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
Top comments (0)