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
Top comments (0)