high-reduction-ratio transmission and external force sensor to enable high-bandwidth transparent physical interaction (arXiv 2608.09198 source); PhysX-CoT rewrites single-image-to-simulatable-3D-asset generation into an explicit, structured physical-reasoning chain (arXiv 2608.08053 source); CMU-Drive and V2V-VLA are closed-loop reasoning benchmarks for multi-vehicle cooperative driving (arXiv 2608.07621 source); Explore, Map, Remember, Decide asks whether embodied VLMs' spatial memory holds up in safety-critical scenarios (arXiv 2608.08077 source).

Open Source · Tools · Benchmarks

· XPolicyLab: cuts the O(NM) integration cost of connecting N policies to M evaluation environments down to O(N+M) via a unified standard and open ecosystem, already incorporating 42 robot policies; onboarding time for one representative policy drops from over 5 hours to 2 hours, and to 30 minutes with a packaged agent skill; the same adapter suite serves RoboTwin, RoboDojo simulation, and standardized real-robot evaluation (arXiv 2608.09892 source)

· LTX-2.5: LTX releases a new open-weight world model, using a new pixel-diffusion pipeline to reduce visual artifacts and improve cross-generation consistency, optimized for local inference on NVIDIA RTX GPUs, downloadable now on HuggingFace and ComfyUI; the LTX model family has surpassed 33 million cumulative downloads on HuggingFace source

· Daimon's open-sourced tactile dataset: the first open-source batch, a 10,000-hour all-modality manipulation dataset including tactile data, has surpassed 4.4 million downloads, ranking first among embodied-AI datasets on ModelScope; also released alongside is RobOmni, a matching all-modality evaluation benchmark including tactile data source

· EsaacSim: a multimodal event-camera plugin for NVIDIA Isaac Sim, filling a long-standing gap in event-vision tooling for mainstream robot simulation platforms (arXiv 2608.08522 source)

· Ego-OSCAR: an open-hardware head-mounted stereo-inertial capture rig, with hardware-synchronized global-shutter stereo cameras + 6-axis IMU + on-board encoding on an embedded Linux board, built for in-the-wild first-person data collection (arXiv 2608.08285 source)

· 360CityArena: an urban navigation benchmark built from real 360-degree video, filling a gap left by existing outdoor benchmarks that are either not realistic enough or not complex enough (arXiv 2608.08814 source)

II. Funding & Deals

Daimon Robotics | Strategic funding | Hundreds of millions of yuan · embodied

Led by Ant Group, with existing shareholders oversubscribing to follow on. This marks Ant's first move into tactile sensing as a physical-interaction layer, after previous bets on Unitree, Galbot, and AgiBot Intelligence (Chinese humanoid/embodied startups). Daimon spun out of a Hong Kong University of Science and Technology team and began commercializing in 2023, closing a hundred-million-yuan Series A and this new round within two months; its shareholder base now spans Inovance's investment arm, China Mobile's supply-chain fund, Legend Capital, China Merchants Capital, and China Telecom. Co-founder and chief scientist Wang Yu was the first PhD graduate of Matthew T. Mason, known as "the father of robotic dexterous manipulation." The same day, the company released Daimon-TWM, which it describes as the world's first tactile-anchored world model, unifying physical cognition, deliberative decision-making, and instantaneous manipulation into a single framework; the company says it has already deployed the system in 3C precision manufacturing and automotive production lines for "feel"-dependent tasks such as USB plug/unplug and fragile-item sorting, with a pencil-case-organizing demo at WAIC achieving a 95% on-site success rate. ⚠️ Vendor-reported figures The hardware-side numbers are harder to verify independently: the company claims to be the world's first vision-tactile sensor maker to reach 10,000-unit-scale mass production and delivery, with over 200 cumulative customers (more than 50 overseas), having already shipped devices to OpenAI, Figure, Physical Intelligence, Meta, BMW, and Google DeepMind. Its home-data-collection network built jointly with China Mobile has deployed 1,000 units in its first phase, with projected annual output of over 1 million hours of real-scenario manipulation data at full capacity.Source: SmartHey source

Linghou Robotics | Series C | Over 100 million yuan · hardware

Led by Chendao Capital, an industrial investment platform affiliated with CATL, with follow-on from an industrial-machine-tool investment fund, Yuxin Electric Technology's affiliated fund, and a dual-currency fund, plus additional investment from existing shareholders Suzhou Venture Group and Guyu Capital. This core-component maker, founded in Suzhou in 2015, has completed five prior funding rounds — including from NIO — with cumulative funding exceeding 500 million yuan; it now has over 700 employees (300+ in R&D), serves over 900 customers in China, and its product line covers machine-vision systems, direct-drive motor systems, industrial robots, and general-purpose robots. Of the three stated uses of the new funds, the most notable is the third — building Jiangsu province's first provincial key laboratory for embodied-AI robot core technology and components, pushing the company from a component supplier toward setting standards and shared common technologies.Source: 智东西 source

Guao Technology | Series A+ | Nearly 100 million yuan · hardware

Investment from Shenzhen Capital Group, Cornerstone Capital, and Guangdong Technology Financial Investment; Shenzhen Capital Group is an existing shareholder from last June's Series A. Funds are earmarked for mass production of industrial embodied robots and building out a robotics business unit. This company, founded in 2019, makes "multi-degree-of-freedom motors" — using a single mover structure to achieve both linear and rotary motion simultaneously via magnetic-field changes, fully direct-drive with no mechanical transmission; its ZR motors are already in the supply chains of Xinyichang, Changchuan Technology, Goertek, and Sunny Optical. Industrial robots built on this motor have already been sold to Bosch in Germany and are in pre-production trials on several major manufacturers' lines; its G-Tools, released this April, can grasp a 0.5mm-diameter fine needle through a micro-hole with ±2μm positioning accuracy and ±0.05N force control. A comment from founder Li Siyang is worth noting: "3C manufacturers basically won't consider using general-purpose humanoid robots" — the gap isn't that the foundation models aren't strong enough, but that base-body motion control can't meet industrial requirements — "the benchmark isn't Optimus, it's ABB." The company's Ganzhou plant currently has capacity of 1,800–2,000 units/month, and it projects revenue of nearly 100 million yuan in 2026 with profitability.Source: 36氪 source

LatentVerse (Chinese world-model startup) | Seed round | Hundreds of millions of yuan · world-model

Led and followed by Hillhouse Ventures, Clearwater Capital, and Innoangel Fund, with AgiBot and RobotEra (Chinese humanoid-robot makers) both taking stakes. The company was only founded in May 2026, spun out of Tsinghua's Institute for Interdisciplinary Information Sciences; founder and CEO Hu Yucheng studied under RobotEra founder Chen Jianyu and previously worked on embodied foundation models at ByteDance Seed. The team's earlier work — DP3, HiRT, VPP — includes VPP, which has been cited by Nvidia's DreamZero and other world-action-model projects. Its in-development model, UTAM, is a unified foundation model combining a vision-language expert, a world-model expert, an action expert, and a high-frequency tactile expert — the action expert generates a Latent Action, which the tactile expert then reconstructs into a closed-loop executable action based on contact force, slip, and deformation.Source: 机器人前瞻 source

Huilun Technology (incubated within GAC) | New funding round | 100 million yuan · humanoid

This embodied-AI robotics company, incubated internally by Guangzhou Automobile Group (GAC) this February, has completed a 100-million-yuan funding round, headquartered in Guangzhou's Huangpu district. Huilun has taken over all of GAC's business in this area and operates as an independent, market-facing entity, with full in-house capability across structural design, systems platform, motion control, and perception/decision algorithms; it has developed four generations of products to date, and plans call for small-batch trial production of its core product this year, with mass production targeted for 2027.Source: 金融界 source

NdotLight (South Korea) | Follow-on investment | KRW 15 billion · adjacent

Naver's early-stage investment arm D2SF has made its third investment in this 3D synthetic-data company for robot training, with this round sized at KRW 15 billion. South Korea has recently sharply increased its focus on the data side of physical AI, filling in a gap after earlier moves in humanoid hardware and safety certification.Source: Chosunbiz source

III. Commercialization & Deployment

Zoox launches the US's first paid steering-wheel-free robotaxi in Las Vegas · autonomy

Amazon-owned Zoox began paid passenger rides in Las Vegas on August 10, the company's first-ever paid operation, having previously offered only free rides to select riders in San Francisco. Following last week's disclosure that it would begin paid rides "starting next week" with approval to deploy 5,000 vehicles within two years, this step turns the plan into actual revenue. The differentiator versus Waymo, which operates in the same city, is the vehicle itself: no steering wheel or accelerator pedal, built instead with a face-to-face carriage-style cabin, and no formal front or rear — it can drive bidirectionally, with passengers adjusting temperature and music via an in-cabin screen. Notably, around the same time as the launch, Zoox's CEO publicly said "we need to be regulated" after one of its robotaxis drove into a fire scene in Las Vegas — an unusual stance for an autonomous-vehicle company to take.Source: Benzinga source · Fox Business source

Hesai Autonomous Driving's H1 revenue nearly doubles to 804 million yuan; over 100 unmanned mining trucks now running routinely at 7 mine sites · autonomy

Following yesterday's disclosure of cumulative unmanned mining-truck shipments exceeding 1,900 units, up 213% year-on-year in H1, the interim results fill in the financial side of this deployment curve: H1 revenue reached 804 million yuan, up 96.98% year-on-year; gross profit reached 212 million yuan, up 204%; gross margin rose from 17.1% to 26.4%; the period loss narrowed 75.23% year-on-year to 113 million yuan, with adjusted net loss down to 19.68 million yuan. The autonomous-driving business contributed 784 million yuan in revenue, 97.6% of total revenue. The scale of deployment is the hardest data point in the report — over 1,900 mining trucks cover nearly 40 mine sites globally, with 7 sites now running over 100 trucks routinely, and one single site running over 220; the 500-vehicle purchase order signed with Guangna Group this February is, as of the reporting period, the largest single formal procurement order in the global unmanned mining-truck sector. Some financial items still bear watching: total borrowings rose to 687 million yuan, the capital-to-liability ratio rose from 29.4% to 41.1%, and the company flagged financial-guarantee and loan-impairment risk related to "Customer K," with impairment provisions of 81.6 million yuan already booked.Source: 集微网 source

Kodiak × Atlas Energy: unmanned heavy-truck fleet expanding from 28 to 100 trucks, public roads early next year · autonomy

Atlas Energy Solutions plans to expand its Kodiak-powered unmanned heavy-truck fleet — which hauls sand to oil and gas well sites in the Permian Basin — from 28 to 100 trucks by mid-2027. The trucks currently run on private industrial roads and are expected to begin operating on public roads in early 2027, though the timeline still depends on operational milestones.⚠️ Company plan, not yet realized What has already changed is the operating model: Atlas has moved from single-loading-point operations to dispatching simultaneously from two yards roughly a 90-minute drive apart in Texas and New Mexico, connected via its 42-mile Dune Express conveyor-belt system for the final leg. Note a discrepancy in the figures — the 28-truck figure here refers to Atlas's own fleet alone, and is not the same statistic as the overall Kodiak fleet size reported last week.Source: Fox News source

Agility's Digit V5 set for December delivery to early customers, pitched as "no fencing required" · humanoid

Agility Robotics will deploy its next-generation Digit V5 to early customers in December; the selling point is not new motions but an upgraded safety system — the robot can work in close proximity to workers without traditional safety fencing, saving on floor space, facility modification, and workflow-redesign costs, which is exactly the most practical barrier to humanoids entering warehouses and factories. The company says the model is designed to support a 20-hour workday and swappable end-effectors, with autonomous tool-changing left for a later stage.⚠️ Company plan, not yet realized Agility previously disclosed plans to go public via a $2.5 billion SPAC, raising over $620 million.Source: Interesting Engineering source

Hong Kong approves "air-rail interchange" autonomous-driving trial in West Kowloon, 4 private cars with backup operators · autonomy

Hong Kong's Transport Department announced on August 11 that it has approved, under the Road Traffic Ordinance, an "air-rail interchange" autonomous-vehicle trial running between the High Speed Rail's Hong Kong West Kowloon station and the Airport Express Kowloon station, involving 4 autonomously operated private cars. A backup operator will remain on board throughout the trial and take over if necessary. The scale is small, but it connects autonomous driving to the transfer route between a cross-border high-speed rail line and the airport express — a new scenario beyond Hong Kong's earlier minibus road trials.Source: 观点网 source

CJ OliveNetworks builds AI logistics system for HD Hyundai Electric's Cheongju plant · industrial

CJ OliveNetworks will build a next-generation smart logistics platform at HD Hyundai Electric's Cheongju site, bringing AI-driven logistics scheduling to heavy-electrical-equipment manufacturing. It's a typical path for South Korean conglomerates: an in-group IT subsidiary taking on smart-logistics upgrades for a sister manufacturing company.Source: The Korea Times source

Tate deploys 58 Hirebotics collaborative welding robots across multiple plants · industrial

Tate, a maker of data-center flooring and airflow-management equipment, has deployed 58 Hirebotics collaborative welding robots across its plants. Welding is one of the few processes where cobots have already scaled from single-site demos to multi-plant replication — deploying dozens of units at once says more about process standardization than a single pilot site does.Source: The Robot Report source

Tencent's ride-hailing service formally integrates WeRide's (Chinese autonomous-driving company) Robotaxi · autonomy

WeRide's robotaxi service is now integrated into Tencent's ride-hailing entry point, part of a push to expand consumer-facing distribution channels. What China's robotaxi operators generally lack isn't vehicles but traffic — integrating into a super-app is another aggregation move following earlier integrations with Amap and the WeChat ecosystem.Source: 36氪 source

IV. Industry Developments

Global humanoid shipments hit 19,100 units in H1, Chinese makers take 97%, but researchers warn hardware is outpacing the underlying technology · humanoid

Following yesterday's report that AgiBot overtook Unitree with roughly 8,400 units for 44% global share, this report fills in a more complete picture: citing data from San Francisco-based research firm Smart Analytics Global (SAG), Southern Metropolis Daily reports global humanoid-robot shipments reached about 19,100 units in H1 2026, up 272% year-on-year from 5,100 in H1 2025, with Chinese manufacturers accounting for over 97% of global shipments and China itself accounting for over 85% of global demand. Unitree grew 170% year-on-year to about 5,900 units, ranking second with 31% share; its flagship G1 model has shipped roughly 11,000 cumulative units as of mid-2026. Behind the top two, Galbot, UBTECH, and AgileX shipped roughly 900, 700, and 600 units respectively in H1 — a clear gap from the leaders. The structural shift matters more than the total: industrial and commercial applications now account for over 70% of global shipments, up from about 50% in H1 2025, with manufacturing and warehousing/logistics becoming the main deployment channels. SAG projects nearly 60,000 units shipped for full-year 2026, with industry revenue of about $1.6 billion. But the report states plainly that "hardware scale is growing faster than the underlying technology is maturing" — real training data, model capability, reliability, manipulation precision, and cost remain constraints on broader deployment; separately, the FCC added "advanced robotics equipment" to its regulated list on July 28, which could create headwinds for Chinese suppliers' expansion into mature overseas markets.Source: 南方都市报 source

Dexterous hands raise over 25 billion yuan in H1 alone, and the industry still hasn't converged on a "standard hand" · embodied

According to incomplete statistics from IT Juzi, China's dexterous-hand sector raised about 16.877 billion yuan in total funding across all of 2025, while H1 2026 alone already exceeded 25 billion yuan — more than the full prior year in half the time.⚠️ Compiled estimate The reason money is piling in here is cost share — Morgan Stanley has estimated dexterous hands account for about 17.3% of Tesla Optimus's bill-of-materials cost, and industry participants generally put hardware cost at 10%–20% of the full unit, with high-end configurations potentially exceeding 20%. But capital enthusiasm and industry chaos coexist: shipment volumes across companies range from hundreds to 100,000 units, unit prices from just over 3,000 yuan to several hundred thousand yuan, and degrees of freedom from 6 to 42, all measured inconsistently. Xuanxiang Robotics CEO Su Jiuyu describes the current landscape as "unfocused" — there's no single dominant mainstream technical direction, application scenarios remain concentrated in research, and this year "could be called the first year of dexterous-hand mass production, and also the early stage of a bubble." One telling mismatch: the company with the highest shipment volume isn't necessarily the one with the highest revenue or valuation.Source: 36氪 source

Tianjin sets a target: over 20 billion yuan core output value for intelligent robotics by end-2028, Beijing-Tianjin-Hebei local-sourcing rate to 60% · industrial

Tianjin's Bureau of Industry and Information Technology, Development and Reform Commission, and Science and Technology Bureau jointly issued the "Tianjin Intelligent Robotics Industry Innovation Development Action Plan (2026–2028)" on August 10. The target for end-2028: core output value exceeding 20 billion yuan, average annual output-value growth over 15%, and the Beijing-Tianjin-Hebei regional local-sourcing rate raised to over 60%, alongside cultivating at least 3 companies with over 1 billion yuan in output value and 20 nationally recognized "little giant" specialized-and-innovative enterprises. Named technical priorities include integrated drive-control controllers, highly integrated joints, high-performance servo motors, high-precision reducers, multimodal fusion sensors, and dexterous hands, with an explicit goal of cultivating single-item manufacturing champions in precision transmission. On the application side, the plan calls for at least 30 demonstration application scenarios in public services. The plan's confidence is backed by existing figures — in H1 2026, Tianjin's above-scale intelligent-robotics enterprises posted total industrial output value of 6.849 billion yuan, up 15.2% year-on-year, with value-added growth up 18% year-on-year, an acceleration of 2.5 percentage points over Q1.Source: 证券时报 source

DYNA-2 claims pretraining on 1 million hours of human video, with zero robot data throughout · world-model

DYNA Robotics has released DYNA-2, architecturally a video-generation-based World-Action Model that, during pretraining on human video corpora, jointly optimizes for "predict the next frame" and "predict the next action." The company says the resulting contact-physics and spatial-reasoning capability transfers across embodiments — to fixed robot arms, humanoid prototypes, and five-fingered dexterous hands — despite none of these robots appearing in pretraining. The company's example: 13 minutes of data taught a pair of five-fingered hands to unscrew a bottle cap; it also claims the video co-training algorithm improves instruction-following scores by 133%.⚠️ Vendor-reported figures These are all self-reported internal comparisons (against the company's own prior VLA model, DYNA-1), with no third-party replication yet; the real point of interest is the claimed scaling law — if the human-video curve holds, scaling to 10 million hours would mean simply collecting more video rather than building out teleoperation robot fleets.Source: Unite.AI source

South Korean media report Samsung is mobilizing company-wide resources for its own humanoid, alongside world-model and actuator work · humanoid

ZDNet Korea reports that, beyond its subsidiary Rainbow Robotics, Samsung Electronics is independently developing its own humanoid robot, with R&D reportedly having reached "a fairly substantial level."⚠️ Report based on unconfirmed sources Two organizational moves are independently verifiable: last month's formation of RX (Robotics eXperience), a business promotion office reporting directly to Representative Director Roh Tae-moon; and the appointment of Vice President Lee Dong-gun from the Planning Group as head of the Robotics Strategy Group — he previously led robotics strategy at Hyundai Motor and was responsible for directing Boston Dynamics' operations. The report also says Samsung is developing the actuators and world model needed for its own humanoid, and plans to build a data factory at its Gumi site to feed existing production-line process data into robot training.Source: IT之家 source

UBTECH proposes an H-share incentive plan while applying for up to 2.6 billion yuan in bank credit · humanoid

UBTECH has proposed adopting an H-share incentive plan with a 10-year term and an authorization cap of 10% of total issued shares, and is applying for up to 2.6 billion yuan in RMB bank credit. This financing-and-incentive package lands the same day SAG listed UBTECH's H1 shipments at around 700 units and expressed reservations about its emotional-companion/bionic-styling approach — it reads more like groundwork for the company's next phase of cash burn.Source: 观点网 source

Automakers and full-machine manufacturers rush to set up new robotics entities: BAIC forms Yuanqi Physical Intelligence, UBTECH invests in a new company · embodied

BAIC Group has formed Yuanqi Physical Intelligence Technology Co., Ltd., with robotics R&D in its registered business scope; the same day, UBTECH and other investors formed a new robotics company with 10 million yuan in registered capital. Combined with GAC-linked Huilun Technology, which also completed a funding round today, spinning off embodied-AI business from the parent automaker into an independent entity has become standard practice — an independent entity makes market-based fundraising and equity incentives easier, while directly transferring the automotive supply chain's cost capabilities.Source: Gasgoo source

Waymo co-CEO: safety gains from a vision-only approach will plateau quickly · autonomy

Waymo's co-CEO publicly stated that safety improvements from a vision-only autonomous-driving approach will hit a ceiling quickly. This is a reaffirmation of the multi-sensor-fusion camp's position amid intensifying robotaxi scaling competition, an executive statement rather than new evidence. The same day, another report noted that while Waymo's robotaxi safety is improving, scaling up is exposing a new class of problems.Source: 디지털투데이 source

Tesla shows off its first Cybercab with an integrated Starlink antenna · autonomy

Tesla has unveiled its first Cybercab with a Starlink antenna integrated into the roof, with Musk saying the configuration will roll out to every vehicle. For a steering-wheel-free robotaxi, the point of an onboard satellite link is remote takeover and dispatch redundancy, not in-cabin entertainment — precisely the reliability issue regulators have been pressing on lately.Source: Electrek source

Hardware · Supply Chain

· Lens Technology · joint-module production line: Jiang Nan, President of Lens Technology's China operations and board secretary, disclosed that the world's first fully automated robot joint-module production line has been completed at its subsidiary Lens Intelligent (Changsha); the company has already delivered core components — humanoid and quadruped full units, joint modules, dexterous hands — to multiple leading robotics customers in China and overseas, with 2026 full-unit robot shipments surpassing 10,000 units, and has developed lightweight solutions including aluminum-magnesium alloy structural components. "World's first" per company statement source

· SGMW Industrial · magnesium-alloy structural components: delivered its first batch of magnesium-alloy "robot skeleton" components, migrating automotive lightweighting materials and processes onto humanoid robot bodies — pointing, like Lens Technology's aluminum-magnesium-alloy approach, at the same bottleneck: body weight reduction directly determines joint torque and battery-life budgets source

· Fullhan Microelectronics · robotics chips: the company confirmed in response to inquiries that it is already supplying chips to humanoid and industrial robotics companies, another Chinese chipmaker confirming shipments to robotics customers, following GigaDevice's robotics MCU disclosure source

FutureX · Physical AI Daily — Issue 85 (08/11)

Shawn — Mon, 10 Aug 2026 14:43:09 +0000

Today's Highlights

· AgiBot (Zhiyuan) shipped about 8,400 humanoid robots in H1, overtaking Unitree to claim 44% of the global humanoid market

· Unitree's Shanghai STAR Market IPO subscription was oversubscribed by 8,288x, with an allotment rate as low as 0.0181%

· Archer acquires Boeing's Wisk, Insitu, and SkyGrid; Boeing gets nearly 20% equity in return

· Sony and TSMC invest about $6.4 billion to jointly build a physical AI image sensor production line

· Pony.ai's autonomous driving mileage passes 100 million kilometers; fourth-generation driverless heavy trucks enter mass production

· Knowin (Chinese embodied-AI startup) raises RMB 500 million in an angel++ round; its home robot folds down to under 40 cm

· A Beijing municipal key laboratory for embodied large models, led by Galbot, is inaugurated

· Hidream Driving (Chinese autonomous mining truck maker)'s cumulative driverless mining truck shipments exceed 1,900 units, up 213% year-on-year in H1

I. Research Progress

SimWAM: treat video generation purely as a training signal, and discard the entire video branch at inference · autonomy

World-Action Models bring video dynamics priors to end-to-end autonomous driving, at the cost of having to generate future frames at inference time, keeping latency stubbornly high. SimWAM uses an isolated attention mask so that action prediction depends entirely on nothing but the past — no future frames required — allowing the video branch to be discarded after training, leaving only a self-contained trajectory planner. The video expert and action expert share no parameters and interact only through a unified attention interface, so the video backbone is swappable and the action expert can be scaled independently. It achieves 91.5 PDMS on NAVSIM, with latency significantly lower than prior WAM-based planners, and transfers zero-shot to nuScenes; code and weights are open-sourced. (HF↑22)

Zongchuang Zhao et al. · arXiv 2608.07468 source

DriveTeach-VLA: first teach the model where to look, then teach it how to drive · autonomy

A team from Tsinghua AIR, Beihang University, and Didi argues that autonomous-driving VLA training data skews too heavily toward text-based VQA and chain-of-thought — the language reasoning looks elegant, but attention wanders: a model may say it will "yield to the pedestrian" while its generated trajectory isn't actually focused on that pedestrian. DriveTeach-VLA uses GroundingDINO to label elements like vehicles, pedestrians, and traffic lights for visual self-distillation, then projects expert trajectories from BEV back into image pixel space as prompts, so the base model's existing visual capabilities can be put to direct use. It reaches 90.4 PDMS on NAVSIM, rising to 92.7 with a trajectory selector added; code and models are fully open-sourced.

Yang Yuguang et al. (Beihang University / Tsinghua AIR / Didi) · arXiv 2607.01658 source · accepted at ECCV 2026

AtlasVLA: using only a single wrist camera, yet beating multi-view baselines · vla

The reactive paradigm in VLA has two hard failure modes on partially observable, long-horizon tasks: perceptual forgetting once an object leaves the field of view, and task-progress forgetting across multi-step execution. AtlasVLA proposes a dual-memory architecture — lifting instantaneous 2D observations into a voxel-hashed 4D persistent world-state memory, paired with a working memory that tracks the agent's own history and task progress, both conditioning a diffusion transformer. It improves absolute success rate by 9.4 percentage points on LIBERO-Long and by 17.5 percentage points on real-robot long-horizon tasks.

Guiyu Zhao et al. · arXiv 2608.06729 source

Move the camera and the VLA breaks — this paper fixes it with cross-view action consistency · vla

A VLA fine-tuned on a fixed-camera scene will fail if the camera is simply repositioned, even with the task, objects, instructions, and robot state all unchanged. The authors reset the same LIBERO demonstration to an identical MuJoCo state, render both a nominal and a perturbed viewpoint, and constrain the predicted action-flow velocity fields of the two to agree at the same flow coordinates. This achieves 87.2% on the LIBERO-Plus camera-perturbation track, 7.4 points above pure flow-matching training on the same data and 12.5 points above naive mixed-camera SFT, with no drop in nominal-view performance; on real robots with unseen camera positions, success rate rises from 53.3% to 74.4%. A shuffled-pairing control collapses to 25.8%, confirming the gains come from action-equivalent pairing rather than data volume.

Bingqi Huang et al. · arXiv 2608.06965 source

C2Dex: extracting stable contact from monocular human-hand video, then transferring it to dexterous hands · manipulation

High-quality demonstrations for dexterous manipulation are expensive and scarce, while monocular videos of humans are abundant — the challenge is that reconstructed hand-object contact sequences are temporally unstable and physically implausible, and task-relevant contact geometry gets lost during retargeting. C2Dex aggregates noisy per-frame observations in the object's canonical space to recover stable object-side contact, which then serves both as a trajectory-level constraint for reconstruction and as a transfer target across different hand morphologies. It achieves end-to-end trajectory success rates of 57.78% and 26.67% on DexYCB and TACO, versus 17.78% and 10.00% for the strongest baseline under the same evaluation protocol.

Jie Ren et al. · arXiv 2608.07045 source

A 2.5×3 mm IMU as a tactile sensor, localizing to millimeter precision within 180 ms · perception

In fine manipulation, the fleeting light collisions between a grasped object and its environment are effortless for humans to localize, but robots often lack the resolution and end up making clumsy hard contact. TECDAR embeds a 2.5×3 mm 6D inertial measurement unit in the gripper fingertip, sampling at 7 kHz to capture sub-millisecond deformation while streaming only 84 KB/s of data; fused with robot pose via an extended Kalman filter, it achieves millimeter-level localization within 180 ms, with average precision around 7 mm on line-contact and point-contact tasks — precise enough for millisecond-scale trajectory correction.

Haowen Zheng et al. · arXiv 2608.07075 source

Forward prediction alone isn't enough: grounding the latent space of JEPA world models in physical state · world-model

JEPA-family world models learn action-conditioned latent dynamics from observation sequences, but a pure forward-prediction objective offers no guarantee that individual latents encode the robot's own physical state, or that latent pairs encode state changes — limiting downstream planning and policy learning. PSG-JEPA adds two grounding objectives that apply only during training — tying individual latents to proprioceptive state and latent pairs to multi-horizon joint-angle changes — at zero added cost to the inference architecture or compute. It outperforms prior latent-space world-model baselines across three levels: probe identifiability, goal-conditioned planning in a frozen latent space, and policy learning in both simulation and on real robots.

Haodong Yan et al. · arXiv 2608.06799 source

When collaboration becomes an attack surface: communication attacks on LLM multi-robot systems · benchmark

Prior embodied-safety research has largely focused on single robots, leaving the communication risks of multi-robot collaboration unexamined. The authors construct both an external-entry-point attack setting and an internal-privileged-attack setting, tested across three communication architectures (DMAS, HMAS-1, HMAS-2), three LLMs, and five embodied tasks; unsafe information turns into unsafe actions across all three architectures — DMAS's entry point achieves a 96.7% acceptance rate with 100% activation once accepted, and HMAS-1's unsafe-action success rate reaches 97.8%. Their proposed traceability-and-verification gate (CPV Gate) drops the violation rate from 70.0% to 36.6% — a substantial reduction, but far from a safe margin.

Zhen Huang et al. · arXiv 2608.06830 source

Capek 0.5: organizing embodied VLM capabilities by execution role, not by dataset · vla

Robot execution is iterative — every action reshapes the scene, continually updating what needs to be perceived, reasoned about, and verified. Rather than organizing training by dataset or task, Capek 0.5 splits capabilities into four families by their functional role in the execution chain — spatial reasoning, temporal understanding, action grounding, and state verification — each first learned by a specialist trained with verifiable-reward reinforcement learning from a shared backbone, then merged via weight merging and routing-based policy distillation into a single model. It's released in 2B and 35B-A3B sizes, along with the accompanying Capek-StateBench state-verification benchmark.

Ying Chen et al. · arXiv 2608.06756 source

Other papers today: EMS, an adaptive VLA inference framework, uses a reinforcement-learned switching policy to schedule between a large and a small decoupled system, pushing effective action frequency to 93.4 Hz on LIBERO (arXiv 2608.06434 source); TEMPO splits a VLA's semantic projection layer and action expert into two time scales for RL post-training, so fast updates no longer overwrite high-level semantics (arXiv 2607.07314 source); PILOT has world-action models explicitly model state-transition tokens as a motion chain-of-thought (arXiv 2608.06994 source); WNM-3D adds 3D scene conditioning to world-navigation models for continuous VLN (arXiv 2608.07267 source); Spatiotemporal Agility uses time-constrained RL to let a quadruped perform vision-guided dynamic ball catching (arXiv 2608.06907 source); LyEvO combines Lyapunov stability analysis with statistical model checking for safe sim-to-real transfer (arXiv 2608.06481 source); SoRoMoX implements a differentiable, GPU-parallel Cosserat-rod model for soft robots (arXiv 2608.06650 source); CrossTracer performs cross-embodiment navigation via trajectory residuals (arXiv 2608.06688 source); LifelongCrossNav adds persistent 3D semantic memory to multi-goal cross-floor navigation (arXiv 2608.07079 source); adaptive gait timing gives fault-tolerant locomotion control for large quadrupeds under hardware failure (arXiv 2608.07328 source); Vernata does self-supervised pretraining for LiDAR point representations (arXiv 2608.06919 source); layer-by-layer probing and pruning of planning tokens in driving VLAs, probing how much language-model depth actions actually require (arXiv 2608.07361 source).

Open Source · Tools · Benchmarks

· DRENDS: a robotic endoscope depth dataset for dynamic scenes, published on the Nature journal platform, addressing the shortage of ground-truth depth data in soft-endoscope scenes source

· VicOne Isaac Sim extension: turns robot-attack research presented at DEF CON 34 into a free NVIDIA Isaac Sim extension, enabling robot security testing in simulation source

· Axis Robotics simulation dataset: a new simulation dataset released for robot manipulation research source

II. Funding & Deals

Archer Aviation × Boeing | Acquisition + Strategic Investment | Boeing to hold nearly 20% equity, plus up to $55M additional investment · adjacent

Archer is acquiring autonomous-flight company Wisk Aero, drone maker Insitu, and airspace-management software company SkyGrid from Boeing in one move; both sides signed a definitive agreement on August 10. Consideration is newly issued Archer Class A shares and warrants, giving Boeing roughly nearly 20% equity, along with cross-licensing rights to use Wisk's core autonomous-flight system on Boeing's own civil and defense platforms, plus a commitment to invest up to $55 million in Archer's next funding round. For Archer, this is more than just filling out its autonomous-flight capability: Insitu has more than 3,500 drone systems in service across 35 countries, contributing a profitable defense business with over $200 million in annual revenue, and the nearly 2 million combined flight hours across the three companies will feed into Archer's ZEE AI platform. Archer's stock jumped roughly 20% on the news.Source: Boeing IR / PR Newswire

Sony × TSMC | Joint Venture | About $6.4 billion (¥1 trillion) | Sony ~60% / TSMC ~40% · hardware

The two companies will form a joint venture this year to produce next-generation image sensors at Sony Semiconductor's Kikuyo, Kumamoto Prefecture plant, targeting mass production by 2029 — an investment on the scale of four years' worth of capital expenditure for Sony Group's semiconductor division, with Sony also in talks with the Japanese government over subsidies. The two firms were previously in a foundry-and-designer relationship; shifting to a joint venture is meant to respond faster to AI-driven demand shifts — image sensors are the "eyes" of physical AI, powering robots and self-driving cars. Sony has long held roughly a 50% share of the CMOS image sensor market, but pressure has mounted sharply since Samsung Foundry won the iPhone image-sensor order last August, with both China's OmniVision and Samsung ramping up large investments to catch up.Source: Nikkei, via Seoul Economic Daily source

Knowin (Chinese embodied-AI startup) | Angel++ round | RMB 500 million · embodied

The single round of RMB 500 million was led by Matrix Partners China, with follow-on from Pinpoint Capital, SenseTime Guoxiang Capital, HTI Group, and L2F Guangyuan Founders Fund; Guangyuan Capital served as exclusive financial advisor. Founded in 2025, the company builds consumer-grade embodied-AI large models, with a self-developed GLOW generative embodied-AI model architecture — it first generates physically consistent, structurally 3D first-person data from limited real-robot interaction to fill the data gap, then has a world model predict how an action will change the environment, feeding task outcomes back into a closed loop. The team has grown to over 200 people in under a year. The KNOWIN-X1 home robot, unveiled at this year's World AI Conference, uses a fully foldable dual-arm design that collapses to under 40 cm in height, small enough to sit on a tabletop, sofa, or washing machine; this round's funds will go toward mass-production readiness.Source: PEdaily source

Acorn Robot | Angel round | Amount undisclosed · embodied

The round was co-led by China Merchants Capital and NIO Capital, with follow-on from the Shuimu Tsinghua Alumni Seed Fund, alongside the launch of its "Natus AGE-0 embodied instinct model." This path is a non-consensus bet in an industry currently crowded around VLA: the model uses tactile perception as its core signal source, extracting the underlying rules of physical interaction from contact-mechanics principles; the company claims this enables zero-data cold start and zero-shot generalization across embodiments and materials, forgoing the large-scale data pretraining paradigm.⚠️ Vendor claim Founded in late 2024, the core technical team consists of 9 PhDs and postdocs from Tsinghua's Department of Mechanical Engineering who have worked together on this direction since 2018; commercially, the company has deliberately dropped high-margin custom integration projects to focus solely on standardized dual-arm flexible production cells, selling "manufacturing as a service" for multi-SKU, small-batch scenarios like FMCG, personal care, and food.Source: PEdaily source

Unitree | STAR Market IPO subscription | oversubscribed 8,288.82x | allotment rate 0.0181% · humanoid

Following its August 6 pricing at RMB 150.80/share and an issuance valuation of RMB 60.993 billion, subscription officially opened today. Valid online subscription accounts numbered 9.7846 million, with 53.637 billion valid shares subscribed for — an initial online oversubscription multiple of 8,288.82x, far above the 100x threshold that triggers a reallocation; the lead underwriter shifted 3.236 million shares from the offline tranche to online per the rules, leaving a final allotment rate of just 0.0181%. Each subscription lot of 500 shares requires RMB 75,400 in payment, with RMB 60,000 in Shanghai market holdings needed for a maximum-size subscription; payment for winning allotments is due August 12.Source: Sina Technology / Global Times source

Henggong Precision × AgiBot | Joint venture | Hengzhike Technology established · hardware

Corporate registry data shows Hengzhike Technology is jointly held by Henggong Weichuang Embodied Intelligence Technology, a wholly owned subsidiary of Henggong Precision, and Zhiyuan Innovation Technology, an AgiBot affiliate, with a business scope covering AI foundational resources and technology platforms and intelligent robot R&D. Henggong Precision's core business is castings and precision machining — this kind of "component maker + integrator" joint venture has become increasingly common in the humanoid supply chain.Source: Guandian.cn source

III. Commercialization & Deployment

Pony.ai's global autonomous driving mileage passes 100 million kilometers; fourth-generation L4 driverless heavy trucks enter mass production · autonomy

On August 10, Pony.ai announced that its global autonomous driving mileage has surpassed 100 million kilometers. The more concrete development is on the freight side: its fourth-generation L4-level pure-electric driverless heavy truck has entered mass production, with vehicles rolling off the line over the coming months for deployment on highway trunk routes, Northwest China bulk-commodity dedicated routes, and ports. Its L4-level driverless light truck, launched at the same time, has entered intensive road testing for express delivery, supermarket supply, and cold-chain use. Company vice president He Xing has previously said the moment for large-scale rollout of driverless trucks has arrived. Notably, per the company, cost reductions in the trucking business are mainly driven by supply-chain benefits and cost dilution from thousand-unit-scale Robotaxi production, rather than standalone breakthroughs in truck technology itself. Separately, Pony.ai and JD Auto Service are building a joint autonomous-driving maintenance ecosystem to support Robotaxi fleet operations.Source: Guandian.cn source

Hidream Driving's H1 revenue reaches RMB 804 million; driverless mining truck shipments exceed 1,900 units · autonomy

H1 2026 revenue reached RMB 804 million, up roughly 97% from RMB 408 million in the same period last year; net loss for the period narrowed 75% to RMB 113 million from RMB 455 million a year earlier, with adjusted net loss falling to RMB 19.681 million. The deployment figures say more about where the company stands than the financials do: cumulative shipments of smart-driving products topped 23,000 units as of June 30, with over 3,400 units delivered or pending delivery in the autonomous-driving business, including over 1,900 driverless mining trucks shipped across nearly 40 mines globally, up 213.57% year-on-year in H1. A 500-unit driverless mining truck purchase order signed with Guangna Group in February was the single largest formal procurement order in the global driverless mining truck sector as of end-June; the company also secured over 400 units in letter-of-intent orders for driverless heavy trucks domestically and abroad in H1, and shipped 34 mining robots.Source: Guandian.cn source

FedEx adds dual-arm robots to trailer-loading lines; Amazon to more than double its robotic-arm count this year · industrial

FedEx announced on July 30 that it is deploying Dexterity's dual-arm Mech trailer-loading system at its Hagerstown, Maryland hub, with real-time package-placement decisions made by its Foresight model. Trailer loading has long been one of the hardest links in package logistics to automate — it requires strength, endurance, and real-time adaptation, and FedEx's US network loads tens of thousands of trailers every day. Per its 2026 investor day, the company plans to expand automated loading and unloading to more than 20 US hubs and thousands of dock doors over the coming years. On the same day, Amazon's CFO said on its earnings call that the company plans to more than double its robotic-arm count in 2026, with cumulative robot deployments across its network topping 1 million units since 2012. Industry association A3 has observed that non-automotive industries have recently overtaken automotive as the largest buyers of industrial robots, with Amazon and FedEx driving that shift.Source: Supply Chain Dive source

51WORLD's physical AI core revenue up 545% year-on-year, now 41.4% of total · adjacent

H1 revenue reached about RMB 124 million, up 129.8% year-on-year, with gross margin rising from 41.1% to 44.8% and net loss for the period narrowing 25.5%. Its physical AI core business, 51Sim, saw revenue grow 545.1% year-on-year, with its share of total revenue jumping from 14.8% to 41.4%, turning it from a supplementary business into the main growth driver. The company launched two new products, AperData and AperOne, targeting the twin problems of data scarcity and low fault tolerance in embodied AI, with commercialization beginning in H2; AperOne is targeted for thousand-unit-scale real-world deployment this year.⚠️ Forward-looking claim It also launched development of ECS-1, billed as the world's first commercial remote-sensing satellite dedicated to physical AI, with Huantian Zhihui.Source: Cover News source

Palladyne AI receives additional $2.9M from the US Air Force, bringing total contract value to over $10.6M · embodied

The US Air Force exercised an option on a STRATFI contract, adding $2.9 million and bringing total contract value to over $10.6 million. The new funding is tied to a set of demonstration milestones: mixed-reality teleoperation between a human operator and a remote expert, human-in-the-loop machine learning that lets a robot autonomously learn complex manipulation tasks from a handful of demonstrations, and autonomous execution in Air Force-relevant mission scenarios. The company's Palladyne IQ platform claims to generalize from just 1 to 5 human demonstrations, requiring no cloud connectivity and being hardware-agnostic. The project began in September 2023, with its first phase completed in October 2024.Source: Auganix source

XPeng's GX model line produces its first mass-production Robotaxi · autonomy

The first mass-production Robotaxi based on XPeng's GX model has rolled off the line. Chinese automakers are taking a different path to Robotaxi than fleet operators like Waymo, going straight from their own vehicle production lines to driverless vehicles — making production-line milestones a more meaningful progress marker here than road-test mileage.Source: Autohome source

IV. Industry Developments

AgiBot overtakes Unitree as the world's largest humanoid robot maker, with about 19,100 units shipped globally in H1 · humanoid

According to a report published Sunday by San Francisco-based research firm Smart Analytics Global (SAG), as relayed by the South China Morning Post, Shanghai-based AgiBot (Zhiyuan) shipped about 8,400 humanoid robots from January to June this year, taking 44% of the global market — up 562% year-on-year — surpassing Unitree to claim the top spot globally for the first time; Hangzhou-based Unitree follows with about 5,900 units and a 31% share. The two companies' product mixes differ sharply: AgiBot has spread across multiple lines, including its full-size bipedal A-series, compact X-series, and wheeled G-series, while Unitree's humanoid shipments remain heavily concentrated in its flagship G1, with a substantial share going to education, research, and performance deployments. Zooming out, global humanoid shipments totaled about 19,100 units in H1, up 272% year-on-year, with Chinese makers supplying over 97% of that volume and Chinese buyers accounting for over 85% of demand; the more significant structural shift is that industrial and commercial applications rose from about 50% of the total a year ago to over 70%, overtaking exhibitions and performances as the main driver of volume growth. SAG projects full-year global shipments will approach 60,000 units.⚠️ Compiled estimate The financial gap between the two companies is equally stark: Unitree's IPO prospectus discloses RMB 1.7 billion in revenue and RMB 591 million in adjusted net profit last year; AgiBot has not disclosed financial details, though founder and chairman Deng Taihua said at a partner conference in April that 2025 revenue exceeded RMB 1 billion, with a 2027 target of RMB 10 billion.Source: South China Morning Post source

Beijing Municipal Key Laboratory for Embodied AI Large Models inaugurated, led by Galbot · world-model

On August 9, the Beijing Municipal Key Laboratory for Embodied AI Large Models was inaugurated in Beijing's Haidian District, holding its first academic committee meeting. It is jointly built by Galbot together with Peking University and CATL, marking a provincial/ministerial-level key laboratory recognized in China's embodied-large-model field. The lab's director is Wang He, Galbot's founder and CTO and a Peking University researcher, with the academic committee chaired by Gao Wen, a Chinese Academy of Engineering academician and director of Pengcheng Laboratory, and including 14 Chinese and international academicians and experts such as Zhang Jianwei, Zeng Wenjun, and Zhang Hongjiang, alongside scholars from Peking University, Tsinghua University, Beihang University, and the University of Science and Technology of China. Research will focus on five tracks: fundamental theory for embodied large models, multimodal data infrastructure and standardization, robot operating systems and middleware, and full-machine and key-component development, with three sequential evaluation-cycle goals — an end-to-end embodied large model with full-body and full-hand control within three years, a functionally complete municipal-level key laboratory within six years, and becoming an original-research hub within nine years. At the meeting, Wang He introduced the company's AstraBrain "large brain, small brain" foundation model system, saying that scaling laws have also been validated in motion control.⚠️ Vendor claimSource: China.com source

52 robotics funding deals in China in July; capital is shifting upstream to components · hardware

According to incomplete statistics from Robot Lecture Hall, China's robotics sector recorded 52 funding deals in July 2026 — 23 seed/angel rounds and 17 Series A-stage rounds, with early-stage deals accounting for roughly 76% of the total; by region, Shanghai led with 14 deals, followed by Guangdong with 13 and Beijing with 8.⚠️ Compiled estimate The notable shift is structural: 13 deals, or 25% of the total, went to core-component companies in July, up roughly 5.5 percentage points from about 20% in June — the fastest-growing category that month. Among those 13, tactile/force sensing and joint/transmission/actuator companies tied for the most deals with 5 each, followed by AI chip and data companies with 2. Humanoid and embodied-AI full-machine makers still led overall with 19 deals and a 36.5% share; Robot Era, Yimu Technology, and Weilai Robotics (unrelated to NIO) each closed rounds in the RMB 1 billion-plus range that month — the upstream heat isn't a sign of full-machine makers cooling off, but rather reflects mass-production expectations at the integrator level driving up demand for sensors, joint modules, and dedicated chips.Source: Robot Lecture Hall source

Hardware · Supply Chain

· Dexterous hand and joint-module capacity: Xinuo Future, founded about six months ago, has closed four funding rounds totaling roughly RMB 1.5 billion, planning for 2027 capacity of 50,000 dexterous hands and 50,000 joint modules; Tashan Technology says it holds over 80% share of the humanoid tactile-sensor segment, with orders on hand as of end-May equal to 4x its full-year 2025 revenue ⚠️ Compiled estimate source

· Xinhongye robot cables: on August 10, the company told investors on an interactive platform that its robot cable products do not currently have a supply relationship with Unitree — a reminder that supply-chain "concept stocks" riding Unitree's IPO rally need this kind of claim-by-claim verification source

· Zhongke Huisi's L1/D1/M1 dexterous hands: a joint venture backed by Lens Technology and others has launched three new dexterous-hand products (previously reported) source

FutureX · Physical AI Daily — Issue 84 (08/10)

Shawn — Sun, 09 Aug 2026 14:43:17 +0000

Today's Highlights

· Waymo filed a recall with NHTSA for nearly 4,000 robotaxis, yet not a single vehicle was taken off the road

· Lotte Mart's Busan center goes live, with 1,000 Ocado robots serving 4 million households

· Lens Technology (Chinese electronics manufacturer) teams up with CASIA Vision to launch a new company, unveiling three dexterous hands at once

· Kong Tao, former head of ByteDance's robotics team, now leads Xiaomi's robotics foundation model team

· ω-0 achieves 81.8% success across 11 real-world household tasks, nearly double the next-best baseline

· A Silicon Valley-adjacent county proposes a licensing scheme for humanoid robots, requiring emergency stop switches and continuous human supervision

· Moore Threads (Chinese GPU maker) posts H1 revenue of RMB 1.736 billion, up 147% YoY, and simultaneously announces plans for a Hong Kong listing

1. Research Progress

ω-0: a latent-space world-action model letting humanoid robots move and manipulate at the same time · locomotion

Household chores rarely offer the clean segmentation of "walk into position, then reach" — humanoid robots must shift their footing, adjust posture, maintain balance, and manipulate objects all at once. Rather than splitting the task into a navigation module plus an arm-control module, ω-0 predicts a whole-body action latent directly from language instructions, visual observations, and proprioceptive state, which is then handed to the controller for execution; the key design choice is to not generate future video, instead learning a compact future-observation embedding that supplies cues about task progress and scene evolution, with whole-body actions then generated via diffusion. Across 11 real-world household tasks, ω-0 achieves an 81.8% success rate, compared with 27.3% for π-0.5 and 22.7% for GR00T-N1.7; the strongest baseline, ψ-0, reaches only 44.5%. The team also released the ω-HOME dataset, over 40 hours of multi-view observations with whole-body action annotations.

Zhe Li et al. (MARS Lab NTU / Peking University / BAAI / HKUST Guangzhou) · arXiv 2608.06375 source

The "data pyramid" for embodied manipulation: a data-recipe survey co-authored by 11 institutions · benchmark

The field widely agrees that data is the bottleneck for embodied manipulation, but there has never been consensus on what mix of data is needed. This survey, jointly produced by 11 institutions, arranges real-robot data, UMI-style handheld capture, first-/third-person video, simulated data, and general vision-language data into a pyramid, with the central axis being the tradeoff between scalability and robot-alignment — the higher up the pyramid, the closer to real-robot data and the harder it is to scale. The survey further uses this data-recipe lens to review recent embodied foundation models, breaking down how each selects, aligns, and blends these sources during pretraining, and maps data composition onto five capability categories: perception, reasoning, planning, action generation, and world prediction.

arXiv 2607.24744 source · Analysis: blog.csdn.net source

2. Funding & Deals

Moore Threads (Chinese GPU maker) | H1 report + planned H-share listing | Revenue RMB 1.736 billion · hardware

Chinese GPU maker Moore Threads disclosed both its H1 report and Hong Kong listing plans on the same day, August 9: H1 revenue was RMB 1.736 billion, up 147.42% YoY, with net loss attributable to shareholders narrowing to RMB 11.5631 million (RMB 259 million less loss YoY, an 95.73% narrowing), and gross margin at 56.95%. The company attributed growth mainly to accelerated commercialization of its Kuae intelligent-computing clusters, naming embodied AI as one of the core scenarios for its compute infrastructure; its ultra-large-scale training cluster based on the MTT S5000 achieved 95% linear scaling, while on the edge side it launched the MTTAICUBE and E300 modules based on its "Changjiang" SoC. R&D spending was RMB 769 million, or 44.30% of revenue. Details of the Hong Kong share offering have not yet been finalized.Source: STCN (Chinese financial news outlet) source

Feiji Kesi | Series A1 | hundreds of millions of RMB · world-model

Existing investor TEDA Kechuang co-led the round with Hisound Capital and Fudan Kechuang, joined by new investors Yida, Youshi, and Yongyue, with existing investors Yunqi and Silicon Harbor increasing their stakes. Founded in 2024, the company builds physical-intelligence infrastructure — its self-developed differentiable physics engine Fysics, an embodied simulation training platform, and physical world models. It sells the "training environment and data" layer rather than robot hardware itself.Source: STCN Venture Connect source

Paxini Tech | Strategic round | RMB 1 billion | valuation over RMB 10 billion · hardware

Following disclosure of this round earlier this month, the investor list and cumulative funding total were filled in this week: the round was co-led by a consumer electronics and semiconductor giant, BOC International Investment, Kunpeng Fund, and Hexin Fangce, with participation from CIC Capital (Baifu Fund) and the Chengdu Jiaozi AI Fund, and existing investor Zhilai Capital increased its stake substantially. The company's cumulative funding has now reached RMB 3.5 billion; it focuses on the full chain from Hall-effect multi-dimensional tactile sensors to dexterous hands and complete humanoid robots, and has filed over 200 patents.Source: STCN Venture Connect source

Weekly primary-market roundup | RMB 8.459 billion disclosed | down 73.17% week-on-week · adjacent

From August 1–7, China saw 150 investment/financing deals (versus 147 the prior week), but disclosed funding fell from RMB 31.533 billion to RMB 8.459 billion — the deal count barely moved while the disclosed amount was more than halved twice over, indicating last week's high total was driven by a handful of large transactions. Advanced manufacturing led with RMB 4.164 billion, and embodied AI remained among the most sought-after sub-sectors; by funding stage, seed/angel rounds accounted for 33% of deals, while Series A led in total amount with RMB 2.506 billion.Source: STCN Venture Connect source

3. Commercialization & Deployment

Lotte Mart's Busan Zetta Smart Center goes live, with 1,000 Ocado robots taking over fresh-food fulfillment · industrial

This is South Korea's first customer fulfillment center to adopt the Ocado Smart Platform, and the first CFC built since Lotte Mart's partnership with Ocado (announced November 2022) with an investment of roughly KRW 200 billion: construction began in December 2023, was completed last August, and after integration testing was completed this July, the center began full-load operations. Across roughly 40,000 square meters, up to 1,000 picking robots ("BOT") travel across a grid at 4 meters/second, with storage stacked 8 to 21 levels high; robotic arms ("OGRP") pick from above the grid, adjusting grip strength via sensors to avoid crushing fresh produce. The chilled zone is kept at 5°C and the frozen zone at -25°C, with robots operating continuously. The operator says the entire facility is staffed by just over 300 people, with robots handling most processes aside from inbound stocking and final packing. Serving 4 million households across Busan and the Yeongnam region, daily delivery slots have increased from about 3 to as many as 13, with 2–3 hour delivery windows available, and an SKU ceiling of 35,000 — roughly 10,000 more than a typical store.Source: Chosunbiz, Asia Economy source

Ligong Industrial's full-size humanoid "Lidian" enters mass application in aerospace and auto-parts manufacturing · humanoid

This company, a nationally recognized "little giant" specialist enterprise with nearly 40 years of precision-manufacturing experience, has developed its own humanoid robot, "Lidian," with 28 degrees of freedom across its whole body and a peak joint torque of 320 N·m, and it has already achieved mass application at leading firms in aerospace and auto-parts manufacturing. Against this year's industry backdrop, where "over 70% of humanoids remain confined to showrooms and labs," a company with roots in parts manufacturing validating the robot on its own production lines first represents a different path from that of standalone humanoid startups.Source: Sohu.com source

Saite Intelligent's medical logistics robots cover over 95% of hospital material-transfer scenarios · industrial

Guangzhou-based Saite Intelligent's "Zhisaila" series now covers over 95% of hospital material-transfer scenarios, having cumulatively served over 1 million patients; its "Zhisaijie" autonomous sweeping vehicles have been deployed at Guangzhou Baiyun railway station, the Baiyun International Convention Center, and multiple stadiums, capable of centimeter-precision edge cleaning. The company holds over 700 proprietary IP rights, and R&D outcomes from the provincial-level medical service robot engineering center it established have generated over RMB 300 million in commercialized value. In-hospital medical logistics is one of the few service-robot scenarios that has already scaled, where success hinges on regulatory compliance and dispatch reliability rather than robot form factor.Source: Sohu.com source

Robot dogs start taking over security guard shifts, at $80,000–$130,000 less per shift annually than human labor · adjacent

According to Business Insider, AI camera-equipped quadruped robots are being used to patrol data centers, farmland, and stadiums, covering a 24/7 post at $80,000 to $130,000 less per year than hiring a human. Asylon, a robot-dog company, said its CEO reports about 50 robots currently deployed across 25 clients — still a tiny fraction relative to the 1.3 million security guards employed nationwide in the US. Researchers at UC Berkeley's Labor Center note that the more realistic near-term effect isn't replacement, but pushing human workers toward the harder, higher-risk portion of the job; the occupation's 2024 median annual salary was $38,390.Source: Futurism source

4. Industry Developments

Waymo files recall for nearly 4,000 robotaxis, yet not a single vehicle stops operating · autonomy

Waymo has filed a recall with NHTSA covering nearly 4,000 robotaxis, over a defect in which its driving software poorly recognized highway construction zones and ramp closures, repeatedly driving into areas that should have been closed off in San Francisco and Phoenix. Waymo's own phrasing was that it "identified room for improvement in highway construction-zone performance" — in substance, the vehicles didn't know that road conditions had changed since the last mapping update. Two signals stand out here. First, self-driving has spent a decade proving it can handle pedestrians, cyclists, and unprotected left turns, yet traffic cones plus a single flagger turn out to be harder — construction zones are scenarios that deliberately break the prior mapping assumptions the system relies on, and standardized, digitized real-time lane-closure feedback from construction crews is still unevenly adopted across the US. Second, the recall framework dating back to the 1966 Motor Vehicle Safety Act presumes an "owner": a letter is mailed, the owner visits a dealer, a part is replaced. Waymo owns and operates its own fleet, so notifying owners means notifying itself, and the fix is simply pushing a software update — meaning not a single vehicle was taken offline, and no rider even noticed; one outlet even had to issue a correction after initially reporting that the vehicles had been taken out of service. Regulatory vocabulary is being stretched to cover a fleet that can be patched overnight. Waymo has not yet opened highway driving in Texas, so riders there were not exposed to this defect.Source: The Auto Wire source

Kong Tao, former head of ByteDance robotics, now leads Xiaomi's robotics foundation model team · world-model

Multiple independent sources confirm that Kong Tao, former head of ByteDance's robotics team, now leads Xiaomi's robotics foundation model team, having joined in summer 2025 along with a number of former ByteDance colleagues. Xiaomi's robotics division now has about 200 people, working across hardware, motor control ("cerebellum"), cognition ("brain"), and even robot operating systems, with different teams sometimes running parallel research on the same direction (such as world models); the foundation model team has its own separate office space. Kong Tao received his PhD from Tsinghua University's computer science department in 2019 and was ByteDance's first employee focused specifically on robotics research; his personal homepage states that as of June 2025 his robotics-related papers had been cited over 15,000 times, and he left ByteDance in June 2024. Xiaomi released Xiaomi-Robotics-0 in February and Xiaomi-Robotics-1 in July this year, with the former sharing the "brain + cerebellum" paradigm used in GR-3, which Kong Tao worked on during his time at ByteDance. Talent flow hasn't been one-directional either: a number of people from Xiaomi's CyberOne and CyberDog 2 teams later joined ByteDance Seed, filling exactly the hardware and motion-control gaps there.Source: 21st Century Business Herald source

Lens Technology (Chinese electronics manufacturer) teams up with CASIA Vision to found Zhongke Huisi, launching three dexterous hands · hardware

Zhongke Huisi Embodied Intelligence (Hunan) was jointly founded in July 2026 by Beijing CASIA Vision, Lens Technology, and Hunan Huaxia Investment Group, and made its debut in Changsha on August 7 with the launch of three dexterous hands — L1, D1, and M1: the L1 takes a lightweight, reduced-DOF approach with 6 active degrees of freedom at a price in the thousands-of-RMB range, positioned as an upgrade replacement for existing grippers, and has already been used to have the CASBOT BAND robot band play guitar, bass, and electronic keyboard; the D1 is a high-DOF general-purpose five-finger hand dedicated to real-robot data collection; the M1 uses a modular "standardized body + functional fingertips" architecture aimed at industrial scenarios. The company also pre-announced a tendon-driven, bio-inspired F-series. Based in Hunan's Xiangjiang New Area, the company says it will build the country's largest dexterous-hand skill training ground and data-collection center. Its strategy is a dual-flywheel model of "multiple robot platforms × data collection" — first lowering the barrier to actually putting these end-effector "hands" to use through a series of hardware platforms, then feeding the resulting data back to train skill models.Source: Sohu.com, STCN source

Silicon Valley-adjacent county proposes licensing scheme for humanoid robots · humanoid

San Mateo County Supervisor Ray Mueller will submit a proposal to the county Board of Supervisors next week requiring a licensing process for autonomous/semi-autonomous humanoid robots deployed in commercial settings such as restaurants and retail. Cited reasons include the need for specialized firefighting equipment against battery thermal-runaway fires, unpredictable robot behavior potentially obstructing evacuations, and pressure on the tax base and social services from job losses in the service sector. Applicants would need to demonstrate that their robots comply with ADA and health standards, are equipped with an emergency stop switch, have no outstanding recalls or history of catastrophic failure, and are continuously supervised on-site by trained personnel during operation; licensing fees would go toward firefighting equipment procurement. Mueller expects state or federal legislation to eventually take over the matter, but argues the county should act first. This is among the first local initiatives to treat humanoid robots as a distinct regulatory subject, following a period when regulatory attention was concentrated mainly on self-driving cars.Source: Palo Alto Daily Post source

BYD's first bipedal humanoid "Xiaodi" debuts in Zhengzhou this month, entering through dealerships rather than factories · humanoid

BYD confirmed it will showcase its first bipedal humanoid robot, "Xiaodi," this month at the Di Space brand experience center in Zhengzhou, where it will interact with customers and introduce vehicle models. BYD Executive Vice President and Americas CEO Stella Li said the goal is "two or three units per dealership," and believes robot sales assistants could become commercially viable within one to two years. According to media reports, the robot stands 1.61 meters tall, weighs 58.5 kg, and has 31 degrees of freedom (7 per arm, 6 per leg), with 1mm hand positioning precision and a 1kg payload — specs better suited to a controlled showroom environment than industrial work.⚠️ Manufacturer's own claimsCompared with Hyundai's approach of betting on mass-producing Atlas at a US factory by 2028, BYD is prioritizing retail service before scaling toward industrial deployment — read by some as reflecting that Chinese factory-floor scenarios are already claimed by players like Unitree, UBTECH, and AgiBot.Source: The Korea Herald source

Turnover in AgiBot's partner roster ahead of IPO; questions remain over Luo Jianlan's status · embodied

Following earlier reports this week that AgiBot's (Chinese humanoid robot startup) official partner roster grew to 9 people, with Chief Scientist Luo Jianlan absent from the list, more of the timeline has since been filled in: on July 24, the same day AgiBot confirmed it had initiated its Hong Kong listing process, Shanghai's Institute for Advanced Study in AI released a video stating that Luo Jianlan had joined the institute as a full-time mentor; references to his AgiBot role on both his personal homepage and X bio have since been removed. AgiBot's response was that he has "not yet left; this is an internal role adjustment," without disclosing a new title.⚠️ Single-party accountThe three newly added partners — Xiong Yan, Zhu Jie, and Zhang Xiuzheng — come from Huawei's Central Research Institute, Tencent Public Affairs, and Huawei Cloud's China region respectively, all with industry rather than research backgrounds; together with founder Deng Taihua, nearly half of the 9 partners now have Huawei backgrounds. Luo Jianlan, a student of Sergey Levine, led the open-sourcing during his tenure of AgiBot World, a dataset of over 1 million real-robot manipulation trajectories, along with the τ0-WM and τ0-VLA models, and at June's Beijing Academy of Artificial Intelligence conference bluntly stated that "chasing leaderboards is meaningless."Source: Caijing Tianxia WEEKLY source

MirroS proposes Physical RSI: moving recursive self-improvement from code sandboxes into the physical world · world-model

Recursive self-improvement (RSI) in the digital world — where an AI edits its own code, runs experiments, and retains changes that work, in a closed loop — carries an implicit assumption: that the environment is known and fixed. MirroS's proposed Physical RSI argues that in the physical world, this must instead be a dual-axis coupling, with the Actor and World Model iterating in tandem: when something unexpected happens, the system first checks whether its world model can reproduce the causal chain of the failure (an "Environment Gap"); only once it can does the system move on to searching for a new strategy (a "Skill Gap"). A related claim is that world models should be built from language abstractions plus executable code rather than end-to-end latent representations, on the grounds that when a black-box model fails, its errors can't be localized, making "improvement" meaningless.⚠️ Single-party accountThis is currently a framework and research roadmap, with no public experimental results yet.Source: Sohu.com source

Guangzhou Baiyun industrial robotics cluster named a provincial specialty industrial cluster · industrial

The cluster currently comprises 75 companies, spanning robot bodies, core components, system integration, and AI large-model and application services; 2025 output value exceeded RMB 24.7 billion, with an average annual growth rate of 18.85% over the past three years, and exports of leading products rose from RMB 3.789 billion in 2024 to RMB 6.34 billion. The cluster includes 1 nationally recognized single-item manufacturing champion and 11 nationally recognized "little giant" specialist enterprises. In 2025, the cluster's R&D spending was RMB 1.238 billion, an R&D intensity of 5.01%, with local supply-chain coverage rising to 73%.Source: Sohu.com source

Hardware & Supply Chain

· Dexterous hand shipments: According to GGII data, China's dexterous hand sales in 2025 were approximately 19,200 units, up 236.84% YoY, projected to reach 70,200 units in 2026, up 265.63%; CASIA Vision's Zhang Zhengtao estimates the industry is growing at roughly 3.7x annually based on this trajectory source

· Dexterous hand cost and lifespan: Zhang Zhengtao says market prices for high-DOF tactile dexterous hands range from the low hundreds of thousands of RMB to RMB 800,000–1,000,000+, and that "Tesla's Optimus hand still costs $6,000," but under continuous industrial use, actual lifespan is only a few weeks to two or three months source

· "Hands without brains" and the data gap: He also notes that no VLA model to date can genuinely drive a 20–30 DOF hand to do real work; open-source data for complete robot systems already includes millions of trajectories, while open-source data for fine dexterous-hand manipulation "amounts to only tens of thousands — several orders of magnitude behind" source

· Foundation Robotics' tendon-driven hand: Motors are positioned in the forearm and pull tendons to open/close each joint, keeping the fingers slim and fast-moving; finger position is inferred from motor rotation angle and cable routing, with magnetic encoders at each joint serving as an angular backup — the hand has already demonstrated catching a thrown baseball in mid-air, though the throw's trajectory was pre-planned and there is no tactile feedback yet; it will be used in the company's factory-oriented Phantom humanoid source

FutureX · Physical AI Daily — Issue 83 (08/09)

Shawn — Sat, 08 Aug 2026 14:45:45 +0000

Today's Highlights

· 116,000 new companies registered in the humanoid robotics sector in H1, up 9.5% YoY

· Per China's Ministry of Industry and Information Technology (MIIT): humanoid shipments in China totaled roughly 23,000 units in H1, with over 70% still confined to showrooms and labs

· Unitree opens subscription on August 10, with nearly half of proceeds earmarked for embodied "brain" R&D

· Waymo adds four cities in one move, having previously operated in 10 cities

· Nvidia open-sources Cosmos 3 world model family, with the largest weights (64B) available for commercial use

· A humanoid robot designated "0755" starts work at a Shenzhen elder-care facility, whose welfare center has opened 21 application scenarios

· A US robotics startup hand-carries Chinese-made components in suitcases to sidestep customs declarations

I. Funding & Deals

Unitree Robotics | STAR Market IPO | Subscription opens August 10 | Issuance valuation nearly RMB 61 billion · humanoid

After setting its issue price at RMB 150.80/share, Unitree Robotics will open retail subscription on August 10, issuing 40.4464 million shares (10% of post-issuance total share capital), with expected total proceeds of RMB 6.099 billion — above the RMB 4.201 billion planned in the prospectus. The use-of-proceeds breakdown disclosed at the August 7 roadshow is notable: nearly half the raised funds will go toward smart robot model R&D, with the rest allocated to hardware R&D, new product development, and manufacturing base construction — this company, which built its reputation on hardware cost-performance, is putting its largest bet on closing the "brain" gap. Founder Wang Xingxing said the company is pursuing both the WMA and VLA technical routes. The financials reveal a gross margin trajectory for the first time: humanoid robot gross margin fell from 87.67% to 63.18% between 2023 and 2025, while quadruped margin rose from 43.71% to 56.72% — the company attributes this to deliberate price cuts to drive volume, offset by in-house core component development and scaled procurement (the base G1 starts at RMB 85,000, the R1 Air at RMB 29,900). Per the prospectus, humanoid shipments exceeded 5,500 units in 2025, ranking first globally, while cumulative quadruped sales from 2023-2025 topped 33,000 units. Source: China Securities Journal source

AgiBot (Chinese humanoid startup) | Hong Kong IPO | Cornerstone target valuation HKD 40-50 billion · humanoid ⚠️ single-source account

The earlier-reported disappearance of Chief Scientist Luo Jianlan from the partner list continues to draw attention during the IPO window; today's new information is a comparison of valuation and shipment figures. At an HKD 40-50 billion valuation, AgiBot's price-to-sales ratio comes to roughly 32-41x, notably above Unitree's approximate 24x and UBTech's approximate 22x — while AgiBot posted approximately RMB 1.05 billion in 2025 revenue and remains unprofitable, versus Unitree's RMB 1.699 billion revenue and RMB 278 million net profit attributable to parent over the same period. The "shipment leader" claim is also disputed: an Omdia report states AgiBot shipped 5,168 general-purpose humanoid units in 2025, or 39% of the global total, while Unitree states its actual pure-humanoid shipments exceeded 5,500 units and cautions against "combining different robot types for comparison." IDC's more granular data shows that in the core category of full-size bipedal units over 1.4 meters, AgiBot shipped only about 1,300 units in 2025. The backdrop is China's embodied intelligence sector raising a total of RMB 93.5 billion in H1, up 137% YoY — the fight over valuation anchors has now landed directly on statistical methodology. Source: Guancha source

Xizhi Technology | Series A | Over RMB 100 million · adjacent

Tongge Venture Capital and Xinhe Capital participated, with existing investor Hangzhou Capital adding to its stake; the company's cumulative 2026 funding has now exceeded RMB 300 million, and it is simultaneously launching a Series B. Proceeds go toward automotive power module production lines at its Hangzhou plant and compute power module production lines at its Suzhou plant, with a core business in high-frequency power electronics and wide-bandgap semiconductors — as robotics and autonomous driving compute demand scales up, Chinese domestic power-supply-side capacity is an easily overlooked but essential supporting line. Source: Gasgoo source

II. Commercialization & Deployment

Waymo adds four cities — San Diego, Las Vegas, Denver, and Tampa launch together · autonomy

Waymo announced service expansion into San Diego, Las Vegas, Denver, and Tampa, having previously operated in 10 cities. As is standard practice, new cities initially open only to Alphabet employees before gradually opening to the public, so this marks the start of a rollout schedule rather than four cities going live for public riders immediately. The notable timing: Zoox just announced it is converting Las Vegas to paid rides, and Waymo is following into the same city — the US robotaxi city race has shifted from single-city pilots to parallel multi-city buildout.Source: Investor's Business Daily source

Humanoid unit "0755" joins Shenzhen elder-care facility as welfare center opens 21 application scenarios at once · world-model

AgiBot (Chinese humanoid startup) partnered with the Shenzhen Municipal Social Welfare Service Guidance Center to deploy a humanoid robot designated Unit 0755 at a Shenzhen elder-care facility, where it handles daily patrols, item delivery, and voice companionship duties; the facility primarily serves Shenzhen household-registered elderly with severe or higher-level disability, under strict care standards. More notable is the supply side of scenarios: the center has now formally opened its first batch of 21 AI and embodied-intelligence application scenarios, spanning child welfare, social placement, elder care, and medical services. AgiBot previously partnered with home-services platform "58 Daojia" to launch in-home robotic housekeeping services in Shenzhen and Beijing, along with a one-month home-residency program. Caregiving is a low-fault-tolerance scenario, and current tasks remain limited to basic actions like patrolling and delivery — still a distance from fine-grained care work. Source: 21st Century Business Herald source

Xino Future dexterous hand lands 10,000-unit order; Flex 2 packs 23 degrees of freedom in one hand · hardware

Following a RMB 500 million Series A+ round in July that valued the company at over $1 billion post-money, this Hangzhou-based company — founded in December 2024 — has now disclosed cumulative orders of 10,000 units for its dexterous hand — a rare scale at a time when most dexterous hands remain stuck at prototype and small-batch delivery stages. Its Flex 2 uses a cable-driven plus direct-drive hybrid scheme: actuators are placed in the forearm with tendon cables pulling the finger joints to preserve grip strength, while micro-motors in the palm handle millimeter-level fingertip adjustments; the palm weighs under 400 grams yet delivers 23 degrees of freedom, with a peak single-hand grasp load of 12 kg, sustained load of 4 kg, repeat positioning accuracy of ±0.1mm, and force control precision of 0.05 N. The motors, motor control, reducers, ball screws, and algorithms are all developed in-house. Source: Hangzhou.com.cn source

Hertz's Oro Mobility surpasses 6 million cumulative miles, set to join Uber's robotaxi network this year · autonomy

Fleet management operations at Oro Mobility, a Hertz-affiliated operating company, have surpassed 6 million miles in cumulative mileage; it currently operates in four markets under a "human-driven managed fleet" model, maintaining and operating vehicles for rideshare drivers. Its first autonomous-driving partnership will be with the Uber robotaxi program in the San Francisco Bay Area, supporting Lucid vehicles equipped with Nuro self-driving technology, expected to launch sometime in 2026. What often limits robotaxi scaling isn't the algorithm but fleet operations — rental car companies are turning cleaning, charging, and dispatch into a new point of entry.Source: Pulse 2.0 source

Star Era robots deployed at over a dozen SF Express and China Post logistics centers · humanoid ⚠️ vendor-reported

On August 7, a "Robotics Development Watch Beijing" press tour visited Star Era (Xingdong Jiyuan) in Haidian District; the company's robots have been "put to work" at more than a dozen SF Express and China Post logistics centers, handling package sorting and packing. The company states that operational efficiency in some scenarios already exceeds that of humans — this claim comes from the company itself, with no third-party measurement cited, so it should be treated as vendor-reported for now. Source: Sina News, The Beijing News source source

III. Industry Developments

SAMR: 116,000 new companies registered in humanoid robotics in H1 · humanoid

China's State Administration for Market Regulation (SAMR) announced on August 8 that 116,000 new companies were registered in the humanoid robotics sector in H1 2026, up 9.5% YoY; over the same period, 55,000 new companies were registered in generative AI, up 28.0% YoY. Nationally, new company registrations across the "8 emerging industries + 9 future industries" totaled 561,000. By comparison, the humanoid sector's company growth rate now clearly trails generative AI's — while total registrations hold at the hundred-thousand level, the marginal trend is shifting from a registration boom toward consolidation of existing entrants. Source: Xinhua News Agency (via Jingbao) source

MIIT: China's humanoid production tops 30,000 units in H1, with over 70% still in showrooms and labs · humanoid

China's Ministry of Industry and Information Technology (MIIT) reports that domestic humanoid robot production in China exceeded 30,000 units in H1 2026, with shipments of approximately 23,000 units — but among shipped units, over 70% are deployed in showrooms and labs, with only a small share actually deployed on production lines. This figure is worth reading alongside the day's funding and IPO numbers: shipment volume is increasingly used as valuation validation, but it currently measures purchase intent more than it does production-line replacement. Source: Guancha source

Nvidia open-sources Cosmos 3 world model family, largest weights (64B) available for commercial use · world-model

Nvidia released Cosmos 3, an open world model family for physical AI, comprising Super (64 billion parameters), Nano (16 billion), and Edge (4 billion) tiers, with the Edge version targeting on-device deployment and running on hardware such as RTX, DGX, and Jetson Thor. Weights are released under the Linux Foundation's OpenMDW 1.1 license, allowing developers to fine-tune with their own data and compute. Unlike previous task-specific model splits, Cosmos 3 folds scene understanding, synthetic data generation, and action prediction into a single family — usable as a vision-language model, as a world simulator predicting future states, or as a foundation for world action models; companion tools Omniverse and OpenUSD handle simulation environments and sensor asset management. Nvidia also claims top rankings on Artificial Analysis' open-weight image-to-video, PAI-Bench world generation, Physics-IQ image-to-video, and RoboLab robot policy leaderboards — these are vendor-self-reported leaderboard results, with no independent reproduction yet. Source: IT Brief Australia source

US robotics startups hand-carry Chinese-made components home in suitcases · hardware

Following last month's FCC restriction, citing national security, on imports of foreign-made humanoid and quadruped robots (previously reported), a supply-chain-side reality has surfaced: employees and investors at several US humanoid startups are directly packing motion control, power management, and signal transmission components into suitcases to carry back to the US, sidestepping the weeks-long formal customs process and tariff costs. Data from Germany's Mercator Institute for China Studies shows that 63% of core suppliers in the global humanoid robot component supply chain are based in China; the ban currently does not cover components, but startups will be directly hit once they move from prototype to mass production, unless final assembly happens in the US. K-Scale Labs, a now-defunct Palo Alto company, repeatedly traveled to China to hand-carry back dozens of actuators, and sourced sensors hard to find in the US via Taobao. Apptronik CEO Jeff Cardenas said the US-China difference comes down to supply chain geography, stating "we want to build America's own 'Shenzhen.'" The article also notes that some vendors print fake battery labels to evade air-cargo lithium battery regulations — itself illegal. Source: cnBeta source

Zhongke Huisi debuts with three dexterous hands at once, previews 37-DOF cable-driven biomimetic hand · hardware

Zhongke Huisi, a joint venture of Zhongke Huiling, Lens Technology (Chinese electronics component maker), and Hunan Huaxia Investment Group, launched three dexterous hands — L1, D1, and M1 — and previewed a cable-driven biomimetic hand from its F series. The L1 weighs about 610 grams with 11 degrees of freedom, opens and closes the full hand in just 0.2 seconds, and can lift a 25 kg load (roughly 40x its own weight); pricing will be kept in the low thousands of RMB range. The D1 weighs under 900 grams with 20 degrees of freedom and piezoresistive tactile sensors at the fingertips; the M1 uses a modular quick-swap approach with over a dozen functional fingertip options. The previewed F series reaches 37 degrees of freedom, with 7 in the thumb alone, and addresses cable-drive's biggest maintenance pain point with a single-tendon-cable design allowing cable replacement in as fast as 3 minutes. Lower price points and maintenance-focused design speak more to dexterous hands shifting from display pieces to consumables than the degree-of-freedom figures do. Source: Robot Outlook (via Ifeng Tech) source

D-Robotics wraps DGP DemoDay, over 40 institutions review 10 teams onsite · adjacent

On August 7, D-Robotics held its "Gravity DemoDay 2026" in Shanghai, drawing over 40 institutions including Hillhouse Ventures, Shunwei Capital, BAI Capital, and Jiuhe Venture Capital. 10 teams were selected from the DGP ecosystem's more than 200 member companies to present, with Aula Wanxiang, Xunling Technology, and Tiaoyue Technology taking the top three spots. Projects spanned personal mobility, exoskeleton hiking gear, home embodied intelligence, low-cost open-source humanoids, and jumping robots, with an average founder age under 30. DGP says it has enabled over 500 robotics startups cumulatively, and its platform now connects over 100,000 developers globally. Investors onsite focused on who would buy, why they'd buy, and whether the pricing holds up — AI capability alone doesn't constitute a reason to purchase. Source: Sohu, Ifeng Tech source

Hardware & Supply Chain

· Robot Battery Joint Lab: China Electronics Standardization Institute (under MIIT) signed an agreement with JD.com and Sunwoda on August 7 to jointly build a lab focused on standards development, product development, and testing/validation for robot batteries; the institute has led drafting China's only current robot battery standard, and two standards discussion meetings were held immediately after signing, with over 60 industry experts attending. JD.com's aim is to standardize battery dimensions, interfaces, and communication protocols to lower R&D and maintenance costs source

· Wanma Co. · robot cabling: In an investor briefing, the company said its specialty smart-equipment cabling is already mature in industrial robot applications and has entered the humanoid robot and robot-dog space, with robot-dog products already in mass production; however, it explicitly noted that robot cabling revenue currently makes up a small share of overall revenue source

FutureX · Physical AI Daily — Issue 82 (08/08)

Shawn — Fri, 07 Aug 2026 14:56:47 +0000

Today's Highlights

· Zoox begins charging for rides in Las Vegas starting next week, cleared to deploy up to 5,000 steering-wheel-free robotaxis within two years

· FCC's robot restrictions draw the line by manufacturing location, not manufacturer nationality — industrial robots explicitly exempted

· Shipbuilding giant HII awards two robotics companies up to $900 million in seven-year production contracts

· Tencent open-sources three embodied foundation models; its VLA is already in testing on a household-goods production line with a success rate above 95%

· Brain Corp's global deployment surpasses 50,000 units, with H1 growth up 68% year-over-year

· Kodiak's driverless truck fleet grows to 35 units, cumulative paid operating hours exceed 40,000

· Unitree (Chinese humanoid robotics maker) allocates RMB 271 million in IPO shares to 171 executives and core employees, signs strategic cooperation memorandum with DeepSeek

· Mech-Mind (Chinese industrial vision/robotics firm) clears Hong Kong IPO filing, cumulative funding exceeds RMB 2 billion, with Meituan, Sequoia, and Intel Capital among its shareholders

I. Research Progress

ω-0: Letting humanoid robots work while they walk, instead of walking then working · locomotion

Nearly all existing humanoid policies split locomotion and manipulation into two separate execution phases, but real household chores require both to happen simultaneously — moving, adjusting posture, maintaining balance, and manipulating objects form one continuous behavior. ω-0 is a whole-body latent predictive world-action model that takes language instructions, visual observations, and proprioceptive state as input and directly outputs whole-body action latents executable by the controller; rather than reconstructing future video, it learns compact future-observation embeddings as a lightweight prediction target. The team also released ω-HOME, a real household humanoid dataset of over 40 hours, containing multi-view synchronized observations, whole-body SMPL motion, and action latents. Across 11 real-robot household tasks, the single model consistently outperforms imitation learning, VLA, humanoid, and world-action model baselines.

Zhe Li et al. · arXiv 2608.06375 source

DyPES-VLA: Cross-embodiment generalist policies no longer need hand-aligned action spaces · vla

One bottleneck in training a generalist VLA that can handle heterogeneous robots is that each robot's action format is mutually incompatible, requiring extensive manual preprocessing to force a unified format. DyPES-VLA sidesteps this in two steps: first, it trains a VLM with a future-prediction objective so the shared representation captures object motion, contact, and interaction-induced scene changes; then it uses an embodiment-specific MoE action head to translate this shared dynamics prior directly into each embodiment's native action space. As a single generalist policy, it achieves a 98.0% success rate on LIBERO, 59.25% on RoboCasa-GR1, and 89.02% on RoboTwin 2.0.

Junfeng Li et al. · arXiv 2608.06374 source · HF trending 15↑

GST-Bench: VLMs can understand a single room, but can't piece together a whole building · benchmark

Existing spatial benchmarks mostly test local perception from a single frame or a few viewpoints, sidestepping the capability embodied intelligence truly needs — maintaining a global map across a continuous, long-duration video stream. GST-Bench builds human-verified VQA questions from 6,790 minutes of synthetic video, requiring models to make spatial inferences from viewpoints not present in the input video and to map first-person observations onto a global top-down view. The best zero-shot score among 22 SOTA models is only 42.68, versus 79.08 for humans; the companion GST-Bench-Local shows models perform well on the local version of the same question type, indicating the failure point is an inability to integrate long-duration observations into a globally consistent scene representation.

Qifeng Zhang et al. · arXiv 2608.05747 source · HF trending 28↑

Hijacking a robot with a single sheet of paper: a systematic measurement of physical prompt injection against VLMs · benchmark

When a VLM is used as a robot planner, adversarial text placed within the camera's field of view becomes an entry point for indirect prompt injection. The authors categorize attacks into four types (indirect identification, task redefinition, authority impersonation, and conflict injection), building a benchmark of 20 attack prompts and running 5,670 trials across 3 scene layouts × 3 instruction phrasings: GPT-4o's attack success rate was 27.0%, Gemini 2.5 Flash's 29.4%, and Qwen3-VL-32B's 5.0%, with authority-impersonation and negation-type attacks transferring across all three models. Reasoning traces show that 99.9% of successful compromises were the model "knowingly" complying; three simple defenses (prompt defense, two-stage verification, and preprocessing text masking) reduced risk to 75-100%, 85-100%, and 100% respectively, at the cost of potentially impairing tasks that require reading in-scene labels.

S. M. Bhagya P. Samarakoon et al. · arXiv 2608.05715 source

GAUGE: Measuring physics engines against real trajectories — none is comprehensively faithful · benchmark

The physical fidelity of simulation engines and video world models has mostly been judged by perceptual similarity or human evaluation, without pinpointing exactly which physical law is being violated. GAUGE uses 22 controlled task families (rigid bodies, flexible cables, fabric, volumetrically deformable objects) with calibrated physical metadata and uncertainty annotations to put Isaac Sim, Genesis, and Newton on the same measuring stick. The conclusion: no engine is comprehensively faithful, with the largest deviations occurring in impact contact, fast fabric motion, and volumetric deformation; 6 image-to-video generation models expose a different kind of distortion — the generated trajectories have the right equation form, but acceleration, momentum transfer, and oscillation timing are all wrong.

Shuai Wang et al. · arXiv 2608.05948 source

JoyAI-RA 0.5: Treating human first-person video as the main axis for scaling · vla

Human first-person video is the easiest data to scale but the furthest from robot data, and directly mixing it into training pools often causes negative transfer. This framework addresses this with "dual action alignment": implicit alignment infers latent actions from visual transitions, letting action-label-free human, simulated, and real-robot data all drive the world model's learning of physical dynamics; explicit alignment unifies reliable human-robot trajectories into a shared camera frame, landing them in the same action space via a chunk-relative end-effector pose representation. On the AgiBot real-robot benchmark, task scores continuously improve as the volume of human first-person pretraining data increases — the paper claims no saturation is observed even at its largest scale, a self-reported scaling trend that awaits independent replication.

RA Team · arXiv 2608.05674 source

W2-VLA: Getting the model to think through "what will the wrist do next" · manipulation

Most VLAs treat the main view and wrist view as parallel inputs, overlooking that the two play different roles: what's actually critical for fine manipulation is anticipating how local wrist contact will evolve within the global task context. W2-VLA uses a set of latent tokens as the interface between the VLM and a wrist predictor, predicting future wrist latents and converting them into look-ahead context for action prediction; a companion W2-CoT pipeline generates structured annotations describing manipulation progress, physical turning-point cues, and wrist evidence as auxiliary supervision. On LIBERO, RoboTwin 2.0, and real-robot tasks, it improves contact-sensitive manipulation while maintaining an action generation frequency above 80 Hz.

Yuhao Pan et al. · arXiv 2608.05369 source · HF trending 14↑

XEWorld: World models are actually still just 2D image matchers · world-model

If action-conditioned world models are only evaluated on the robots used in training, it's impossible to tell whether they've learned physical dynamics or merely memorized visual patterns. XEWorld isolates the embodiment variable, evaluating held-out robots in physically identical scenes, and the conclusion is fairly unforgiving: current models' generalization is dominated by visual similarity rather than kinematic similarity, and they struggle to translate abstract joint-value actions into coherent visual trajectories; zero-shot rendering of an unseen embodiment must rely on strong grounding cues like pixel-space actions and explicit spatiotemporal alignment, while forcing this threshold with few-shot adaptation triggers catastrophic forgetting of previously seen embodiments.

Yixiang Chen et al. · arXiv 2608.05799 source

Other papers today: HOPE estimates hand-object contact pressure from monocular video, placing pressure predictions directly on hand mesh vertices to free itself from dependence on object shape or sensor layout (arXiv 2608.06192 source); near-sensor computing turns visuotactile reconstruction into a fully streaming hardware pipeline, with power consumption around 347 mW and a fixed per-frame first-depth-value latency of 0.211 ms, compressing a robot's protective reflex loop from 169.9 ms to 28.3 ms (arXiv 2608.05725 source); KILVO fuses joint encoders, IMU, LiDAR, and camera for humanoid robots via an asynchronous-sequential hybrid ESIKF odometry (arXiv 2608.05647 source); TRACE uses end-to-end learned proprioceptive odometry to handle unreliable contact conditions in legged robots (arXiv 2608.05975 source); GeniWorld decouples embodiment kinematics from environment dynamics, generating diverse manipulation trajectories within the world model to feed back into policy learning (arXiv 2608.06332 source); Adaptive-WAM lets a video-diffusion driving model dynamically decide how many layers to run and when to skip future video synthesis based on trajectory quality (arXiv 2608.06008 source); In-Context VLA argues that what VLAs need is not generating language but consuming grounded language (arXiv 2608.05738 source); IcFuzz is the first fuzzing method targeting Isaac Sim (arXiv 2608.06088 source).

Open Source · Tools · Benchmarks

· NVIDIA Cosmos 3: A next-generation open-weight physical AI model family launches, positioned to let machines understand and predict what will happen in the real world, not just describe what a camera sees; in July, Nvidia co-signed an open letter on "open weights and U.S. AI leadership" with 200+ organizations. Claims of topping benchmarks come from the vendor's own reporting source

· Tier IV × Cosmos: Japanese autonomous driving company Tier IV discloses its use of NVIDIA Cosmos to build the data foundation for its autonomous driving datasets — a rare third-party deployment disclosure for the Cosmos series source

· RLWRLD dexterous hand benchmark: South Korean physical AI company RLWRLD partners with three domestic robotics and simulation firms to build a benchmark for humanoid robot hand dexterity source

· MoWorld-3D: A newly released 3D world model positioned as a middle layer to push AI from "conversational interaction" to "physical interaction" source

II. Funding and Deals

Unitree (Chinese humanoid robotics maker) | STAR Market IPO strategic placement | Employee asset management plans, RMB 271 million · humanoid

Following yesterday's confirmed issue price of RMB 150.80/share and an issuance valuation of RMB 60.993 billion, placement details landed today: Employee Asset Management Plans No. 1 and No. 2 were allocated 1.3561 million and 444,300 shares respectively, corresponding to roughly RMB 204 million and RMB 67 million, totaling approximately RMB 271 million, with lock-up periods of 12 and 36 months respectively, covering 161 and 10 participants — 171 senior executives and core employees in total. This offering of 40.4464 million shares represents 10% of post-issuance total share capital, at an issuance P/E of 219.23x, expected to raise approximately RMB 6.099 billion; founder Wang Xingxing's direct and indirect stake will be diluted from 33.3583% to approximately 30.02%. On the same day, DeepSeek was allocated 933,900 shares worth RMB 140.8 million and also signed a strategic cooperation memorandum with Unitree aimed at jointly developing AI models for humanoid robots — previously disclosed details covered only the placement and 36-month lock-up. Wang Xingxing, speaking at an online investor exchange session, called the listing "a brand-new starting point" and attributed the decline in humanoid robot gross margins to multiple factors.Source: National Business Daily source, Guandian.cn source

Mech-Mind (Chinese industrial vision/robotics firm) | Hong Kong IPO filing approved | Cumulative funding exceeds RMB 2 billion · industrial

Mech-Mind (Xiong'an) Robotics Technology received IPO filing approval today, preparing to list on the Hong Kong Stock Exchange, with a shareholder roster including IDG Capital, Meituan, Sequoia China, Source Code Capital, Intel Capital, and Qiming Venture Partners, with cumulative funding exceeding RMB 2 billion. Founded in 2016 by Shao Tianlan, the company's core business is industrial robotics solutions centered on AI and 3D vision, having built a standardized "eye-brain-hand" product line combining "3D cameras + proprietary AI algorithms + software platform + dexterous hands." Its deployments at scale cover workpiece loading/unloading, depalletizing/palletizing, high-precision assembly, defect inspection, and welding, serving over 100 Fortune 500 clients globally across nearly 50 countries and regions. Amid a wave of humanoid full-stack robot makers filing for listings, this is a rarer candidate approaching embodied intelligence from the industrial vision side.Source: LeiDD source

X Square Robot | Planned Hong Kong IPO | Post-money valuation RMB 20 billion · embodied ⚠️ unconfirmed reports

Multiple outlets report that X Square Robot has confidentially filed for a Hong Kong listing, though the company has made no official response and no searchable filing exists at the exchange. Founded in late 2023, this company is the only Chinese robotics firm to have secured investment from all four of Meituan, Alibaba, ByteDance, and Xiaomi, with each taking turns leading Series A, A+, A++, and B rounds; in late June it announced completing B+, B++, and C rounds within just over two months, pushing its valuation to RMB 20 billion. On the technical side, it has pursued a generalist rather than task-specific model from day one, developing a proprietary unified architecture called WUM; on deployment, it partnered with 58 Daojia (Chinese home services platform) in June, putting robots into real households in Shenzhen to work alongside human cleaning staff.Source: Chaodian Robotics source

Ommo Technologies | Series A | Tens of millions of USD · adjacent

Hong Kong's VMS Group co-led the round with an unnamed fund, with Kanghao Capital participating and Dianshi Capital serving as exclusive financial advisor. Ommo works on permanent-magnet field positioning — generating a distinctive magnetic field via mechanical rotation of a permanent magnet, sidestepping the two major weaknesses of optical positioning (fails when occluded) and traditional electromagnetic positioning (vulnerable to metal interference, requires bulky coils), achieving sub-millimeter precision within a defined range, with the smallest magnetic sensor at about 0.8 mm — narrower than a grain of rice — and small enough to embed in a robot end-effector or medical device. The technology was first validated in surgical navigation scenarios and is already used by over a hundred medical device companies in China and abroad; this year it pivoted toward embodied intelligence, with its first data-collection glove used to capture human hand manipulation postures, targeting the industry-recognized scarcity of 3D physical manipulation data.Source: 36Kr source

DEEP Robotics | STAR Market IPO | Review status updated to "inquiry stage" · humanoid

According to the Shanghai Stock Exchange's website, DEEP Robotics' (Chinese quadruped and humanoid robot maker) STAR Market IPO review status has been updated from "accepted" to "inquiry stage." Following its filing in late July, this quadruped and humanoid robot maker has entered the exchange's substantive review process.Source: Beijing News source

Zhuxian Technology | Second Hong Kong exchange filing · autonomy

Zhuxian Technology (Chinese L4 trunk-logistics autonomous driving company) has filed again with the Hong Kong Stock Exchange. Filings in this sector have been coming in dense clusters recently, but commercialization is still stuck in the "valley of death" between demonstration operations and profitable scale — on the same day, Pony.ai also announced its move into autonomous light trucks, seeking a second monetization path for its robotaxi technology.Source: Ebrun source

Vangrid | Seed token round | $9 million · adjacent

DePIN project Vangrid completed a $9 million token round funding round to build a decentralized spatial-data network for physical AI. The idea is to use distributed incentives to address the high cost and sparse coverage of 3D spatial data collection — an alternative approach on the supply side of embodied data.Source: The Block source

III. Commercialization and Deployment

Zoox gets NHTSA clearance, begins charging for rides in Las Vegas next week · autonomy

Amazon-owned Zoox received approval last week from the U.S. National Highway Traffic Safety Administration to deploy up to 5,000 steering-wheel-free custom robotaxis within the next two years, with permission to charge passengers; the company confirmed it will start charging fares in Las Vegas next week. This marks Zoox's transition from long-running free test rides to commercial operation — it has already carried over 1 million passengers and driven over 3 million miles across Las Vegas, San Francisco, Austin, and Miami without being able to charge fares. The significance is twofold: first, a custom, steering-wheel-free vehicle has for the first time received federal approval for paid deployment at scale; second, Waymo now has its first comparably sized competitor. According to data from mobile app tracking firm Apptopia, Zoox's monthly active user share rose from 15% to 25% between January and June 2026, while Waymo's fell from 79% to 69% over the same period (third-party figures, not disclosed by the operators themselves).Source: CleanTechnica source

Shipbuilding giant HII awards two robotics companies up to $900 million in seven-year production contracts · industrial

HII, the largest U.S. military shipbuilder, has signed long-term performance-based production agreements with Path Robotics and GrayMatter Robotics, planning to issue up to $900 million in shipbuilding work to the two companies over seven years, covering aircraft carriers, submarines, destroyers, amphibious ships, future frigates, and unmanned surface vessel programs. The agreement has two phases: a development phase to jointly research and certify high-precision processes such as autonomous welding, grinding, sandblasting, spraying, assembly, and inspection, integrating them into an autonomous production line; and a delivery phase in which HII begins purchasing from the two companies based on cost, schedule, and quality performance, starting with small steel structures and expanding to sections and modules. This is not an upfront order but a long-term demand signal conditioned on technology and manufacturing readiness milestones — it's precisely because of this signal that the two companies feel confident making long-cycle investments in robotics, facilities, and workforce. The broader context: HII plans to outsource over 2.5 million hours of shipbuilding work in 2026, a 30% increase from 2025.Source: Marine News source

Brain Corp's global deployment surpasses 50,000 units, H1 growth up 68% year-over-year · industrial

Robot operating system maker Brain Corp announced that robots running BrainOS have surpassed 50,000 units deployed globally, with H1 2026 global deployment growth up 68% year-over-year; over the same period, the fleet logged over 5.3 million hours of autonomous operation (up 24% year-over-year), covering more than 3.7 billion square feet — 400 million more square feet of autonomous coverage than the same period last year. These robots are deployed across retail, logistics, airports, commercial cleaning, and inventory management. CEO David Pinn's framing is notable: enterprise buyers are no longer evaluating whether a single robot can complete an isolated task, but whether an autonomous platform can deliver predictable performance, secure data handling, and unified fleet management across hundreds or thousands of sites — the company's recent SOC 2 Type II compliance certification was earned precisely to meet that procurement bar.Source: Robotics & Automation News source

Kodiak adds 7 driverless trucks in Q2, fleet reaches 35 units, launches Gen7 · autonomy

Kodiak AI added 7 new driverless trucks in Q2, bringing customer-owned fleet size to 35 units, and launched its seventh-generation Kodiak Driver platform — with nearly 50% more compute than the prior generation, using a more compact modular design to support lower-cost daycab configurations; its autonomy readiness metric rose to 91% in July. Operationally, cumulative paid operating hours passed 40,000 (up 71% from the end of Q1), with over 300,000 tons of freight hauled in the quarter. Financials, however, show this remains a cash-burning phase: Q2 revenue was $3.5 million, up 91% quarter-over-quarter, but operating cash burn was $34.1 million and free cash outflow was $38.1 million, with $151.1 million in cash and securities at quarter-end. The company's goal is to achieve its first driverless long-haul run by year-end.Source: Automotive World source

Guangzhou Post uses 8 embodied robots for parcel sorting, up to 1,200 parcels/hour · embodied

China Post's Guangzhou district center in Guangdong has been deploying 8 embodied robots at its Jianggao road-transport export center for automated parcel-sorting tests since March 2026; after multiple rounds of data training and on-site tuning, peak sorting efficiency has reached 1,200 parcels/hour. The planned roadmap is more concrete than most pilots: raising efficiency to 1,600 parcels/hour by Q2 2027, and completing full data connectivity and coordinated operation between humanoid robots and sorting machines by Q3. It remains a pilot stage and has not yet moved into standard production.Source: China News Service source

UBTECH lands humanoid robot order worth over RMB 50 million from an AI foundation-model company · humanoid

UBTECH announced an order worth over RMB 50 million for humanoid robots from an AI foundation-model company. With the buyer being a foundation-model maker rather than a traditional industrial customer, this points to a full-robot procurement use case for data collection and model training — a demand line that is becoming a sizable cash-flow source for full-stack humanoid makers this year.Source: Futu Niuniu source

Tate boosts per-welder output 12x with 58 collaborative welding robots · industrial

Data-center infrastructure maker Tate has deployed 58 Hirebotics Cobot Welder collaborative welding systems across factories in Arkansas, Virginia, and Kentucky, raising per-welder output on critical structural components 12-fold, while its team of certified welders has actually expanded. This 120-year-old company decided in 2024 to move from manual welding to a cloud-connected automation network, using the no-code Beacon Pro platform to program, run, and monitor its cross-plant fleet from a phone or tablet. The Arkansas plant has also implemented a distinctive process: collaborative robots handle upstream spot-welding and pre-assembly precision steps, feeding parts to downstream industrial robots to keep the line's pace steady.Source: Robotics & Automation News source

IV. Industry Developments

FCC clarifies scope of robot restrictions: line drawn by manufacturing location, industrial robots explicitly exempt · adjacent

The International Federation of Robotics (IFR) issued a fact sheet today clarifying the boundaries of the FCC's July 28 addition of "foreign-manufactured advanced robotics equipment" to its Covered List — previously, the market only knew that new models from Chinese manufacturers like Unitree might be unable to be sold in the U.S.; the actual scope is much broader. Three key points: first, the restriction is based on where a product is manufactured, not the manufacturer's nationality, so companies headquartered in the U.S., Europe, or Asia are equally covered if they manufacture outside the U.S.; second, coverage applies to mobile ground robots capable of autonomous operation without an operator, equipped with sensors and network connectivity — including service robots, AMRs, humanoids, quadrupeds, and even some consumer robots — while articulated, Cartesian, SCARA, and parallel-kinematic industrial robots are explicitly excluded; third, the rule is not retroactive — products already FCC-certified may continue to be imported, sold, and used, and software/firmware updates that maintain function and cybersecurity are also permitted. The FCC also offers a conditional-approval pathway, but applicants must submit extensive business information including equity structure, investors, supply chain, manufacturing arrangements, software/firmware architecture, and cybersecurity controls. In effect, what looked like a ban targeting a single country has become a repricing of manufacturing-location footprint across the global robotics industry.Source: International Federation of Robotics (IFR) source

Tencent open-sources three embodied foundation models, splitting the robot into three "brains" running at different frequencies · world-model

Tencent formally open-sourced three embodied foundation models at WAIC 2026: Hy-Embodied-VLM-1.0 (spatial and scene understanding), Hy-Embodied-RxBrain-1.0 (cognitive planning and future-state imagination), and Hy-Embodied-VLA-0.5 (high-level goals to continuous actions). Chief scientist and Robotics X Lab director Zhang Zhengyou identifies an underdiscussed pain point in current VLA design: cognition, perception, and execution are forced to share the same computational pipeline running at the same frequency, yet physical-world time scales vary enormously — tactile feedback operates at around 1 millisecond, while low-level motion control runs at 500Hz to 1000Hz. Tencent's approach lets the upper cognitive system wake on demand, the mid-level perception-action system continuously ingest multimodal information at about 15Hz, and the lower-level execution system handle collisions and loss of balance like a reflex at higher frequency, paired with its proprietary agent framework TairosAgent to compress task response time to 2–3 seconds. RxBrain's differentiator is that it doesn't just generate task steps in language — it also predicts what the physical world "should look like" after each subtask is completed, folding visual state into the cognitive pathway. The hardest data point is the deployment result: Hy-Embodied-VLA has entered a household-goods factory for production testing, achieving a success rate above 95% on a high-mix, low-volume production line with frequent SKU changes, with a cycle time under 6 seconds per unit and less than 3 days allotted for data collection and post-training when a new SKU is introduced. Zhang Zhengyou's assessment of the industry's current state is blunt: "A demo scoring 80 or 90 out of 100 that never gets deployed is basically worth zero." The claim that the model ranks first overall on the RoboDojo benchmark comes from Tencent's own reporting.Source: InfoQ source

MagicLab breaks ground on 10,000-unit production line in Wuxi, with quadrupeds making up the vast majority of capacity · humanoid ⚠️ planning figures

MagicLab (Chinese humanoid and quadruped robotics maker) broke ground on its headquarters factory in Wuxi's Liangxi Science City, planning 4 joint-module production lines and 2 full-robot assembly lines. What's worth noting is the capacity mix — according to Gasgoo, of the designed annual capacity of 10,000 units at full production, 9,000 will be small quadrupeds and 1,000 large quadrupeds, which doesn't match the intuitive read of a "10,000-unit humanoid production line" that headlines might suggest. Founded in early 2024, this company unveiled a full-size humanoid, MagicBot X1, an industrial wheeled humanoid, MagicBot D1, and a light-industrial quadruped, MagicDog T1, all at WAIC 2026 in one go, and released its native world model Magic-Mix in April. Across the industry, companies founded three to four years ago that have genuinely crossed the 10,000-unit production threshold remain in the single digits, with most second-tier players still at planning stages in the thousands — capacity ramp-up, supply chain, and cost all remain unresolved. The groundbreaking is a fact; hitting full production is a target.Source: Gasgoo source

RhOS-World: Khora — Khora claims to scale co-present agents in a world model to 1,024 · world-model ⚠️ vendor claim

RhOS.ai and Ophilus.AI released a world-model architecture called RhOS-World: Khora, along with a playable online multiplayer 3D shooter demo — with visuals, collisions, and hit feedback reportedly all rendered in real time by an AI world model, with no traditional physics-engine code involved. Previous world models were largely confined to a "single-observer viewpoint," where multi-viewpoint consistency either collapsed or its compute cost grew exponentially; Khora's approach introduces a shared spatiotemporal blackboard called STBoard, shared by all agents, which records no pixels but only maintains the absolute physical state of a 4D world — agents each write local changes and read back the world state to render their own viewpoint, without needing to communicate directly with each other. The team says this decouples model structure from the number of viewpoints, allowing a model trained on data from 10 agents to generalize with zero fine-tuning to 1,024. If this capability holds up, it points toward multiple robotic arms on a flexible production line sharing a common physical-spatiotemporal intuition; but what's currently public is still a game-environment demo, with no third-party replication or real-robot results yet.Source: 36Kr source

Waymo co-CEO publicly questions the vision-only approach · autonomy ⚠️ conference remarks

Waymo's co-CEO, speaking in a keynote, directly criticized Tesla's camera-only robotaxi approach, arguing it faces major obstacles to achieving true driverless operation. Around the same time, Tesla narrowed its Robotaxi operating hours in Austin to end at 10pm, and a separate incident where a vehicle stopped at a stop sign was rear-ended has been filed with the NHTSA. The multi-sensor versus vision-only debate has run for years, and the operational data from each side still doesn't constitute directly comparable evidence.Source: 36Kr source

South Korean army rolls out robots and unmanned equipment across its entire military logistics chain · industrial

South Korea's army announced the introduction of robots and unmanned equipment across its entire logistics supply chain, aimed at speeding up nationwide supply turnover. This aligns with a broader thread of recent South Korean government moves: a prior proposal to deploy 1,000 AI robots annually and develop dedicated humanoids for 10 major industries, and Daegu's establishment of the country's first humanoid robot safety certification center. Military logistics is one of the rare scenarios that demands extremely high reliability while being willing to bear early-stage costs at scale — its procurement standards often spill over to become a reliability baseline for civilian products.Source: Chosunbiz source

BAIC Group establishes Yuanqi Physical Intelligence Technology company, business scope includes robotics R&D · embodied

BAIC Group has established a new company, Yuanqi Physical Intelligence Technology, with a business scope covering robotics R&D. Automakers setting up standalone physical-intelligence entities has become a routine move this year — carmakers hold mass-production manufacturing capability, supply-chain bargaining power, and accumulated autonomous-driving perception stacks all at once, making them one of the few player types in humanoid full-stack competition that doesn't need to build these capabilities from scratch.Source: D1EV source

Hardware · Supply Chain

· Paxini (Chinese tactile-sensing company) PX-FOOTRIX: launches the world's first multi-dimensional plantar tactile sensor, based on a proprietary 6D Hall-array tactile technology, achieving full-area 3D array force sensing and 6-axis force/torque sensing on the sole, letting bipedal robots adjust gait in real time based on force changes on complex terrain; it will debut at the 2026 World Robot Conference. The company recently raised another RMB 1 billion in a strategic round, bringing cumulative funding to RMB 3.5 billion source

· Weihong Corporation: shipped over 50,000 dexterous-hand joint motors in H1 — dexterous hands are among the hardest components in a full humanoid robot to bring down in cost and up in yield, and this volume suggests the joint-motor segment has entered mass production ahead of the full robot itself source

· Ligong Leike "Shanhai" series: launches an intelligent computing module for edge AI scenarios, built on Chinese-made CPUs and GPUs, targeting robot on-board compute — a segment that has historically been heavily import-dependent source

FutureX · Physical AI Daily — Issue 81 (08/07)

Shawn — Thu, 06 Aug 2026 14:59:11 +0000

Today's Highlights

· Unitree Robotics set its IPO price at RMB 150.80/share, implying an issue-date market cap of RMB 60.993 billion

· DeepSeek received a strategic placement allocation of over RMB 140 million in Unitree's IPO, with a 36-month lock-up

· Hadrian raised nearly $1.4 billion, valuation quadrupling in seven months to $7.9 billion

· LexisNexis's first humanoid patent report: China holds 73% of body-structure patents

· South Korea plans to deploy 1,000 AI robots per year, building purpose-built humanoids for 10 major industries

· GigaDevice (Chinese chipmaker) launched two robot MCUs in succession, with roughly 3 million units shipped for robotics applications in H1

· Nucleus emerged from stealth: deployed a humanoid robot into a German factory within 90 days

· DreamWAM lifted real-robot unseen-scenario success rate from 55.6% to 74.4%

I. Research Progress

DreamWAM: World Action Models Shouldn't Just Predict RGB · world-model

World Action Models (which learn action representations by predicting how the environment will evolve) typically predict the future in RGB pixel space, where task-relevant state changes get entangled with texture, lighting, background, and viewpoint noise. DreamWAM instead jointly trains four complementary future representations — appearance, motion, geometry, and semantics — while all these auxiliary branches are switched off at inference, keeping deployment purely RGB-based. The gains show up mainly out-of-distribution: under LIBERO-Plus perturbations, success rises from 51.36% to 63.44%, and on a real robot facing unseen lighting, backgrounds, and object placements, average success reaches 74.4%, versus 55.6% for the Fast-WAM-Joint baseline from the same group. Code and models are open-sourced.

Shanglin Yuan et al. (Huazhong University of Science and Technology, hustvl) · arXiv 2608.04996 source

MobileWAM: Moving World Action Models from Tabletops to Mobile Manipulation · world-model

Prior WAMs have been almost entirely confined to tabletop pick-and-place, while mobile manipulation demands simultaneous locomotion and whole-body manipulation amid scene-scale dynamic changes — something still largely handled with dynamics-free vision encoders plus hand-written coordination logic. MobileWAM uses a mixture-of-transformers to jointly attend a pretrained video diffusion transformer with a lightweight action expert layer by layer, while splitting the feed-forward layers into shared, locomotion, and manipulation experts that are softly routed by action intent. A training-time Chain-of-Foresight mechanism has intermediate representations predict future latent chunks progressively; at deployment the entire foresight chain and video generation are discarded, leaving only policy-level overhead. It outperforms existing mobile manipulation policies on ManiSkill-HAB and was fine-tuned and validated on a real ARX Lift2 robot.

Zehua Fan et al. · arXiv 2608.04657 source

BridgeVLA++: Giving 3D VLA a Spatiotemporal Memory · vla

Existing 3D VLA models are data-hungry, generalize poorly under distribution shift, and lack any explicit memory of past observations — making memory-dependent tasks (e.g., recalling where an object was placed after being occluded) largely infeasible. Built on top of BridgeVLA's original multi-view projection and heatmap intermediate representation, BridgeVLA++ adds a unified spatiotemporal memory module, achieving SOTA on two memory-dependent manipulation benchmarks without sacrificing the original version's data efficiency and generalization, and validated in bimanual settings and on an additional real-robot platform. Same-day HF trending score 7↑.

Peiyan Li et al. · arXiv 2608.05042 source

SiMDex: More Human Egocentric Video Isn't Always Better — Selection Matters · manipulation

Scaling human egocentric video for robot training is now conventional wisdom, but which portion of that data actually helps dexterous manipulation has remained unclear. SiMDex treats human-data selection for VLA post-training as a recommendation problem, using a retrieve-rank-rerank pipeline to select similar clips for each robot demonstration from a pool of roughly 32 million egocentric samples, operating entirely in an embodiment-agnostic action space without modifying the VLA architecture. Against a strong baseline using an equal amount of randomly sampled data, it uses only about 1.49 million clips (under 5% of the sample pool), yet raises overall success rate from 47.7% to 61.1%.

Nie Lin et al. · arXiv 2608.04196 source

Mind-VLA: Spatial Alignment Should Target the Object the Instruction Names · vla

Aligning representations with 3D scene geometry is a common way to improve VLA generalization, but such methods align the entire scene uniformly, overlooking exactly the geometry of the target object named by the language instruction — causing failures in fine manipulation and when the target is occluded. Mind-VLA first localizes the instruction-specified target, then aligns the VLA's latent representation with VAE and VGGT features from three views of that target. Using a compact 345M-parameter backbone, it achieves 93.9% on LIBERO and 4.47 on CALVIN; on real-robot occlusion tasks, average success reaches 54%, 32 percentage points above the best instruction-agnostic comparison method.

Xingyu Ding et al. · arXiv 2608.04633 source

Tactus: Open-Vocabulary Recognition Even from the Cheapest Pressure Arrays · perception

Tactile representation learning has for years centered on optical sensors that image gel deformation, while resistive pressure arrays — the cheapest and highest-volume sensor type — have been comparatively neglected. Tactus answers text queries using only pressure data, achieving a top-1 accuracy of 0.771±0.062 (four runs) on the STAG benchmark (27 object categories), matching and in the best case exceeding the 0.76 achieved by a supervised closed-set CNN on the same dataset, without training a classification head. The recipe is thoroughly small-data: 187 training recordings and 144,000 unlabeled same-sensor frames for masked-autoencoder pretraining, with the sensor's built-in calibration affine transform contributing more gain than all architectural changes combined. The paper also reports its failures — cross-sensor pretraining pooling yields no benefit, and joint training with vision actually hurts tactile performance; weights and code are open-sourced.

Abdul Basit Tonmoy et al. · arXiv 2608.04043 source

Deltoris: Real-Time Inference for Diffusion VLAs via Algorithm-Hardware Co-Design · vla

Diffusion-based VLAs deliver the best action quality and generalization, but at the cost of requiring 50–200 Hz control frequencies, imposing hard latency and energy constraints on edge devices. Deltoris exploits the temporal similarity between adjacent inputs, computing only the difference between consecutive frames to eliminate redundant bit-level operations, then amortizes data loading across multiple control steps via speculative inference, paired with a dedicated accelerator using a 1D systolic bit-serial PE array. It delivers up to a 34.2× speedup relative to a mobile GPU, with accuracy essentially unchanged.

Zheng Liu et al. · arXiv 2608.04428 source

PRIMAL3: Pushing Multi-Agent Pathfinding to City-Scale, 100,000 Agents · autonomy

Multi-agent pathfinding failures cluster at topological choke points — bottlenecks, dead ends, and locations of repeated conflict — where decisive coordination is essential. PRIMAL3 constructs features from cut vertices, dead-end regions, shortest-path distances, and blockage estimates, modeling same-direction following and opposite-direction conflict via separate graphs; during training, policy entropy identifies uncertain agents, LaCAM3 intervenes on trigger and supplies label-smoothed imitation targets, and LaCAM3 is no longer needed at inference. The approach scales to city-level deployments of 100,000 agents and was feasibility-tested on a real robot system.

Chengyang He et al. · arXiv 2608.04905 source

Talk2Sensors: First 3D Visual Grounding Dataset Including 4D mmWave Radar · perception

3D visual grounding (identifying a target in 3D space based on a language description) has largely been confined to indoor settings; outdoor extensions have mostly relied on monocular images alone, even though in real outdoor perception, cameras, LiDAR, and 4D radar each capture complementary texture, geometry, and motion properties. Talk2Sensors is the first multi-sensor 3DVG dataset built on camera + LiDAR + 4D radar, containing 8,682 language instructions and 20,558 referenced targets, with prompts deliberately aligned to physical cues specific to each sensor. The accompanying TSFormer dynamically routes appearance, geometry, and motion cues based on query semantics, outperforming the strongest baseline by 8.05 mAP on this dataset.

Runwei Guan et al. · arXiv 2608.04568 source

Other papers today: Faster-WAM, computing future representations once and sparsely reusing them throughout action denoising (arXiv 2608.04404 source); SAFECAST, improving training and calibration of VLA latent-state risk probes via contrastive-set perturbations (arXiv 2608.04246 source); GUARD, detecting failures by measuring diffusion VLAs' visual-language grounding degree via ablation, without modifying the pretrained policy (arXiv 2608.04510 source); CofactVLA, using counterfactual intervention to resolve VLA's "vision overriding language" confound (arXiv 2608.04396 source); Retrieve in Time, Correct in Frequency, a training-free test-time correction framework requiring no parameter updates or repeated inference (arXiv 2608.04527 source); RORA, end-to-end reconstruction of articulated simulation assets from a single static object video (arXiv 2608.04842 source); SpikingNav, a spiking neural policy for embodied navigation with lower parameter count and per-step compute, more robust under visual corruption (arXiv 2608.05078 source); EventKitchen, a large-scale stereo event-camera benchmark of kitchen human activities (arXiv 2608.04865 source); SSC, converting long-horizon manipulation demonstration sub-task annotations into automatically verifiable state-transition chains (arXiv 2608.04425 source).

Open Source · Tools · Benchmarks

· MAGI-2 Preview: Sand.ai open-sourced a MoE unified audio-video generation model with roughly 114B total parameters but only about 6B activated per forward pass, releasing weights, code, and training system together; it ranks 6th globally on the Artificial Analysis text-to-video leaderboard — video generation backbones are precisely the foundation underpinning today's world action models source

· Orbbec (Chinese 3D vision company) Robot-Free Data Collection Platform: Debuted at ROSCon JP 2026 in Japan, covering egocentric capture, hand-object interaction, and near-field wrist capture via its EGO, UMI, and WristCam lines; the demonstrated Dual-Ego prototype uses dual 2-megapixel fisheye cameras, sub-1ms sync precision, a six-axis IMU up to 1000 Hz, and weighs 200 grams total source

II. Funding and Deals

Unitree Robotics (Chinese humanoid maker) | STAR Market IPO Pricing | RMB 150.80/share | RMB 60.993 billion issue-date market cap · humanoid

Following the launch of preliminary book-building on August 5 (when market estimates put the price around RMB 104/share and market cap above RMB 40 billion), Unitree Robotics set its final IPO price at RMB 150.80/share, implying a listing market cap of about RMB 60.993 billion — notably above book-building estimates. The offering covers 40.446434 million shares, or 10% of post-issue total share capital, expected to raise roughly RMB 6.099 billion gross, or about RMB 5.917 billion net of issuance costs. The company proactively flagged valuation risk in its filing: the price implies a diluted P/E of 219.23× based on the lower of 2025 pre- or post-non-recurring-items earnings, and a static P/S of 35.89×, both above comparable industry levels. Total strategic placement shares are 8.089286 million; Hangzhou DeepSeek, controlled by Liang Wenfeng, received a placement allocation of over RMB 140 million with a 36-month lock-up, grouped alongside Tencent-affiliated Shanghai Qishan Investment under the category "strategic partnership or long-term cooperation intent with the issuer"; also in this tier are China National Petroleum Corporation's Kunlun Capital, China Southern Power Grid's industrial-financial holding arm, and China Telecom Capital. The company explained that the purpose of bringing in these investors is joint R&D focused on large AI models and embodied intelligence. Unitree also disclosed expected H1 2026 revenue of RMB 1.052–1.128 billion and net profit attributable to parent of RMB 258–306 million; subscription opens August 10.Sources: Beijing Daily App source; Sina Finance source

Hadrian (US) | New Round | Nearly $1.4 Billion | $7.9 Billion Valuation · industrial

Baillie Gifford and J.P. Morgan's Strategic Investments Group led the round, with Washington Harbour Partners and others participating. Hadrian builds highly automated factories for defense and aerospace customers; the valuation is more than quadruple its level in January this year, up from just $260 million in last year's round (which included Founders Fund). CEO Chris Power said the software platform driving its factories, Opus, is already in use by the US Army and Navy, and that the new funding will target capacity bottlenecks in the submarine, munitions, and drone industrial base. This round lands amid a dense stretch of defense-tech fundraising — Shield AI raised at a $12.7 billion valuation in March, autonomous-vessel maker Saronic at $9.25 billion, and Anduril's valuation more than doubled to over $60 billion two months later.Source: CNBC source

Avatar Robotics (US) | Seed Round | $6.5 Million · embodied

AlleyCorp led the round; the earlier pre-seed round was led by defy.vc, with Headline and others participating. The company builds warehouse humanoids around "remote human teleoperation plus progressive autonomy": robots perform picking, packing, kitting, sorting, and inventory tasks on customer sites while an operator remotely takes over from elsewhere, continuously generating data that feeds back into autonomous policy. Since launching in December 2025, its deployed robots have handled the packing and sorting of over 900,000 items, with customers including a top global beauty retailer, and it has entered a post-pilot expansion phase with a warehouse operator worth several billion dollars.Source: PR Newswire source

Strutt (Chinese robotics startup) | Pre-A+ Round | Nearly $100 Million Raised Cumulatively · autonomy

Founder Hong Xiaoping was formerly head of DJI's optoelectronics division and previously built the world's first mass-produced automotive-grade LiDAR. Its product, the Strutt EV1, takes the form of a wheelchair but is built around an L2-level autonomous driving core: two LiDARs, ten ToF sensors, six ultrasonic sensors, and two cameras, on an automotive-grade four-wheel-drive chassis with independent suspension on all four wheels and ABS, top speed 13 km/h and 32 km range. Overseas pricing is $7,499, with a planned Chinese price of RMB 29,999; overseas mass production and delivery have begun, with entry into the Chinese market planned for H2 2026. The company says the return rate on its first delivery batch was zero, and the test-drive-to-paid-purchase conversion rate is nearly 50%.⚠️ Single-party accountSource: Robotics Lecture Hall source

Shelfmark (Pittsburgh, US) | Seed Round | $3.5 Million | ~$5 Million Raised Cumulatively · industrial

Armory Square Ventures led the round, with Grand Ventures, Hyde Park Angels, and others participating. The company provides in-line quality inspection for continuous-flow manufacturing — production lines making industrial film, webbing, paper, flooring, and coiled metal at high speed and volume, where defects often aren't caught until large batches of material are already wasted. Its platform packages in-line industrial cameras, spatial sensing, and a proprietary deep-learning vision model into a managed service, and causally correlates defects with environmental variables like temperature, humidity, and pressure; the company says its customer deployments achieve 99.5% defect-detection accuracy and roughly a 50% reduction in manual inspection labor costs.⚠️ Vendor accountSource: Pulse 2.0 source

III. Commercialization and Deployment

Japan's First Hospital-Validated Commercial Humanoid Goes on Sale, Built on a Unitree G1 Chassis · humanoid

Tokyo AI company ZEALS's subsidiary Omakase Robotics launched D1 for sale on August 5 at ¥5 million (about $32,000) per unit — Japan's first humanoid robot to pass a hospital proof-of-concept and go on public sale. The design reflects compliance-driven tradeoffs throughout: standing 129.3 cm tall with a base roughly 48 cm wide, sized to pass through standard door frames, fit into elevators, and navigate narrow hospital corridors without building modifications; it uses a wheeled base rather than bipedal legs — per Omakase CTO Shuichi Nagao's remarks at the Tokyo Humanoids Summit this past May, bipedal humanoids currently cannot be sold in Japan for safety and regulatory reasons; each arm joint has a torque sensor that stops the robot instantly on contact with a person or object; battery life is about 8 hours, matching a standard work shift. Notably, it runs on a Unitree G1 chassis, a platform that was added to the US FCC's Covered List on security grounds just 8 days before D1's commercial launch — what a "quasi-domestic" product means when its supply chain still runs through a Chinese manufacturer is a question now facing the Japanese hospitals being asked to sign deployment agreements.Source: Tech Times source

Video Action Models Enter an Audi Plant: Robots Begin Learning Directly from Watching Humans · embodied

The video action model FLUX-mimic, released jointly by German company mimic robotics and Black Forest Labs, has entered an Audi factory environment for application validation. The approach is to understand human operating motions from video and translate them into executable robot behaviors, targeting assembly, loading/unloading, flexible grasping, and quality inspection tasks that aren't fully structured — tasks that traditionally require engineers to repeatedly demonstrate and hand-write rules, and to redo the tuning whenever the model or station changes. Hard constraints for industrial deployment remain unsolved: factories need deterministic, traceable actions with safety margins, and the model can't just perform well on average — force feedback, simulation validation, and safety controllers are all still required.⚠️ Application validation stageSource: AGV.com.cn (reprinted via Sina Finance) source

Zoox's Steering-Wheel-Free Robotaxi Begins Paid Rides in Las Vegas August 10 · autonomy

Following last week's approval of the first US commercial exemption for a steering-wheel-free robotaxi (previously reported), Zoox has set August 10 as the date it begins charging for rides, starting in Las Vegas. Since opening free rides across its test cities in September 2025, it has accumulated over 500,000 rides, with a waitlist of comparable size. Miami and Austin are its next named destinations.Source: TechCrunch

PIA Automation Builds E-Drive Assembly Line for BMW's Steyr Plant · industrial

The project is located at BMW Group's Steyr plant in Austria, spanning two floors totaling over 5,600 square meters, with 78 processing stations, 27 robotic cells, and 46 industrial robots — one of PIA's largest projects to date. The line first assembles the EV transmission, then "marries" it with the motor in a robotic cell where 12 synchronized electric screwdrivers tighten it in one pass into a complete E-Drive unit; automatic gripper changeover lets different transmission variants switch without manual intervention, supporting single-piece batches. The entire process is validated in advance using digital twins and virtual commissioning of axes, sensors, material flow, and on-site equipment behavior; the final sections of the line are scheduled to enter commissioning in 2027.Source: Robotics & Automation News source

Nucleus Emerges from Stealth: Humanoid Robot Into a German Factory in 90 Days · humanoid

European robotics startup Nucleus came out of stealth on August 5, saying it deployed a humanoid robot to the site of a leading German industrial manufacturer in under 90 days; the robot performs factory tasks under human supervision while accumulating data to feed subsequent autonomy development. The founding team comes from CERN, the European Space Agency, 1X, NEURA Robotics, and Agile Robots. The customer's name, robot unit count, and task scope were not disclosed.⚠️ Single-party accountSource: Interesting Engineering source

IV. Industry Developments

GigaDevice (Chinese chipmaker) Launches Two Robot-Specific MCUs, Bringing EtherCAT On-Chip · hardware

GigaDevice released the GD32H77R robot-specific chip and the GD32F50MxxG highly integrated motor control chip, targeting humanoid joint modules/servo drives and lightweight servo joints respectively. The GD32H77R is a 600 MHz Arm Cortex-M7 with 640 KB of tightly coupled memory running at CPU frequency for zero-wait-state instruction fetch; it integrates an EtherCAT slave controller and dual 100 Mbps Ethernet PHYs directly on-chip, eliminating two external industrial-grade PHY chips, achieving one-way transmission latency as low as under 50 ns and DC synchronization cycle precision of 62.5 μs; the company says dynamic power consumption is only 20%–50% that of comparable-architecture M7 competitors. The GD32F50MxxG integrates a proprietary three-phase gate driver and a four-channel op-amp into the same package, eliminating the need for external standalone drivers and sampling op-amps to simplify BOM; the 8mm×8mm QFN80 variant enters mass production this December, and the 7mm×7mm BGA100 variant ships in volume next March. A company marketing representative said humanoid-robot-related MCU shipments reached roughly 3 million units in H1 this year, with volumes expected to grow further in H2. Shares closed up 8.26% at RMB 385.44 that day, with a market cap of RMB 270.481 billion.⚠️ Vendor accountSource: ChipInsight source

South Korean Government Sets Physical AI Direction: Deploying 1,000 AI Robots Per Year · industrial

South Korean Deputy Prime Minister and Minister of Economy and Finance Koo Yun-cheol chaired an emergency economic headquarters meeting and an economic structural reform ministerial meeting on August 6, announcing the "K-Asset Innovation Plan" and economic restructuring direction. The robotics-related portion includes: advancing AI process innovation for core industries like steel, petrochemicals, and shipbuilding, developing purpose-built humanoid robots for 10 industries, and pushing an annual deployment scale of 1,000 AI robots, aiming to gain first-mover advantage in physical AI. The same package includes a comprehensive overhaul of the 1950 State Property Act — historically focused on real estate — into a "Basic National Asset Act" that expands the scope of the already-1,402.7-trillion-won national asset base to include securities equity, intellectual property such as patents, and virtual assets.Source: Seoul Economic Daily source

Pony.ai Gives Robotruck Mass-Production Timeline · autonomy

Pony.ai targets deploying 500–1,000 unmanned heavy trucks over the next 2–3 years, with unmanned light trucks aiming for a fleet of 100,000 vehicles before 2030. Verifiable facts: the trucking business contributed nearly half of the company's revenue in 2025; the company obtained China's first nationwide permit in April this year for driverless-following operation in autonomous truck platooning. The company says its robotaxi, unmanned light-truck, and heavy-truck businesses share over 95% synergy in driverless hardware, policy frameworks, and supply chains, and that unmanned pure-electric light trucks can cut per-kilometer freight costs by 40%–50%. The industry backdrop is a driver shortage — the "2025 China Road Freight Industry Talent Survey Report" says the number of truck drivers nationwide fell by roughly 3.8 million from 2019 to July 2025, with those under 40 accounting for less than 10%.⚠️ Plan-stage figuresSource: Caijing Tianxia WEEKLY (reprinted via Sina Finance) source

LG Chairman to Meet Jensen Huang Again in Silicon Valley Next Week, Agenda Points to Robotics Cooperation · adjacent

According to multiple South Korean outlets, LG Group Chairman Koo Kwang-mo will travel to Silicon Valley next week to meet with Nvidia CEO Jensen Huang, discussing concrete cooperation plans in AI infrastructure and robotics — about two months after the two met in Seoul this past June. Around the same time, Nvidia is also deepening AI and robotics cooperation with Hyundai Motor Group. Details and outcomes of the meeting have not been announced.Source: Yonhap News Agency source

DOBOT Unveils Consumer Humanoid DOBOT LUMO · humanoid

DOBOT unveiled a humanoid robot, LUMO, marketed around "embodied all-terrain" capability, standing nearly 1.3 meters tall; its promotional video shows continuous walking across grass, sand, and cobblestone terrain, plus outdoor sparring, home companionship, and dance-mode switching. The company positions it as a consumer-grade companion robot and has not disclosed a mass-production timeline or pricing. The showcased content is a demo video, not a mass-production or scaled-delivery metric.⚠️ Vendor accountSource: 36Kr source

Hardware · Supply Chain

· PaXini PX-FOOTRIX: Released a multi-dimensional plantar tactile sensor based on 6D Hall-array tactile sensing technology, providing full-area 3D array force sensing and six-axis force/torque sensing on the sole of the foot, supporting customizable IP67/68 protection with claimed 1000% shock overload tolerance, and compatible with SPI/UART/RS485/CAN FD interfaces; the company calls it "the world's first" and says it will debut at WRC 2026 ⚠️ Vendor account source

· Molex Hybrid Connector: Launched a compact hybrid connector for humanoid robots and industrial automation, addressing the longstanding problem of limited wiring space inside joints source

· GAM Enterprises × Schaeffler: GAM will manufacture PSC planetary reducers for Schaeffler — after Schaeffler recently placed orders on the finished-humanoid side in Europe, upstream capacity is now beginning to look outward for contract manufacturing source

· Horizon Robotics (Chinese embodied-AI chipmaker) × GigaDevice: The two companies jointly released a Chinese-made "large brain/small brain" solution for six-axis robotic arms, integrating the controller and main control chip into one domestic supply chain source

V. This Week's Watch

California Truckers' Union Sues DMV, Seeks to Overturn New Driverless Heavy Truck Rule · autonomy

Teamsters California filed suit against the California DMV in Alameda County Superior Court on August 5. The dispute centers on procedure: the DMV first adopted the rule in April this year, allowing commercial vehicles over 10,000 pounds, including trucks, to operate fully autonomously on California public roads. The union alleges the DMV used an expedited rulemaking process intended only for regulations whose first-year implementation costs or benefits fall below $50 million, in order to bypass legally required economic-impact studies and public disclosure — and that in its filing the DMV claimed the rule would not eliminate any jobs or businesses, even though reducing demand for commercial truck drivers is itself an openly stated goal of autonomous freight companies. The complaint also alleges the DMV failed to adequately consider the safety risks of sharing roads with driverless trucks that are still being tested and have not been fully validated, and asks the court to find the rulemaking process fundamentally flawed and to vacate the rule. This marks the first time in the US that a driverless heavy-truck approval process has been challenged in court at the state level, and the outcome will directly affect the timeline for California, the largest testing ground for the technology.Source: Courthouse News Service source

LexisNexis's First Humanoid Robot Patent Report: China Leads on Volume, the US on Per-Patent Quality · humanoid

Legal and IP information firm LexisNexis released the first global innovation report on humanoid robots, covering over 26,000 patent families across three technology areas: interaction systems, body structure design, and control/planning architecture. On volume: as of 2025, China holds 73% of body-structure-design-related patents, versus South Korea 11%, Japan 7%, and the US 5%. Switching to a weighted patent asset index accounting for technical relevance and market coverage, China still leads at 63%, followed by the US at 11%, Japan at 10%, and South Korea at 7% — but the US ranks first in average per-patent competitiveness, a conclusion only visible when volume and quality are examined separately. Over the past decade, China's patent activity has grown roughly 5x in control/planning architecture and about 2.5x in body structure and 2x in interaction systems. In terms of patent-holder structure, companies account for 92%, 83%, and 86% of patents in Japan, Germany, and the US respectively, versus 58% companies and 39% research institutions in China. The top ten by patent asset index are Fanuc, Alphabet, UBTECH (Chinese humanoid robotics maker), Kawasaki Heavy Industries, the Chinese Academy of Sciences, Nvidia, Strong Force Innovation, Harbin Institute of Technology, Intuitive Surgical, and Samsung, with nine of the ten having humanoid-related patents accounting for under 10% of their total patent portfolio — UBTECH is the only specialized company exceeding 25%. The startup ranking's methodology deserves note: it evaluates only unlisted companies that have publicly demonstrated a bipedal-walking humanoid, excluding Tesla, Boston Dynamics, and the already-listed UBTECH — under this criterion the top five are all Chinese, in order: Fourier Intelligence (Chinese humanoid maker), Agibot (Chinese humanoid startup), Limx Dynamics (Chinese humanoid startup), Pudu Robotics (Chinese service robot maker), and Unitree, with AgiBot (LeJu Robotics) ranking 9th. The report's cited market forecasts diverge sharply: Grand View Research and Goldman Sachs project roughly $40 billion by the early-to-mid 2030s, while Morgan Stanley projects the figure, including supply chain and maintenance, could exceed $5 trillion by 2050.Source: Robotics Outlook (reprinted via ifeng Tech) source

Production Racing Toward 100,000 Units, But Under 5% Actually Reach the Factory Floor · humanoid

Gan Xiaobin, Deputy Director of the Ministry of Industry and Information Technology's Science and Technology Department, gave figures at WAIC in July this year: China's humanoid robot production was roughly 20,000 units in 2025, already over 40,000 units in H1 2026, and is expected to surpass 100,000 units for the full year. A more noteworthy set of figures from the same period comes from industry statistics: China's market sold roughly 23,000 humanoid robots in H1 2026, with about 30% going to research institutions, 30% used for trade shows, over 30% exported, and under 5% actually entering factory production lines. Taken together, these two figures show that production growth and depth of industrialization are not currently the same thing — the boom in shipment figures has not yet translated into shop-floor hours.⚠️ Third-party compilationSource: AI Research Notes Pro source (WeChat, CN)

This Week's Supply Chain Roundup: The Cost-Reduction Curve Is Bottlenecked at Transmission Components, With Tightest Supply in Harmonic Reducers and Lead Screws · hardware

This week's in-depth component coverage converges on the same conclusion: the main battlefield for cutting humanoid costs isn't in algorithms or the finished robot, but in transmission components. On a comparable basis, Optimus's bill-of-materials cost is about $131,000 without relying on the Chinese supply chain, dropping to about $46,000 when it does — nearly a threefold gap; actuators, lead screws, and reducers together account for about 68% of its BOM. By component: Chinese-made harmonic reducer prices run 40%–60% of Harmonic Drive's (Japanese maker), with Chinese maker Leaderdrive's unit price falling from about RMB 1,900 in 2017 to about RMB 1,000 in 2025, with potential to reach RMB 400–500 at scale. The planetary roller screw has the highest barriers — each unit needs 10–14 screws, priced RMB 5,000–10,000 imported versus about RMB 2,000 once localized — with the bottleneck being 0.5-micron-level machining precision and raising yield from 60% to 85%; foreign suppliers still hold over 70% of this market. Six-axis force sensors are about 55% localized, with each unit requiring 4 six-axis sensors plus 28 joint torque sensors.

The tight spots are equally specific. Per Soochow Securities, global humanoid harmonic reducer demand is projected at roughly 8.8 million units by 2027, against effective capacity of only about 5 million units — a shortfall exceeding 20%. Leaderdrive signed a three-year exclusive agreement with Unitree covering reducers for 110,000 finished robots, with orders booked through Q3 2027 and capacity utilization around 95% — yet its gross margin has already fallen from 41% in 2023 to 33.6% in Q1 2026, as capacity expansion and price competition happen simultaneously. Lifespan is another unresolved hurdle: planetary roller screws require precision held within ±6 microns while joint service life is only about 1 year, and precision components like reducers generally last around 8,000 hours — under an 8-hour workday, that's under three years before replacement is needed, still short of the 5+ years industrial applications require. Tesla's guidance to suppliers targets 1,000 units/week by September and 2,000–2,500 units/week by year-end (equivalent to a year's supply of components for 100,000 units), though several supply-chain sources believe the year-end target won't be met; Elon Musk himself said on the Q2 earnings call on July 22 that Optimus will be the hardest product Tesla has ever mass-produced, citing the fact that "there is no supply chain."⚠️ Brokerage research compilationSources: Tech Dissection source (WeChat, CN); AI Research Notes Pro source (WeChat, CN)

FutureX · Physical AI Daily — Issue 80 (08/06)

Shawn — Wed, 05 Aug 2026 14:52:35 +0000

Today's Highlights

· Wayve×Uber secures London ride-hailing license, beating Waymo and Baidu to supervised Robotaxi launch

· Nvidia open-sources autonomous driving VLA Alpamayo 2 Super, 34B weights ready for direct commercial use

· Xiaomi's embodied foundation model Xiaomi-Robotics-1 officially open-sourced, trained on 100,000 hours of real robot data

· Hai Robotics' shelf-climbing robot reaches 10,000 units in global deployment

· Amazon's largest Kansai hub goes live, deploying roughly 3,000 transport robots in a single warehouse

· Unitree begins preliminary price inquiry today, estimated IPO price around RMB 104/share, valuation above RMB 40 billion

· Daegu, South Korea completes the country's first humanoid robot safety certification center, with KRW 18.7 billion invested over five years

· Tsinghua AIR×BAIR continuous-time world model generates at arbitrary frame rates and can even run inference backward

I. Research Progress

Quo Vadis, World Modeling? Where Should World Models Go Next · world-model

This position paper, trending at 27↑ on Hugging Face today, is worth reading for how it relocates the definition of "world model" entirely: classic world models predict physical state transitions, but a continuously self-improving agent actually needs "usable information" — execution outcomes, retrieved experience and skills, verification signals. Based on this, the authors propose Agent-Centric Interactive World Proxies, organizing six proxy categories by feedback modality — dynamics, space, execution, memory/experience, skills, and reward/verification — and mapping three tiers of function: in-context guidance at inference time, producing rewards and synthetic rollouts at training time, and co-evolution between agent and proxy. This is a clear course correction against the current tendency to equate world models with "video prediction."

Yu Yang et al. · arXiv 2608.02713 source

Light-Loco-Parkour: A Single Policy for Whole-Body Humanoid Parkour · locomotion

Previously, whole-body humanoid control either tracked expressive motions but was terrain-blind, or reacted to terrain in real time while leaving arms and torso idle; this work solves both sides with one deployable policy. The policy takes only onboard depth maps and velocity commands as input, deciding on its own whether to walk, balance, climb, descend stairs, or vault over obstacles — with no reference inputs, skill labels, hardcoded state machines, or runtime motion generators. Skills are learned by expanding from a small set of seed motions into terrain-matched feasible references, with successful zero-shot transfer to real hardware both indoors and outdoors.

Hongming Chen et al. · arXiv 2608.02653 source

How Should VLAs Actually Use Proprioceptive State? · vla

Nearly all VLAs consume robot proprioceptive state, but the methods vary wildly — serialized into text prompts, projected into vision-language prefixes, or fed directly to the action expert — and mostly using only the current frame, with no systematic comparison to date. Holding backbone, data, action representation, and evaluation protocol fixed, this paper implements 5 interface variants, compares them across 45 atomic tasks + 20 compound tasks, and sweeps state history length from 1 frame to 96 frames, answering three open questions: whether state actually helps, how much history helps, and where it should enter the model. These are conclusions with immediate engineering applicability.

Yiren Zhao et al. · arXiv 2608.03052 source

DRIFT: The Adversarial Robustness of Flow-Matching VLAs Is an Illusion · vla

It was previously widely believed that flow-matching VLAs like pi0 are inherently more robust to adversarial perturbations than autoregressive VLAs — this paper points out that's only because past attacks ignored the multi-step denoising ODE. The authors attach a universal adversarial patch to a robot gripper to directly attack the denoising velocity field, with a counterintuitive core finding: attacking only the first denoising step is stronger and cheaper than attacking a wider step window (stemming from gradient conflicts specific to input-space optimization, the exact opposite pattern from training-time backdoors). Across all four LIBERO suites on pi0 and pi0.5, a single small patch nearly breaks every task that was previously solvable.

Hoseong Tae, Jong-Seok Lee · arXiv 2608.03207 source

Shooting for Contact: Turning Kinematic References into Dynamically Feasible Trajectories · locomotion

Motion retargeting typically only cares about kinematic similarity, producing references that are extremely hard for RL policies to reproduce, especially for contact-rich behaviors. This paper embeds a differentiable simulator into a nonlinear program to perform contact-implicit multiple shooting, solving contact, friction, collision, self-collision, and joint limits simultaneously without needing to pre-specify contact timing. The result is faster imitation learning training and lower tracking error, and on Unitree G1 it achieved zero-shot sim-to-real transfer for contact-rich crawling and a 180-degree jump turn.

Sergio A. Esteban et al. · arXiv 2608.03116 source

Functional Safety Certification for Industrial Humanoids: Where the Fail-Passive Gap Lies · locomotion

What's holding back industrial humanoids isn't whether they can walk or grasp, but that functional safety certification for legged platforms remains immature — this paper precisely names the core problem the fail-passive gap: a legged robot's safe state is one "actively maintained by control," while the underlying assumption of ISO 13849-1 / EN 60204-1 is that de-energizing equals safe — but cutting power to a walking biped instead creates a new hazard source, falling. The authors use an externally quantifiably-certifiable safety chain (light curtains, e-stops, fail-safe PLC, wireless PROFIsafe) as a probe to localize the uncertifiable residual gap to the robot's own reaction chain, and validated feasibility on a semi-enclosed pick-and-place work cell using the Unitree G1 EDU. The authors explicitly do not claim end-to-end PL e / SIL 3 — this restraint itself signals that the barrier to humanoids entering production lines still sits on the standards side.

Caiwu Ding et al. · arXiv 2608.02809 source

ODEWorld: A Continuous-Time Embodied World Model · world-model

Mainstream video/latent-space world models are built on discrete time, understanding the world at whatever cadence the sampling rate happens to be, while real physics evolves continuously. Tsinghua AIR and UC Berkeley BAIR propose PT-Flow (Physical Time Flow), having the model learn "which direction the world is changing right now, and how fast," then integrating along the velocity field with an ODE solver to obtain the latent state at any point in time. A companion dynamics-representation decoupling hands the static background off to the initial-state memory, so the latent space only describes change; using Jacobian-vector products to project actual visual change velocity provides first-order supervision, sidestepping the JEPA approach's long-standing reliance on bolt-on regularizers to prevent representation collapse — under the RankMe protocol, its 768-dimensional dynamics latent space achieves an effective rank of 425.2, higher than the equally 768-dimensional DINOv2 CLS features (376.1) and the 1024-dimensional V-JEPA 2 (203.7). In practice, the same model can generate at any frame rate without retraining, fill in missing intermediate frames from sparse observations, and even run inference backward by reversing the integration direction.

Tsinghua University Institute for AI Industry Research (AIR) × UC Berkeley BAIR · Analysis: Tsinghua University Institute for AI Industry Research source

Bimanual Manipulation on an 8GB Budget: Real-World Deployment Testing on Entry-Level Jetson · manipulation

Imitation learning policies are almost always evaluated on workstation- or datacenter-grade GPUs, and the real cost of deploying them onto embedded hardware has barely been measured. This paper crams an entire bimanual SO-101 system onto a Jetson Orin Nano Super (8GB), using a desktop GPU only for offline training. Three findings are useful for engineering: ACT converges to usable performance within 100,000 steps (19/20 trials), while Diffusion Policy still hasn't converged after twice as many steps (0/10); after conversion to TensorRT, FP16 cuts inference latency from 114.02ms to 17.93ms, and INT8 to 12.65ms without dropping success rate; and TensorRT's generic INT8 calibration only quantizes the ResNet18 backbone, with none of the 145 transformer layers quantized — explaining why INT8 is still 28% faster than FP16 despite barely shrinking in size.

Ekansh Singh et al. · arXiv 2608.03938 source

Other papers today: UVT modifies not the VLA architecture but the supervision objective, using a unified visual-motor latent goal to jointly encode actions and scene transitions, with the largest gains under constrained training budgets (arXiv 2608.03563 source); SARF defends against "attention hijacking" from physical adversarial patches by fine-tuning only the vision encoder, cutting OpenVLA's failure rate on LIBERO from 100% to an average of 28.6% (arXiv 2608.03231 source); Track4Action distills a world-frame 3D tracker into a VLA, requiring no tracker at deployment time (arXiv 2608.03727 source); Continue or Replan? lets a VLA adaptively decide when to replan rather than on a fixed cycle (arXiv 2608.03483 source); CUDA MPC moves the entire optimization loop onto the GPU, achieving a real-time-feasible horizon one to two orders of magnitude longer than CPU solvers (arXiv 2608.03051 source); LiLa-WAM compresses a world-action model down to end-to-end trainable on a single 24GB GPU (arXiv 2608.03701 source); SLAMFormer-∞ claims to be the first geometric transformer with no explicit distance upper bound on either front-end or back-end (arXiv 2608.03429 source).

Open Source · Tools · Benchmarks

· Xiaomi-Robotics-1: Xiaomi's embodied foundation model was officially open-sourced today, covering the full pipeline from real-robot post-training to deployment along with benchmark reproduction code; Xiaomi states it was trained on over 100,000 hours of real robot manipulation data plus over 10,000 hours of cross-embodiment post-training, achieving 57.4% average success rate on RoboCasa365 and 74.5% on RoboCasa, with weights and code posted to GitHub and Hugging Face source

· OpenETA: an open-source implementation that brings the digital-agent paradigm into the physical world — the Planner selects one Tool call at a time, the Interface controls execution, and the World returns results and new observations, letting robots verify outcomes, revise plans, and turn both success and failure into reusable experience; offers a swappable Planner, composable Tools/Skills, auditable memory, and replayable trajectories (arXiv 2608.03924 source)

· Principles of Robot Autonomy: a robotic autonomy textbook compiled by Marco Pavone's team drawing on years of teaching at Stanford, bridging classical robotics and modern physical AI under a unified conceptual framework, with a Jupyter notebook and implementation-based exercises for each topic (arXiv 2608.03496 source)

· HandTok: a reproducible test bed for a hand-motion tokenizer released alongside DigitCode, organizing symbolic representations hierarchically by the hand's anatomical units, cutting quantization error by three-quarters (arXiv 2608.03127 source)

II. Funding and Deals

Moove | New Round | $250 Million | $2.1 Billion Valuation · autonomy

This company builds the "infrastructure layer" of autonomous driving mobility — vehicle assets, financing, and fleet operations, the heavy lifting that Robotaxi scaling can't avoid but that no one wants to do — and this round's funding is explicitly earmarked for expanding that layer globally. Capital in the Robotaxi space is spilling over from the algorithm side toward the asset and operations side; Uber's $1.25 billion vehicle-supply agreement with Rivian is another facet of the same trend.Source: Business Wire source

Estun Automation (002747.SZ) | Acquisition | RMB 487 Million | 100% Stake in Estun Cool-Drive · industrial

Following its mid-July disclosure of a planned all-cash acquisition, this filing sets the total deal value at RMB 487 million; upon completion, the company will indirectly hold 100% of Estun Cool-Drive and consolidate it into its financial statements. Estun states the acquisition is meant to fill out its "heavy-duty industrial robots + lightweight collaborative robots + embodied intelligence robots" full-scenario product lineup. Industrial robot-arm manufacturers using acquisitions to fill in their embodied-intelligence gap is a common path among Chinese production-line suppliers this year.Source: Sohu source

Unitree Robotics | STAR Market IPO Preliminary Price Inquiry | Targeting RMB 4.202 Billion Raised · humanoid

Following the underwriter's valuation range, August 5 is the preliminary price-inquiry day, with the issue price to be set on August 6, and online/offline subscription opening simultaneously on August 10. The public offering covers 40.4464 million new shares, 10% of total share capital post-issuance. According to CCTV Finance, the market estimates an IPO valuation above RMB 40 billion, implying an estimated issue price of roughly RMB 104/share and a subscription deposit of RMB 52,000 per lot — but this is a back-calculated estimate based on projected valuation, with different estimates ranging from RMB 40 billion to RMB 56 billion, and the actual pricing has yet to be determined. Per the prospectus, the company reported 2025 revenue of RMB 1.699 billion and net profit of RMB 278 million.⚠️ Estimated figuresSource: Sina Technology source

Xihu Digital Intelligence | Seed Round + Angel Round Series | Over RMB 100 Million Cumulative · world-model

Investors include Innoangel Fund, Oriental Jiafu, Zhengxuan Investment, Tianqi Capital, Xihu Innovation Investment Fund, and Jinma Investment. The company, incubated from Westlake University's Autonomous Intelligence Lab, was founded only this January and builds a general-purpose physics engine: pulling physical simulation out of the graphics pipeline to perform dynamics computation directly in latent space, predicting physical causality like fluid sloshing or rigid-body collisions without needing to render photorealistic imagery — the company says this cuts compute requirements by 1 to 2 orders of magnitude. Founder Yu Kaicheng previously led the scaling of visual perception algorithms at Alibaba DAMO Academy. This round's funding will go toward building a dataset of tens of millions of physical interaction trajectories.⚠️ Company's own accountSource: Stockstar source

Mariana Minerals (US) | New Round | $310 Million · industrial

Builds autonomous mining systems; this round's amount ranks among the top tier in the current mining automation space. The company's CEO framed this round explicitly in the context of the critical minerals race — the driving force behind mining automation is shifting from cost reduction to supply chain security.Source: CNBC source

Foundational Industries (US) | Seed Round | $25 Million · industrial

Aims to build "AI-native factories" — organized from the ground up around AI and robotics, rather than automation bolted onto existing factories. This is a rare from-scratch factory-building approach amid the current industrial embodied-AI wave, and the $25 million seed round also signals capital's acceptance of the underlying thesis.Source: AI Insider source

A Personal Mobility Robotics Company | Nearly $100 Million Raised to Date · adjacent

Founded by a former DJI executive, focused on personal mobility robots. Per 36Kr's exclusive report, cumulative funding has reached nearly $100 million — the spillover of drone supply chains and engineering teams into embodied AI continues, and this is already the second company with a DJI-background founding team to secure funding above the $100 million level recently.⚠️ Single-source accountSource: 36Kr source

III. Commercialization and Deployment

Wayve × Uber Secures London Ride-Hailing License, Beating Waymo and Baidu to Launch · autonomy

Wayve's autonomous vehicles have obtained a London Private Hire Vehicle (PHV) license — the same category of license held by ride-hailing operators — meaning the service can now operate on the road as a compliant commercial offering rather than remaining confined to testing. Under the current authorization, each vehicle still carries a safety operator, hands off the wheel but ready to intervene when needed; fully driverless operation would require a different license and further approval. Users who have signed up can begin trying it this summer, ahead of full rollout. The launch uses Wayve's development vehicle, a Ford Mustang Mach-E, equipped with cameras and radar; Wayve says the future fully driverless version will be based on a Nissan-manufactured model. This move clarifies the three-way race in London: Waymo is already testing on London streets and plans to launch a driverless version this year via its own app; Baidu's Apollo Go began testing last week and expects to operate publicly through Freenow by 2027. The timing is especially significant for Uber — its exclusive US partnership with Waymo expires in 2028, and Uber has already stated it plans to offer autonomous mobility with partners including Wayve and Nuro across more than 10 cities worldwide this year.Source: The Business Times source

Hai Robotics' Shelf-Climbing Robot Reaches 10,000 Units in Global Deployment · industrial

On August 3, Hai Robotics (Chinese warehouse robotics maker) announced its shelf-climbing robot has reached 10,000 units in global deployment. The shelf-climbing design is a warehouse robot category where the system uses only this single robot model, paired with shelving, totes, workstations, charging stations, and the HaiQ management platform in a "5+1" configuration, requiring no positioning codes to be deployed — plug-and-play, keeping implementation costs below traditional solutions. Customers span footwear/apparel, fast-moving consumer goods, retail, beauty, pharma, and cross-border third-party logistics, including Anta, Li-Ning, Panasonic Logistics, Itochu Logistics, and Zongteng Cangchu overseas warehouses. Reaching 10,000-unit scale with a single robot model is rare, solid evidence of real scaled deployment in warehouse robotics.Source: Guandian.cn source

Amazon's Largest Kansai Logistics Hub Goes Live, Deploying Roughly 3,000 Transport Robots in a Single Warehouse · industrial

The Mukogawa FC in Amagasaki, Hyogo Prefecture, which opened on August 5, is Amazon's largest logistics facility in the Kansai region, with total floor space of roughly 110,000 square meters and inbound/outbound processing capacity of over 500,000 items/day each. At its core is the Amazon Robotics system: roughly 3,000 drive robots deployed throughout the warehouse, moving about 30,000 dedicated shelving units — the company says this allows for up to 40% more product storage compared to fixed shelving. Automated paper-bag packing equipment has also been introduced in the packing process. Amazon now operates more than 25 FCs and more than 65 delivery stations in Japan, with investment in Hyogo Prefecture alone exceeding JPY 44 billion in 2025. This is the most solid real-world deployment in this issue — not a pilot, but full-capacity operation live from day one.Source: BigGo Finance source

Waymo Removes Dallas Waitlist, Opens Citywide · autonomy

Anyone can now download the app and hail a ride, though rides to Dallas Love Field airport are not yet supported — that stretch of road is still under testing. Waymo opened public service in Dallas this February but with a waitlist requirement; during that period, roughly 150,000 riders used its app. The fleet consists of Jaguar I-Pace vehicles, with charging and maintenance handled by Avis Budget. Removing the waitlist follows the same playbook Waymo has reused in Phoenix, Los Angeles, and San Francisco. Worth noting is its Texas track record: this past May, service was temporarily suspended in Dallas, Houston, San Antonio, and Atlanta due to issues handling heavy rain and flooded roads.Source: TechCrunch source

Humanoid Robot Daily Rental Business Takes Shape: AgiBot's SHAREBOT Surpasses 5,500 Orders in Three Months · humanoid

AgiBot (Chinese humanoid robotics startup) established rental subsidiary SHAREBOT in December 2025; in its first three months live, the platform logged over 5,500 orders, with daily rental prices starting around RMB 3,500, and that price includes transport and a human operator on-site. One e-commerce livestreamer in Hangzhou rents out her own robot at RMB 3,000/day for trade shows, event performances, and even marriage proposals, having purchased the unit for roughly $30,000 in 2025; an investor in Jiangsu bought 8 units specifically to rent out. "What's actually being rented is a small service team, not an autonomous laborer" — the fact that the operator is included is itself telling: the current commercial loop runs on event-exposure value rather than autonomous work capability. AgiBot projects this market could reach $1.5 billion by the end of 2026, a company forecast.⚠️ Vendor's own accountSource: Bitcoin News source

Aurora Expands Driverless Freight into Refrigerated Transport · autonomy

Aurora has expanded its partnership with Charger Logistics to bring driverless truck capacity into refrigerated freight. Cold chain demands higher timeliness and temperature-control continuity than general freight, marking an extension of autonomous long-haul trucking into a higher-value category.Source: FleetOwner source

EACON Autonomous Driving System Installed on New-Energy Truck Fleet at Xinjiang Mining Site · autonomy

Easy Control Intelligent Driving's (Chinese mining autonomy company) autonomous driving system has been deployed on a new fleet of Lovol pure-electric and KNOW-HOW methanol trucks in Xinjiang. Mining sites are among the closed-scenario applications where autonomous driving has already achieved commercial viability in China, and the notable point here is that the system was installed simultaneously on two new-energy chassis types — pure-electric and methanol — rather than retrofitted onto conventional diesel trucks.Source: International Mining source

IV. Industry Developments

Nvidia Open-Sources Alpamayo 2 Super, 34B Autonomous Driving VLA Weights Ready for Direct Commercial Use · autonomy

The significance of this release isn't the parameter count but the license: the weights are released under the Linux Foundation's permissive OpenMDW-1.1 license, code under Apache 2.0, covering fine-tuning, derivative models, and commercial redistribution — and Nvidia is applying this license retroactively across the entire Alpamayo family, meaning versions previously available for research only can be deployed directly starting today. The model itself is a 34B vision-language-action model, combining a 32B VLM backbone built on Cosmos 3 Super Reasoner and post-trained with reinforcement learning, plus a 2.3B diffusion-based action decoder; a single forward pass takes surround-view camera video and simultaneously outputs a planned trajectory, a causal explanation for that trajectory, and meta-actions (yielding, lane changes, etc.). Training data comprises roughly 115,000 hours of multi-camera driving video and about 3.7 million Chain-of-Causation annotations. Nvidia reports a Lingo-Judge score of 79.2 on LingoQA, ranking first among nearly 40 evaluated models — 17.0 points above Qwen2.5-VL 72B and 15.1 points above Gemini 2.5 Pro; open-loop minADE₆@6.4s is 0.911 meters. The design explicitly targets long-tail scenarios: temporary rule changes at construction zones, vehicles cutting in illegally — situations that can't be exhaustively enumerated. Wu Xinzhou, Nvidia's head of autonomous driving, stated the strategic positioning explicitly for the first time — "Nvidia wants to be the Android of autonomous driving" — meaning it won't operate as a Robotaxi operator, but will provide the foundation of "three computers": in-vehicle inference, cloud training, and simulation verification; open-sourcing the model, the Cosmos world model, data tools, and evaluation framework together is precisely the move to lower the barrier to entry into this ecosystem. Both the leaderboard results and the claim of cutting "annotation cycles from months to days" are Nvidia's own self-reported figures.⚠️ Vendor's own accountSource: MarkTechPost, 36Kr source

Daegu, South Korea Completes the Country's First Humanoid Robot Safety Certification Center · humanoid

The city of Daegu announced on the 5th that starting this month it will fully launch the "Humanoid Robot Safety Certification Center" project, selected through a competitive process by the Ministry of Trade, Industry and Energy, which will establish for the first time in South Korea a safety verification and international standards certification system for humanoid robots. Over the five years through 2030, KRW 18.7 billion will be invested (including KRW 9.7 billion in national funding), located at the National Robot Testing Field in Dalseong-gun, led by the Daegu-headquartered Korea Institute for Robot Industry Advancement, with five other specialized institutions participating. The center will establish three evaluation systems — dynamic stability, AI reliability, and cybersecurity — linked with the National Robot Testing Field to provide one-stop support spanning test evaluation, demonstration, and certification, aiming to help Korean companies obtain safety certification and open overseas markets. Read alongside today's paper on the "fail-passive gap," this kind of national certification infrastructure is precisely the missing piece for humanoids entering industrial settings.Source: Seoul Economic Daily source

Samsung's Humanoid Push Revealed: Data Factory in Gumi, In-House World Model and Actuator Development in Parallel · humanoid

According to Korean media reports on the 5th, Samsung Electronics has confirmed its subsidiary Rainbow Robotics (Korean humanoid robotics maker, KAIST spinoff) is independently developing humanoid robots. Rainbow Robotics originated from KAIST's HUBO lab, with Samsung progressively acquiring a stake and gaining control in 2023–2024; its main revenue currently comes from the RB series collaborative robot arms, while it advances commercialization of the RBY1 wheeled bimanual humanoid platform. The newly disclosed development is Samsung's plan to build a data factory in Gumi, North Gyeongsang Province, integrating process data from existing production sites into a system to assist robot learning; earlier reports also indicated Samsung plans to build a humanoid mass-production base in Gumi, with total investment of roughly KRW 13 trillion. On the model side, industry sources say Samsung is also developing a world model; on the hardware side, Samsung is developing actuators in-house, drawing on its home-appliance motor technology. Organizationally, the RX (Robot Experience) business promotion office, established July 21, reports directly to CEO Roh Tae-moon, headed by Lee Dong-gun, who previously led Boston Dynamics' robotics strategy at Hyundai Motor Group.⚠️ Single-source accountSource: Cailianshe source

Mitsubishi Motors to Convert Idle Kyoto Engine Line, Targeting Up to 1,000 Humanoids per Month · humanoid

Mitsubishi Motors is partnering with Tokyo-based startup Highlanders to convert an idle engine production line at its Kyoto plant into a robot assembly line, targeting production start in early 2027 with a planned monthly capacity of up to 1,000 industrial humanoid robots. The first batch of robots will initially be deployed on Mitsubishi's own engine production lines to generate operational data, with external sales to follow validation. Highlanders is a University of Tokyo spinoff founded in 2023, whose flagship platform HL Human was released in 2025 with 19 degrees of freedom. Both capacity figures and timeline are per the companies' plans.⚠️ Planned figuresSource: Interesting Engineering source

Dobot Unveils Humanoid Robot Lumo, Targeting Home and Educational Companionship · humanoid

Dobot (Chinese robotics/robot-arm maker) unveiled the humanoid robot Dobot LUMO on August 5, standing nearly 1.3 meters tall and built on the company's in-house Kongyi embodied foundation model. Official demonstrations show it walking across outdoor terrain including lawns, sand, and cobblestones, along with emotion recognition, dancing, and backflips, while also being positioned as an embodied-AI teaching platform supporting further development. The company describes it as "the world's first fully all-terrain embodied humanoid robot," with the official demo video claiming no special effects were used — but "all-terrain" (quan qi) is a self-coined category term by the company, and its capability boundaries and mass-production timeline remain undisclosed.⚠️ Vendor's own accountSource: Securities Daily source

China Association of Automobile Manufacturers Establishes Autonomous Vehicle Industry Development Joint Council · autonomy

The China Association of Automobile Manufacturers has launched the establishment of an Autonomous Vehicle Industry Development Joint Council. Coming right after China's mandatory autonomous driving national standard just took effect (previously reported), this standing industry-association-level coordination mechanism typically corresponds to industry-wide alignment work on implementation details and compliance practices for the standard.Source: Sohu source

JD.com and Wuba Intelligence Reach Strategic Partnership · adjacent

The two parties will jointly advance the industrialization and application of embodied intelligence robots. E-commerce and logistics platforms continue to bring in embodied AI capability through strategic partnerships rather than in-house development — essentially locking in supply for their own warehousing and fulfillment scenarios.Source: Cnbeta source

Hardware · Supply Chain

· GigaDevice (Chinese chipmaker) GD32F50MxxG: a highly integrated motor-control MCU for humanoid joints, collaborative arms, and dexterous hands, packaging an in-house three-phase gate driver and 4-channel rail-to-rail op-amp into a single 252MHz Cortex-M33 chip, eliminating the need for external standalone drivers and sampling op-amps; already certified to IEC 61508 SIL2 functional safety, with the 8mm×8mm QFN80 variant now sampling and slated for mass production in December 2026, and the 7mm×7mm BGA100 variant planned for volume supply in March 2027 source

· Molex MiniMix: a hybrid connector combining 15.0A power delivery and 1Gbps automotive-grade Ethernet into a single interface, with a 5.65mm wiring profile that the company says takes up to 50% less routing space than alternative solutions, intended to replace the currently common manual wiring and hand-soldering inside humanoid joints — assembly labor time, not component cost, is the mass-production bottleneck, which is where this part's value lies (vendor's own account) source

· D-Robotics (Chinese embodied-AI chip maker) × GigaDevice: jointly released a full-stack control solution for a six-axis collaborative robot arm based on the Sunrise S600 processor plus GD32H77R MCU, with a Chinese compute chip and a Chinese control MCU now being supplied together as a complete package to robot-arm manufacturers source

· Hesai Technology (Chinese lidar maker): selected for Morgan Stanley's "Humanoid 100" list, with a lidar manufacturer now included in the humanoid supply-chain investment pool, reflecting the growing weight of its robotics business in its valuation narrative source

FutureX · Physical AI Daily — Issue 79 (08/05)

Shawn — Tue, 04 Aug 2026 15:03:35 +0000

Today's Highlights

· China's mandatory autonomous driving national standard lands — L3 systems must include driver takeover-capability monitoring, effective July 2027

· RoboArena score-gaming exposed: a single account handled 58% of Spirit v1.6's evaluations

· Single-leg standing test, 90 moves — eight top-tier general humanoid policies pass zero

1. Research Progress

Eight top-tier general humanoid policies, zero out of 90 single-leg standing moves · locomotion

Humanoid policies have made fast progress on dancing, rolling, and whole-body tracking over the past two years, yet collectively fail at one of the most basic demands: standing still on one leg. On the authors' single-leg balance benchmark, eight published SOTA general policies clean-completed 0 of 90 test moves — they recover from imbalance via stepping and small hops rather than avoiding imbalance in the first place. Achieving "prevention" requires the capture point (a center-of-mass position extrapolated using velocity), which had never driven a real-robot policy before because it depends on base linear velocity that no onboard sensor can provide; this paper finds that expressing it in the support-foot frame cancels this exact term, making it reconstructable from encoders and an IMU alone. Feeding this deployable dynamic center-of-mass observation directly into an actor running on hardware, paired with a reward library translated item-by-item from human posture control, the policy FDDC cleanly completes 86 of 90 held-out moves and transfers to a real Unitree G1. Ablation shows this observation is the single largest contributor — removing it drops performance by 40 points.

Yikai Zhou et al. · arXiv 2608.00500 source

Why does action chunking work? Three popular explanations debunked · manipulation

Predicting and executing multi-step action chunks in a single pass is now nearly standard in robot policies, but why it works has never been settled. After systematically testing in simulation and on real hardware, this paper argues that none of the three mainstream hypotheses — temporal consistency, shortened decision horizon, and representation learning — explains its success; what actually matters is stronger non-Markovian expressiveness and lower compounding error, both of which can, in many settings, be fully reproduced by a "delayed policy" (predicting a single action at each step based on observations from k steps earlier). The authors further identify a previously unnamed benefit — implicit ensembling: an action-chunk policy simultaneously learns multiple temporal relationships such as a_t|o_t and a_t|o_{t-1}, behaviorally equivalent to a model ensemble, which improves robustness and generalization. Building on this, they use a "randomly delayed policy ensemble" to match chunking's performance without chunking at all, and offer a class of policies that explicitly instantiate the ensemble, clearly outperforming chunking itself across multiple domains.

Filippo Lazzati, Kyle Stachowicz, William Chen et al. · arXiv 2608.02547 source

Why VLAs fail on contact-rich tasks: two failure modes, two targeted fixes · vla

Prior remedies mostly added force-sensing branches or training-time regularization, with little investigation into root causes. The authors split failures into two categories: precision failures stem from training misalignment in flow-matching policies, while force failures stem from the structural properties of force signals themselves. Each gets a targeted mechanism, combined into FACT, which achieves an average 66% success rate across five contact-rich tasks versus 41% for the best prior baseline, evaluated over nearly 2,500 real-robot rollouts.

Carlota Parés-Morlans et al. (Stanford / KTH) · arXiv 2608.01402 source

Ego2Robot: converting first-person human video into robot training data at scale — 18,500 hours · vla

Retargeting first-person human video into robot data had previously only been validated for small-scale, single-task policies, and whether it could offer pretraining benefits for VLAs remained an open question. Ego2Robot chains action retargeting, robot-arm visual synthesis, and multi-stage quality filtering into a scalable pipeline that consumes both curated datasets and in-the-wild video, producing 18,561 hours of training data spanning 15 robot embodiments — currently the largest ego-to-robot dataset. To test generalization, the authors add four decoupled perturbation axes on RoboTwin2.0 — visual appearance, scene layout, embodiment, and task semantics — showing that joint pretraining on synthetic and real-robot data consistently improves out-of-distribution generalization across multiple perturbation types, verified in real-robot deployment.

Ye Wang et al. · arXiv 2608.02580 source

Showing ground-truth trajectories to the teacher when labeling driving-VLM chain-of-thought teaches the model to confabulate · autonomy

Today's only paper with community buzz signal (HF↑25). The authors point out that existing autonomous-driving CoT annotation pipelines commonly feed recorded ground-truth future trajectories to the teacher model, inducing a "trajectory anchoring bias" — the teacher doesn't infer decisions from scene evidence but instead fabricates justifications for a known outcome, degrading causal faithfulness and notably worsening hallucination in causally complex scenarios. But simply removing the ground-truth trajectory entangles high-level decisions with geometric synthesis and low-level dynamics. So the authors reframe planning as selecting among explicit candidate trajectories (AD-MCQ) and propose DEFT-RLVR, which turns the future trajectory from an "anchor before the decision" into a "verification target after the decision," improving driving reasoning while preserving or even enhancing general vision capability.

Zixuan Huang et al. · arXiv 2608.01755 source

MiniWorld: a streaming video world model trained from scratch on a single 8-GPU machine in days · world-model

Current video world models mostly repurpose pretrained video generation models, resulting in complex pipelines, expensive compute, and a mismatch between bidirectional pretraining and causal streaming inference; the community has lacked a lightweight, transparent, fully reproducible from-scratch baseline. MiniWorld trains a block-causal Video Diffusion Transformer with flow matching in the latent space of a pretrained Video VAE, building on Diffusion Forcing but with a block-wise non-decreasing noise schedule and two-stage continued training; at inference, it combines a rolling KV cache with pipelined asynchronous denoising. The entire model can be trained in a few days on a single 8-GPU server, with training and inference code plus pretrained weights released together.

Yian Zhao et al. · arXiv 2608.01127 source

3D-printed electrical impedance tomography e-skin aims to free humanoid whole-body touch from "rebuild for every robot" · perception

Traditional taxel arrays are hard to scale up in area, wiring complexity, and robot-specific curvature, which is why humanoid whole-body touch has been slow to roll out. The authors use a geometry-adaptive additive manufacturing process to produce conformal electrical impedance tomography tactile skin: a flexible conductive TPU layer forms a continuous sensing field, contact-induced changes couple with conductive patches to alter boundary voltage, which is then reconstructed with a one-step Gauss-Newton EIT solver. Validated on three prototypes — a flat surface, a U-shaped curved surface, and an iCub facial shape — the curved sensor achieved a mean localization error of 6 mm across 18 contact positions, with no supervised post-processing required.

Haofeng Chen et al. · arXiv 2608.02080 source

Letting machines "suffer a bit internally" first makes them more resilient in unfamiliar environments · locomotion

Biological organisms are composed of diverse sensing-motor parts spanning multiple scales and adapt to new environments extremely fast, whereas machines are built from inert materials at smaller scales and break down when faced with the unexpected. The authors propose and validate a mechanism that hadn't been clearly articulated before: tethering morphologically diverse, originally independent agents together with physical connectors. As the connectors learn to recover behaviors disrupted by their own presence, they trigger and "tame" a sufficiently diverse range of internal perturbations; disruptions later encountered from a new environment happen to fall within this already-tamed range, so the collective can continue normal behavior without any additional learning or adaptation. The more participating individuals and the greater their diversity, the stronger the resilience to new environments.

Alice Hein, Josh Bongard · arXiv 2608.02395 source

Other papers today: Faster-WAM questions whether world-action models truly need deep action modules (arXiv 2608.02365 source); SelfWAM conditions future prediction on executed actions rather than just task prompts (arXiv 2608.00725 source); CoWAM encodes synchronization, role compatibility, and collision convergence as "coordination contracts," intervening in bimanual policies only when necessary (arXiv 2608.02578 source); DynamicWAM uses dual-path motion conditioning for dynamic object manipulation (arXiv 2608.00793 source); an empirical study on asynchronous deployment overlaps model inference with action execution to eliminate stalling (arXiv 2608.01880 source); Open-DiffLoco is the first open-source framework to train deployable quadruped locomotion policies via differentiable simulation, training in 20–60 minutes on a single RTX 5080 (arXiv 2608.02069 source); a survey organizes robot learning along "frozen weights vs. self-written skill code" lines, reviewing 77 representative systems (arXiv 2608.01851 source); the first systematic study of prompt injection attacks on multi-agent robot systems, showing attacks can propagate between agents via shared prompt structures (arXiv 2608.00747 source); Rapid Embodiment Adaptation infers embodiment parameters online from short interaction histories to handle hardware variation (arXiv 2608.01506 source); a language-conditioned latent-space safety filter lets safety constraints be specified in natural language (arXiv 2608.00315 source).

Open Source · Tools · Benchmarks

· Amap (Alibaba's mapping unit) ABot-World-0: Alibaba-owned Amap's general-purpose interactive world model has been upgraded and open-sourced, claiming stable continuous inference for 24 hours on a single consumer-grade GPU with no noticeable degradation in image quality, physical consistency, or interaction responsiveness. Mainstream world models generally cap out at 30 seconds to 1 minute of continuous single-pass generation, with the bottleneck being error accumulation from autoregressive generation; Amap's approach is a LongForcing long-horizon training paradigm — continuously feeding the model's own generated output back as subsequent input during training, while constraining it to stay within a plausible world distribution. ⚠️ Vendor claim source

· Open-source robotic guide dog: a research team has released an open-source design for a robotic guide dog aimed at blind and visually impaired people, pushing this type of assistive form factor from closed products toward a reproducible design source

2. Funding & Deals

Kaiwang Data (KWDATA) | New strategic funding round | Over RMB 100 million · adjacent

Led jointly by the Beijing Yizhuang Industrial Upgrading Fund, Huafang Capital, and Tianji Capital, with Xinding Capital, Mairui Asset Management, and other institutions following on. More notable is who joined the follow-on round: Deep Robotics (Chinese quadruped/humanoid robot maker), Songyan Dynamics (Chinese humanoid startup), AgiBot's spin-off X² Robot, and Zhiyuan Robotics's (Chinese humanoid startup) subsidiary Mifeng Technology — marking the first time this full-chain AI data infrastructure company, founded in 2022, has brought in embodied-AI peers as strategic investors. The company's core business began in autonomous driving data, covering passenger vehicles, commercial vehicles, logistics vehicles, and algorithm firms across the board; over the past two years it has moved into robot data collection, with data categories now covering first-person bare-hand video, tactile sensing gloves, UMI grippers, and third-person panoramic video, deployed in settings including supermarkets, restaurants, homes, and auto production lines. The company says its stable monthly effective-data output capacity is now approaching 100,000 hours, and this round will fund an integrated data trading platform, specialized talent recruitment for world models, and capacity expansion in three key areas. Founder and CEO Yu Xu previously worked at Uber, Momenta, and ByteDance.

Sources: Zhidx source, Cyzone source

Taoshi (Shenzhen) | Series A | RMB 140 million | Valuation surpasses RMB 1 billion · hardware

Participants included Qianhai Financial Holdings, Guangzhou Guokong, Haichuan Juyi, and Hangzhou Zhongshen, with TPG (德太资本) serving as financial advisor. Founded in 2016, Taoshi pursues an unusual niche — micro-scale toroidal-envelope worm-gear reducers — integrating deceleration and 90-degree turning into a single structure, roughly 40% smaller than the conventional "reducer plus separate angle converter" approach, with precision around ±0.5 arc-minutes and a lifespan of up to 10,000 hours. Its products have entered supply chains at Foxconn, Lens Technology, Luxshare Precision, and Piotech, and its dexterous-hand joint modules have already signed a 100,000-unit supply agreement, with customers including Linker Hand and Chaoweit Dynamics. On capacity, its existing ~20,000-square-meter facility produces 500,000 to 700,000 key modules per year, with a target of 1 to 1.5 million once new facilities come online. Actuator systems account for roughly 45% of the cost of a full humanoid robot, making reducers a key link in China-made substitution for imports.

Source: Ifeng Tech · Hardcore source

DISCOVER Robotics | Angel+ round | $100 million · embodied

Participants included IDG, Xinglian, Wuyuefeng, Fortune Venture Capital, Joyoung, Huaying Capital, and the Binhu District Industrial Group, with existing shareholders all adding on and Gengxin Capital serving as lead financial advisor. The notable part is the timing: less than a month after its first angel round of over $100 million closed, the two rounds together total over $200 million; the company positions itself in consumer-grade embodied AI. ⚠️ Single-party claim

Source: Sina Finance source

Unitree Robotics | STAR Market IPO | Underwriter valuation exceeds RMB 50 billion · humanoid

Following the previously disclosed listing valuation of roughly RMB 42 billion, an August 5 book-building date, and an August 10 subscription date, underwriter CITIC Securities has now put out a research-desk assessment putting post-listing valuation at over RMB 50 billion — higher than the valuation implied by the offering terms. The company plans to issue 40.45 million shares, about 10% of post-offering equity, raising RMB 4.202 billion. ⚠️ Broker claim

Sources: Reuters source, Lianhe Zaobao source

Exclaim Robotics (Zurich, Switzerland) | Seed round | €4.29 million · adjacent

The company has come out of stealth with a focus on robots for maintaining AI data centers — in a field crowded with humanoid and factory-automation plays, this is a niche demand pulled directly out of the compute buildout.

Sources: EU-Startups source, AI Insider source

AI² Robotics (Zhipingfang) | Planning Hong Kong listing | Post-money valuation over RMB 20 billion · embodied

Reports suggest the company could IPO as early as next year, with its valuation surpassing RMB 20 billion after its most recent funding round. If it goes through, it would be another embodied-AI foundation-model company, alongside AgiBot and Unitree, putting a valuation payoff on the calendar. ⚠️ Rumor-sourced

Sources: iyiou.com source, The Standard source

3. Commercialization & Deployment

Pony.ai reveals Robotruck production timeline for the first time in two years: 500–1,000 heavy trucks, 100,000 light trucks · autonomy

Over the past two years, Pony.ai's public narrative has been almost entirely dominated by Robotaxi, with its trucking business rarely surfacing. At an August 3 media briefing, the company systematically disclosed its trucking progress for the first time, setting a production target of 500 to 1,000 heavy trucks and 100,000 light trucks. The reason for speaking up now is that cost and engineering maturity have both arrived at once: the seventh-generation Robotaxi fleet has surpassed 1,000 vehicles, and the cost of the self-driving kit has fallen 60%–70%, while the sensor and compute platforms for heavy and light trucks are nearly identical to those used in Robotaxi, letting the truck line directly capture that cost dividend and deployment experience. He Xing, VP and head of the trucking business unit, said the company has accumulated a full set of experience partnering on mass production with OEMs, and that heavy trucks have "reached the stage where large-scale pilot production is possible," with the fourth-generation heavy-truck production line launching alongside a parallel rollout of light trucks. Looking back also explains why it took so long: the first-generation vehicles were "workshop-style" retrofits, and the two prototype vehicles exposed consistency and reliability issues within less than a year of use — "every vehicle had its own unique hardware faults." It wasn't until choosing the FAW Jiefang J7 for the second generation in late 2020 that the biggest pain point of the R&D phase was resolved. ⚠️ Plan-stage claim

Source: 36Kr source

FedEx moves autonomous loading and unloading from pilot to hub scale, Dexterity's dual-arm robots take over trailer loading · industrial

FedEx's partnership with intelligent robotics company Dexterity has scaled up from pilot to full operational deployment: the Mech trailer loading/unloading system has moved into a 325,000-square-foot hub warehouse in Hagerstown, Maryland, at what FedEx describes as a "significantly larger" operating scale than before. The two companies have partnered since 2023, with the predecessor system DexR first deployed at an initial test site in Tracy, California. The dual-arm Mech has a wingspan of nearly 18 feet, a vertical reach of 7.5 feet, and can handle a single payload of up to 60 kilograms, while remaining compact enough to autonomously drive into a trailer to operate. It's powered by Dexterity's physical AI model, Foresight, trained on more than 100 million autonomous actions according to the company, reasoning across three spatial dimensions plus time to balance space utilization, stacking stability, and speed. FedEx says the partnership is testing how physical AI can be woven into broader hub operations — destination planning, trailer allocation, maintenance, and labor workflows.

Source: WWD · Sourcing Journal source

South Korea's a2z lands 11 billion won UAE order, fully replacing Chinese-made autonomous vehicles · autonomy

Autonomous driving company Autonomous a2z has signed a "Smart Mobility Convergence Project" contract with Abu Dhabi's AI company Space42, worth 27.91 million dirhams, roughly 11 billion won. The key detail is the direction of the swap: Space42's TXAI service, previously operated with Chinese-made autonomous vehicles, will be fully replaced with a2z vehicles. The order covers 19 vehicles total, including 8 units of a2z's self-developed L4 driverless shuttle ROii, 5 modified Kia PV5 units, 5 modified Carnival units, and 1 modified MAN bus; the ROii will be manufactured in South Korea and shipped as complete vehicles in the second half of the year, with the rest locally procured and retrofitted. a2z will also build an integrated control facility including a control system and remote driving cabin, and integrate with Space42's ride-hailing app. Shuttle trial operations will begin in the second half of the year on Saadiyat Island and Yas Island, transitioning after 2027 to demand-responsive shuttles and tourist routes. The export had previously been delayed after the China-U.S. conflict with Iran temporarily cut off Middle East sea and air routes.

Source: Seoul Economic Daily source

XPeng's IRON humanoid begins small-batch trial production at Guangzhou plant, mass-production line in final debugging · humanoid

XPeng's humanoid robot IRON has begun small-batch trial production at its Guangzhou plant, with the mass-production assembly line simultaneously entering its final debugging phase; the company plans to reach full mass production in Q4 2026. ⚠️ Vendor claim

Source: Autohome source

South Korean nuclear waste disposal facility deploys seven-robot fleet to keep workers out of high-radiation zones · industrial

A radioactive waste disposal facility in South Korea has deployed a fleet of seven robots to reduce worker radiation exposure — one of the most straightforward cases of robot value in scenarios like nuclear waste handling, where "humans shouldn't linger."

Source: Tech Times source

4. Industry Developments

China releases its first mandatory national standard for advanced autonomous driving: L3 must include driver takeover-capability monitoring, effective July 2027 · autonomy

The "Intelligent Connected Vehicles — Safety Requirements for Automated Driving Systems" (GB 44721—2026), drafted under the organization of China's Ministry of Industry and Information Technology, has been approved and released by the State Administration for Market Regulation and the Standardization Administration of China on July 30, and is set to take effect on July 1, 2027. The standard applies to M-class and N-class vehicles equipped with L3 or L4 systems, but not to automated parking systems. Of its four categories of requirements, the two strictest are: first, the safety level of an automated driving system must at least match that of a qualified, attentive human driver currently performing the dynamic driving task, with specified rules for triggering and executing minimal-risk maneuvers; second, on human-machine interaction, it adds a hard requirement for driver takeover-capability monitoring in L3 systems, requiring the entire activation-and-exit process to be safe and status prompts to be standardized, with automakers also required to clearly communicate capability boundaries and responsibility allocation through official websites, in-vehicle terminals, and other channels. The standard also requires automakers to build safety-assurance capabilities spanning design, development, and post-deployment across four dimensions — safety policy, risk management, safety assurance, and safety improvement — and to build a testing system combining "enterprise assurance-capability inspection + safety-file review + confirmatory testing," with third-party bodies conducting confirmatory assessment across proving-ground, road, and simulation tests. Compared with the UN Global Technical Regulation on Automated Driving Systems (ADS GTR), which China led and which was approved in June 2026, the national standard sets more detailed technical requirements for L3/L4 and builds a unified, standardized test-scenario system. Multiple people interviewed converged on the same point: as regulation shifts from "soft guidance" to "hard constraints," room for exaggerated marketing claims narrows, companies that fail to meet the bar will be weeded out, the supply chain may face a fresh reshuffling, industry competition will shift from a performance race to a safety race, and rising compliance costs will accelerate differentiation.

Sources: 36Kr · NBD Auto source, Beijing Business Today source

RoboArena score-gaming controversy gets a data backing: a single account handled 58% of Spirit v1.6's evaluations · embodied

Following last week's scrutiny of embodied-AI leaderboard credibility, this storyline now has concrete evidence. According to the South China Morning Post, Hangzhou-based Spirit AI (Qianxun Intelligence) briefly topped Nvidia to reach the top of RoboArena in early June with Spirit v1.6 — a benchmark co-built by Nvidia with Stanford, UC Berkeley, and other institutions, evaluating how well general robot policies translate digital instructions into real-world actions. Days later, the leaderboard operators restructured their methodology and removed that model along with several others from the official rankings; also removed was X Square Robot's model, previously ranked fourth. Pranav Atreya, a Berkeley PhD student and lead author on the project, said on X that the team had "retroactively removed evaluation results from organizations found to be engaging in leaderboard manipulation," without naming specific companies. The Chinese tech blog "肉身算法" (Embodied Algorithm), analyzing over 4,600 evaluation records prior to June 3, found that a rater account named ECUST Robot Lab completed 58% of all evaluations for Spirit v1.6, under which the model's win rate was 99%, versus only 66% among independent raters; the blog also claimed that the top five models (including Nvidia's DreamZero) each had at least half of their evaluations coming from a single specific account. The vulnerability lay in the original design: anyone could register as a rater, and evaluated models were hosted on developers' own servers, letting developers selectively power on their servers only when their own team was evaluating, boosting the chance of being matched against their own model. As of Tuesday, Nvidia's DreamZero had reclaimed the top spot, with the remaining top-nine slots held by U.S. players including Google DeepMind and Physical Intelligence.

Source: South China Morning Post source

WeRide sets up internal "Superfluid Lab," reportedly led by a former core DeepSeek multimodal researcher · autonomy

WeRide's newly established Superfluid Lab has begun hiring, with a focus on large-model algorithms, simulation algorithms, and AI infrastructure. According to people familiar with the matter, the lab was formed internally in May this year and has already completed groundwork such as infrastructure restructuring; it is reportedly to be led by Ruan Chong, a former core DeepSeek member and core multimodal-technology researcher — Ruan joined DeepSeek back in the High-Flyer (DeepSeek's parent quant fund) era and was a core contributor to multimodal work including Janus-Pro. The lab centers on foundation-model R&D, with a focus on VLA, while also advancing world-model and multimodal directions in parallel; organizationally it sits independent of the R&D department, without overemphasizing rank or role boundaries, with researchers collaborating around a shared model and jointly accountable for its final capabilities. This reflects WeRide repositioning itself: internally, the company now positions itself as a foundation-model company rather than a driver-assistance supplier. ⚠️ Source: people familiar with the matter

Source: LatePost·Lanjing (Blue Whale Finance) source

Waymo co-CEO: Physical AI has no undo button, "move fast and break things" is over · autonomy

In an interview at Y Combinator, Waymo co-CEO Dmitri Dolgov said that once AI moves from digital products into machines operating around people, Silicon Valley's familiar playbook is no longer acceptable and should be replaced with "move fast, ship safely" — safety must be built into the model, training methods, and system architecture from the start, not bolted on after deployment. His framing: a chatbot error can usually just be retried, but vehicle failures cost "lives, not tokens", with "no undo and retry button at all." The weight behind this comes from Waymo's own scale and scar tissue: the company has logged over 220 million fully driverless miles and delivers over 500,000 fully autonomous rides per week; a study comparing 56.7 million fully driverless miles against a human baseline showed statistically lower rates of injury reports and airbag deployments. But in May the company recalled nearly 3,800 vehicles because its software could potentially direct vehicles into flooded roads, and later recalled nearly 3,900 vehicles and temporarily paused highway service after more than a dozen incidents of vehicles entering closed highway construction zones.

Source: Benzinga source

Xu Xin: without spending $1 billion a year, you don't get a ticket to embodied AI · adjacent

In an interview, Xu Xin (founder of Capital Today, a prominent Chinese VC), said embodied AI is harder than robot hardware itself, because the "brain" component demands extremely heavy compute and data — not every team can pull it off. She put the threshold at $1 billion a year. ⚠️ Interview statement

Source: CNFOL source

The spillover effects of U.S. robot import restrictions: South Korean parts suppliers named as beneficiaries · hardware

More than a week after the FCC's covered-list expansion and import restrictions took effect, Korean media are now tallying a different set of consequences: as the U.S. reshoring policy squeezes out Chinese suppliers, it's opening space for Korean parts makers, with KB Securities naming LG Innotek as a beneficiary with improved prospects. China's Ministry of Commerce has previously said it would firmly retaliate, and neither side's posture has shifted this week. ⚠️ Broker claim

Source: Chosunbiz source source

Hardware · Supply Chain

· Samsung Electro-Mechanics, LG Innotek: the two companies have jointly secured exclusive supply of camera modules for Tesla's Cybercab — Robotaxi mass production is turning exclusive sensing-hardware slots into firm orders source

· igus energy chains: launched a robot energy chain supporting 600-degree rotation, targeting wear and interference in industrial robot wrist and turntable cabling under large-angle rotation source

· Grinm Advanced Materials (Youyan): net profit up over 40% year-on-year in H1, with the company saying NdFeB magnet demand from humanoid robots could become a new growth driver for rare-earth magnetic materials — for now this remains a forecast rather than actual orders ⚠️ Forward-looking claim source