Today's Highlights
· Zoox secures the first US commercial exemption for a steering-wheel-free robotaxi, fleet capped at 2,500 vehicles per year
· Unitree Robotics sets subscription date for August 10, targeting a raise of RMB 4.202 billion and an IPO valuation of roughly RMB 42 billion
· TurboVLA hits 97.7% on LIBERO with just 0.2B parameters, running at 32Hz on a single RTX 4090
· China's hundred-ton-class fully electric autonomous mining truck enters operation, cutting per-unit-output energy costs by 65%
· China's Ministry of Commerce responds to US robotics ban: obstinate pursuit will be met with firm countermeasures
I. Research Progress
TurboVLA: A real-time VLA with 0.2B parameters running at 32Hz on a single RTX 4090 · vla
Mainstream VLA models typically project vision into a large language model first, then decode it into actions — every policy call pays a compute tax for the large model. TurboVLA instead restructures V→L→A into V+L→A: vision and language are encoded independently with only lightweight bidirectional interaction, and a compact decoder outputs continuous action chunks directly. The result is 0.2B parameters achieving a 97.7% average success rate on LIBERO, with 31.2ms inference latency and under 0.9 GB of GPU memory — matching or exceeding much larger policies. This means high-frequency real-robot control no longer needs to be tied to a data-center-grade GPU. Code is open-sourced.
Hengyi Xie et al. · arXiv 2607.27205 source · HF↑114
HumanCLAW: Nine top VLMs given bodies, and even the best only gets it right 16.8% of the time · benchmark
Evaluating "can a VLM act through a body" has always had a confound: is task failure due to bad model decisions, or did the underlying locomotion controller simply fall over? HumanCLAW separates the two — the model only issues step-by-step atomic skill commands, which a controller translates into sub-second continuous whole-body motion subject to real gravity and collisions, leaving "action intelligence" as the sole measurable quantity. Across 41 indoor scenes and 1,218 long-horizon first-person "find-walk-interact" tasks, none of the nine SOTA models cleared the benchmark; the best score was just 16.8%. The bottleneck isn't recognizing the target — it's embodied self-awareness: models lose track of their own bodies, unable to say where they are, whether they've arrived, or whether they've collided with something.
Siyao Li et al. · arXiv 2607.27180 source · HF↑63
Using a general-purpose agent directly as a robot brain, zero-shot navigation matches industrial-grade policies · autonomy
Vision-language navigation has long relied on specialized policies or task-specific workflows to sustain long decision loops. This paper takes a third path: letting a general-purpose software-engineering agent hold the loop itself, with input limited to monocular RGB and discrete actions. Under this minimal interface, opus-5 at default configuration achieves 70.7% success, and fable-5 at maximum effort reaches 78%; with an optional waypoint tool added, the hybrid agent achieves 76.7% using half the environment steps and under a quarter of the wall-clock time. The authors' controlled experiments indicate capability is largely determined by the model itself, with limited harness influence — and forcing a waypoint interface actually drags down stronger models. Limitations are clear too: success rate drops sharply as tasks lengthen, and latency plus context bloat constrain sustained operation.
Jian Zhou et al. · arXiv 2607.26148 source
CheckVLA: An "execution-time quality check" for open-loop action chunks · world-model
When VLAs perform long-horizon mobile manipulation, they typically issue an entire action chunk at once, so mid-course deviations can't be caught by confidence at dispatch time, and remaining actions keep amplifying the error. CheckVLA uses a separately trained, frozen action-conditioned world model for execution-time verification, using a conformally-calibrated risk threshold to decide when to intervene, and a latency-aware hard prefix to limit rewrites to only those actions still replaceable in time. On RoboCasa365, average success rate is 36.1%, 8.5 points above the 27.6% of periodic replanning; at a matched 5% false-positive rate, the action-conditioned approach lifts timely recall to 77.9%, versus 48.6% for an observation-only baseline. Results are currently from simulation.
Yushan Liu et al. · arXiv 2607.26789 source
What physical quantities does a latent world model actually "know"? The answer depends on what you make it predict · world-model
The core assumption behind world models is that "predicting the future forces the representation to internalize physics." This paper uses controlled interventions to test that claim: in POKEWORLD, where objects look identical but each hides a different mass, friction, and contact stiffness, the authors first show a given parameter can be recovered from raw observations, then test whether it actually enters the latent space. The conclusion: what's kept depends on the prediction target, not just what's in the input — stiffness only enters the latent when the model is trained to predict touch (R²=0.50; simply adding the same signal to the input yields −0.02); friction is recoverable at 0.89 accuracy from input, yet stalls around 0.13 under all deterministic prediction targets. The same pattern replicates on 4,258 real-robot trajectories from RH20T: branches lacking either the right information or the right predictive pressure stay flat even with five times more data.
Kaizhen Tan et al. (NYU / CMU) · arXiv 2607.27017 source
HERO: Starting from zero human demonstrations, consolidating interaction experience into policy through "practice" · manipulation
Existing systems largely learn skills statically — trained once for a specific task and layout, then fixed, never evolving further through physical interaction. HERO orchestrates heuristic reasoning, exemplar reuse, and reflective execution into a self-improving hierarchical embodied agent that bootstraps manipulation experience from zero human demonstrations, progressively consolidating recurring interactions into closed-loop visuomotor policies, and dynamically scheduling capabilities according to the stage of accumulated experience. Experiments show it substantially reduces the human involvement needed in data collection.
Jialiang Li et al. · arXiv 2607.26809 source
Genie Sim PanoWorld: A single 360° panorama generates a freely navigable simulation scene · world-model
Building indoor scenes for embodied training previously meant either per-scene optimization or multi-view capture, and long-distance movement often exposed large occlusion holes. This two-stage feed-forward pipeline first plans an SE(3) navigation trajectory via NavMesh and injects it into a latent-space video diffusion model, generating a controllable panoramic video in just 4 CFG-free denoising steps, then lifts it into a 3D Gaussian scene via a feed-forward reconstructor — directly usable as a simulation asset — and generalizes zero-shot to unseen indoor scenes.
Yongxin Su et al. · arXiv 2607.26646 source
Trustworthy embodied AI: A 37-page systematic framework and trust-level classification · benchmark
Once an embodied system fails, the damage is immediate and physical, so "task completion rate" alone can't establish trustworthiness. This paper defines the goal as sustained safe success and breaks it into four interdependent layers: the model layer produces action proposals with calibrated uncertainty, the system layer ensures authorized actions land reliably with fault-tolerant fallback, the evidence layer supports bounded capability claims through verification and traceability, and the deployment layer maintains claim validity via runtime monitoring and controlled updates. The authors also propose a non-normative trust-level rating for cross-comparison and future standardization — at a moment when China's national humanoid robot standards are being intensively drafted, this kind of classification framework carries more real-world weight than the paper itself.
Xinyu Yang et al. · arXiv 2607.26121 source · Commentary: 具身智能之心 (Embodied Intelligence Heart, Chinese robotics community) source (WeChat, CN)
Other papers today: CG-World, deriving roughly 850,000 temporally aligned world-state sequences from industrial CG production pipelines, with explicit contact events and counterfactual branches (arXiv 2607.26452 source); NeoRacer, an open-source standardized 1:12 autonomous racing platform, priced at $2,699 pre-built with triple the compute of comparable products (arXiv 2607.26855 source); MoMo, abstracting "how an action is executed" into a reusable continuous motion-pattern condition across tasks (arXiv 2607.26315 source); ContactFlow, a cross-embodiment transferable video action condition targeting contact constraints world models struggle to learn (arXiv 2607.26579 source); Enfold, folding the "compute for constructing the future" of world generation models into predictive representations, saving an expensive generation branch (arXiv 2607.26657 source); RL²-VLA, test-time-scaled adaptive latent-space compositional guidance (arXiv 2607.26991 source); DLAM, adding temporal constraints to latent actions extracted from action-free video for joint generation with real actions (arXiv 2607.27138 source); explicit kinematic guidance, injecting 3D structural priors into VLA via analytical concepts (arXiv 2607.26513 source); SymmGrid, using parallelized symmetry to cut wall-clock training time for real-robot reinforcement learning (arXiv 2607.26985 source); VidMap, using video temporal structure for SfM to recover metric pose (arXiv 2607.27194 source); StructureGS, structure-aware Gaussian splatting for articulated-object reconstruction (arXiv 2607.26889 source); Speech2Grasp, efficiently transferring text-conditioned grasp-detection data to speech input (arXiv 2607.26567 source).
Open Source · Tools · Benchmarks
· Hands-on embodied robotic arm deployment tutorial: covers the full pipeline for key algorithms like VLA and reinforcement learning in three days, with complete code included source (WeChat, CN)
· RoboCasa365 benchmark leaderboard: 365 tasks spanning roughly 60 kitchen activity categories, sub-steps ranging from 2 to over 15 segments, purpose-built to test long-horizon semantic judgment and temporal constraints source (WeChat, CN)
II. Funding & Deals
Unitree Robotics | STAR Market IPO | Targeting RMB 4.202 billion | Estimated IPO valuation ~RMB 42 billion · humanoid
Announced the evening of July 30: the STAR Market listing application has been approved by the Shanghai Stock Exchange listing committee and registration approved by the CSRC. The company plans to publicly issue 40,446,434 shares, or 10% of total shares post-issuance, combining strategic placement, offline inquiry-based placement, and online fixed-price issuance. Preliminary inquiry is set for August 5, with both online and offline subscription on August 10; payment date is August 12. The company has special voting rights arrangements; after the issuance, founder Wang Xingxing and the holding platform he controls will together hold 31.29%. Based on a public float ratio of no less than 10%, the estimated IPO valuation is roughly RMB 42 billion. This marks the first formal pricing event for a general-purpose robot manufacturer in the domestic Chinese capital market, following Zhiyuan Robotics' Hong Kong listing.Source: Science and Technology Daily source (WeChat, CN)
Holiday Robotics (South Korea) | Series A | KRW 155 billion (approx. $106 million) · humanoid
Led by IMM Investment, with follow-on from existing investors including Stonebridge and Atinum, and participation from policy-oriented financial institutions such as the Korea Development Bank — setting a record for the largest Series A in South Korean startup history. Founded in 2024 by Sualab founder Song Ki-young, the company's flagship product FRIDAY is a wheeled humanoid — 176cm, 115kg, 64 degrees of freedom, with the hands alone accounting for 40 — with fingertip tactile precision of 0.05N, taking a "concentrate all resources on the end effector" industrial approach; on the software side it also rejects end-to-end VLA in favor of a skill-unit architecture that supports regression testing. The funds will scale R&D to over 100 staff and build annual production capacity exceeding 10,000 units, with plans to mass-produce 100 units in the second half of this year and 1,000 by 2027.Source: 六观阿尔法 (Liuguan Alpha, Chinese industry outlet) source (WeChat, CN)
Vicarious Surgical (US) | Investors vote for bankruptcy liquidation | Approx. $300 million raised cumulatively · adjacent
Investors voted to cease operations and proceed with bankruptcy liquidation. Founded in 2014 by two MIT students and a surgeon, this surgical robotics company was once seen as a challenger to the da Vinci system, with investment from Bill Gates, Eric Schmidt, and Jerry Yang through their respective funds; after going public via SPAC in 2021, its market cap briefly exceeded $1.2 billion, and its products had also received FDA Breakthrough Device designation. Since then, clinical trials and regulatory approval have repeatedly slipped, with losses of $63 million and $50.2 million in 2024 and 2025 respectively; as of late March this year, cash and short-term investments stood at only about $3.7 million, and the company received an NYSE delisting notice. Technically, its overly aggressive decoupled-drive design created yield problems, and having only two robotic arms limited procedure coverage — at a time when embodied-AI startups routinely raise three or four rounds a year, this stands as a counterexample of "buildable but unsellable."Source: 首席财经社 (Chief Finance Society, Chinese financial outlet) source (WeChat, CN)
MUKA Robotics | Angel round | Amount undisclosed · world-model
Funded by Lanchi Ventures and Zhijian Power. Founder Chi Xiaowei, first author of the WoW embodied world model and a HKUST PhD, previously served as chief algorithm lead at the Beijing Yizhuang Humanoid Robot Innovation Center before founding the company in Beijing's Haidian district this April. The company develops the LJM latent joint-conditioned world model in-house, focused on low cost and compatibility with Chinese-made compute. At WAIC, he assessed that China's available compute is only one-fourth to one-fifth of that of North American peers, but data collection is cheaper and hardware supply chains respond faster; world models don't need to achieve perfect generality — mastering local physical regularities is enough to empower physical products.Source: AI智件 (AI Smart Device, Chinese tech outlet) source (WeChat, CN)
Tesollo (South Korea) | IPO launched | KB Securities as lead underwriter · hardware
This specialized dexterous-hand manufacturer has formally launched its listing process. Dexterous hands have this year graduated from a humanoid accessory to a standalone core subsystem, drawing intense private-market interest, but public markets still lack a pricing reference point — Tesollo will be the first test case.Source: PR Newswire source
Lanse Yongxian | Angel+ round | Tens of millions of RMB · adjacent
Funded by Hillhouse Ventures, Qiming Venture Partners, and others. Founded in 2024, the company applies robotics technology to marine scenarios; its first product is a power-assisted fishing reel for sea angling, currently in user beta testing, with funds going toward product iteration and post-launch channel building.Source: Hillhouse Ventures source (WeChat, CN)
Geek+ | Lock-up expiry share sale | Warburg Pincus cashes out approx. RMB 800 million · industrial
Warburg Pincus, a backer for nine years, exited after its restricted shares became eligible for sale. Warehouse robotics is one of the few segments in the embodied-AI space that has already achieved scaled revenue; an early-stage financial investor choosing to exit at this point offers a real-world benchmark against the sector's long-horizon growth narrative.Source: Sina Finance source
III. Commercialization & Deployment
Zoox secures first US commercial exemption for steering-wheel-free robotaxi, set to begin charging fares in Las Vegas · autonomy
NHTSA approved a temporary exemption for Amazon-owned Zoox on Thursday local time, allowing its custom-built robotaxi to charge passengers; the decision has been published in the Federal Register. Zoox's vehicles have no steering wheel or pedals, and this exemption covers 8 federal motor vehicle safety standards, including windshield defrosting and light-vehicle braking systems. Nearly a year ago, NHTSA had granted an earlier exemption allowing demonstration rides and free passenger transport in cities including San Francisco and Las Vegas, but without charging fares — this new exemption is the final federal hurdle. Conditions attached include: commercial fleet size capped at no more than 2,500 vehicles per year, for a two-year period (some foreign media report this cumulatively as "up to 5,000 vehicles over two years," referring to the same cap), and a strengthened regulatory framework that can be adjusted as the technology evolves. The company says it will begin charging fares first in Las Vegas, with other markets requiring state-by-state commercialization requirements; its home state of California still needs deployment permits from the CPUC and DMV for driverless operation. CEO Aicha Evans called it "the first commercial exemption for a purpose-built robotaxi." The same day, NHTSA also updated its overall exemption process.Source: TechCrunch source
China's hundred-ton-class fully electric autonomous mining truck enters operation, per-unit-output energy costs 65% lower than diesel · industrial
According to CCTV News, State Power Investment Corporation's hundred-ton-class bidirectional fully electric autonomous mining dump truck has entered operation at a mine site in Inner Mongolia, capable of autonomously completing the full closed loop of loading, driving, and precision unloading. The powertrain uses a custom fully electric drivetrain paired with semi-solid-state batteries and an "integrated charging-and-swapping" system, addressing the long-standing problem of slow energy replenishment in electric heavy trucks affecting continuous operation; structurally, a 4×4 all-wheel-steering, bidirectional-drive design eliminates the need for the wide-radius turns required by conventional mining trucks. Disclosed figures: per-unit-output energy costs are 65% lower than a comparably sized diesel dump truck, continuous single-vehicle operating range is no less than 6 hours, and overall equipment dispatch rate exceeds 85%. Open-pit mining is one of the few scenarios where autonomous driving already pencils out economically; the notable point here isn't autonomy itself but electrification opening a new gap in operating costs.Source: 快科技 (Kuaikeji, Chinese tech outlet) source
Chery's Aimoga brand surpasses 2,000 cumulative overseas deliveries, expanding from 30 to over 60 countries in one year · humanoid
Aimoga, the robotics brand incubated by Chery, announced cumulative overseas deliveries of over 2,000 units, with products and services now covering more than 60 countries and regions; a shipment ceremony was held Thursday in Wuhu. For comparison, as of the end of 2025 the company had delivered 300 humanoid robots and 1,000 robot dogs to over 30 countries. It has obtained EU CE and US FCC certification, with deployment concentrated in automotive marketing services, public tour guiding, smart policing, and commercial reception; 110 police-use robots were also deployed across multiple Chinese cities in the first half of the year. Its flagship store on JD.com prices the humanoid at RMB 285,800 and the robot dog at RMB 15,800. Chinese robotics brands with genuine overseas delivery volumes reaching four digits are still rare; Aimoga's path leans on Chery's nearly three decades of accumulated overseas distribution channels and localized operations, rather than leading-edge robot hardware technology.Source: CnEVPost source
Waymo resumes highway operations after software fix · autonomy
After a pause of more than two months, Waymo's robotaxis are back on highways. Highway driving imposes very different requirements on speed range and takeover windows than city streets do; the resumption itself says more about the speed of Waymo's validation process converging than the original launch did.Source: TechCrunch source
Midea-owned KUKA signs strategic partnership with Winner Medical and Winner Recycling Materials, worth over RMB 300 million total · industrial
The partnership has a total scope of over RMB 300 million, targeting automation upgrades for medical consumables and recycled-materials production lines. Amid a landscape flooded with letters of intent for humanoid robots, industrial robot production-line contracts with clearly stated deal sizes like this one remain one of the few directly verifiable revenue figures in the sector.Source: 财闻网 (Caiwenwang, Chinese finance outlet) source
BMW deploys humanoid robots for automotive parts sorting · humanoid
BMW showcased a humanoid robot handling automotive parts sorting. ⚠️ Single-source claim Public information has not disclosed the number of units deployed, cycle times, or continuous operating duration on the production line, so this should be understood as a capability demonstration rather than mass-production deployment.Source: stuff.tv source
National-Local Center partners with Bailian Group to build humanoid robot applications for retail · humanoid
The National-Local Joint Engineering Center for Humanoid Robot Innovation has partnered with Bailian Group to advance humanoid robots into retail scenarios. The center operates a heterogeneous humanoid robot training ground and has led the drafting and publication of over 90 industry standards; the real value of such partnerships often lies not in the initial number of units deployed but in converting retail-scenario data and operational protocols into reusable standardized modules.Source: 人形机器人世界 (Humanoid Robot World, Chinese industry outlet) source (WeChat, CN)
IV. Industry News
China's Supreme People's Court rules two patent lawsuits against Unitree constituted malicious litigation, awards RMB 80,000 · humanoid
China's Supreme People's Court ruled that two patent infringement lawsuits filed against Unitree's Go2 and A2 robot dogs constituted malicious litigation, dismissing the plaintiff's infringement claims. The patent in question, titled "an electronic dog," was obtained by plaintiff Luweimei just 5 days before the first lawsuit was filed; the claimed damages started at RMB 500, were raised to RMB 80 million during the second-instance proceedings, then reverted back to RMB 500. The patent was declared invalid by the patent authority on March 12. In the A2 case, Unitree's countersuit was upheld, and Luweimei was ordered to pay RMB 80,000 in reasonable rights-defense costs plus a RMB 3,700 case acceptance fee. With robotics valuations running high, patent trolling is emerging as a new category of operational risk — this also marks the first time a leading Chinese robotics company has secured a judicial finding of "malicious litigation."Source: TechNode source
Credibility of embodied-AI leaderboards under scrutiny: new company claims top spot, industry peers question "everyone claims first place" · world-model
MUKA Robotics, founded less than four months ago, claims its embodied world model LJM ranked second globally on the WorldArena leaderboard, achieved using only 32 domestically produced Zhenwu 810E accelerator cards; the model's approach reasons about causal relationships among actions, scenes, and future states in continuous latent space first, then hands off to a video world model for rendering. ⚠️ Vendor claim The same day, industry commentary directly called out the "everyone claims first place" phenomenon on embodied-AI leaderboards, arguing that leaderboard-gaming incidents have decoupled rankings from real capability. Leaderboards are one of the few public reference points available for cross-comparison at this stage of the field; once evaluation protocols and submission practices become opaque, the information value of "reaching the top" collapses toward zero — for investors and procurement teams alike, the questions worth asking are whether the evaluation set is public and whether the reproduction scripts are available.Source: 具身智能之心 (Embodied Intelligence Heart) source (WeChat, CN), 适才猎头 (Shicai Recruiting, Chinese headhunting outlet) source (WeChat, CN)
Samsung evaluating acquisition of overseas robotics startup · humanoid
Samsung is evaluating the acquisition of an overseas robotics startup to strengthen its humanoid robot strategy. ⚠️ Single-source claim Samsung previously took a controlling stake in Rainbow Robotics with a KRW 267 billion investment and established a Future Robotics office; adding another overseas asset most likely points toward capabilities in models or data rather than hardware.Source: 机器人大讲堂 (Robotics Lecture Hall, Chinese industry outlet) source
Li Auto rolls out OTA 9.1, adds two new modes to its Mach VLA large model · autonomy
Li Auto began rolling out OTA 9.1 today, with an update to its in-vehicle Mach VLA large model adding two new driving modes. In-vehicle VLA is currently the largest real-world user validation ground for the VLA paradigm — the scale of data returned from hundreds of thousands of production vehicles is something no lab-based real-robot fleet can match.Source: Autohome source
BMW and Qualcomm sign ten-year chip supply agreement for autonomous driving and smart cockpit · autonomy
The two companies reached a ten-year chip supply agreement covering autonomous driving and smart cockpit systems. A ten-year term is unusually long even by automotive electronics standards, reflecting how automakers are shifting their autonomous-driving compute platforms from a "multi-vendor bid comparison" purchasing model to a "locked-in architectural foundation."Source: Autohome source
Cross-embodiment data collection and training center unveiled in Beijing, alongside new Beijing-Tianjin-Hebei embodied-AI supply chain alliance · adjacent
A cross-embodiment data collection and training center was unveiled in Beijing, with a Beijing-Tianjin-Hebei coordination alliance established the same day to advance the embodied-AI robotics supply chain. Data and supply chain remain the two hardest bottlenecks at present — total global high-quality real-world physical interaction data amounts to roughly 500,000 hours, while training a general-purpose embodied model with baseline generalization ability requires at least tens of millions of hours, a gap of over 90% — the emphasis on "cross-embodiment" here is precisely aimed at data reuse.Source: Sina Finance source
Mimic Robotics partners with Audi to bring video-action models into factory automation · industrial
Swiss robotics company Mimic Robotics has partnered with Audi to apply frontier video-action models to factory automation. The appeal of video-action models is bypassing task-by-task teaching, learning executable action distributions directly from operation videos — deploying this into an automaker's assembly line, with its extremely demanding cycle-time and consistency requirements, is a fairly rigorous stress test for the paradigm.Source: ZDNet China source
Shunde Embodied AI Training and Application Industrial Park unveiled, just two days after the local development bureau launch · adjacent
Following the launch of China's first Embodied AI Development Bureau in Shunde on July 28 (previously reported), the Embodied AI Training and Application Industrial Park was officially unveiled on July 30 at the Yunlu Listed Companies Town's Xinxing Zhicheng Robotics Center. Local government moved from establishing an agency to opening a physical site in just two days — training grounds and data collection facilities are becoming standard infrastructure for regional investment promotion.Source: Shunde Release source (WeChat, CN)
Hardware · Supply Chain
· MediaTek: entering the robotics chip race with a 3nm platform, landing in the 50TOPS tier; vendors are currently commonly "mixing and matching" — Zhiyuan Robotics' (Chinese humanoid startup) Lingxi X2 base-compute board has already replaced its earlier Jetson design with dual Rockchip RK3588 chips, while its high-compute board still uses Jetson Orin NX source (WeChat, CN)
· LG Electronics: has moved up mass production of its in-house actuator Axium, targeting orders in the second half of the year; the company produces roughly 45 million motors annually, with motor design expertise dating back to 1962 source
· Samsung Electro-Mechanics: has begun mass production of a humanoid-robot-specific camera, formally naming physical AI as a key strategic focus source
· Nidec: repositioning its precision reducer product line toward collaborative robots, humanoids, and automation, moving directly into traditional Japanese-manufacturer territory source
· Naxin Micro: launched a high-compute EtherCAT real-time control MCU for energy storage and robotics, targeting the millisecond-level force-position hybrid control bus needed by dexterous hands and joint modules source
V. Weekly Watch
US-China robotics controls escalate: one week after FCC's Covered List takes effect, Ministry of Commerce warns of "firm countermeasures" · adjacent
On July 28, the FCC announced it would add foreign-manufactured power inverters and advanced robotics equipment to its "Covered List" — new models covered will be unable to obtain FCC certification authorization and barred from sale in the US market, with the scope covering mobile robots including humanoids and quadrupeds. On July 29, China's Ministry of Foreign Affairs voiced opposition to "generalizing the concept of national security," and on July 30 the Ministry of Commerce further stated that the measure, while nominally framed as "non-discriminatory," is in substance a discriminatory crackdown on Chinese companies and products, and that should the US side persist, China will take firm countermeasures. Two new points of note this week: first, the ban's coverage is broader than headlines suggest, also sweeping in consumer products like robot vacuums; second, industry voices are entering the fray — Boston Dynamics' CEO has publicly cited "data leakage" as grounds for restricting Chinese robots. For Chinese manufacturers, the near-term impact is concentrated on the certification pathway for new models; meanwhile, over 70% of the humanoid robots Unitree exported in 2025 went to university research institutions, with under 10% entering actual industrial applications — actual revenue exposure is out of proportion with the intensity of the public debate.Source: Guancha (Observer Network), Sputnik, DeepTech深科技 source (WeChat, CN)
NHTSA simultaneously eases exemption process, opening an institutionalized pathway for vehicles without controls · autonomy
On the same day it approved Zoox's commercial exemption, NHTSA updated its overall exemption process — which allows automakers to temporarily sell limited quantities of non-compliant vehicles, mainly for testing new technology. The significance here is converting "case-by-case negotiation" approvals into a predictable channel: previously, every steering-wheel-and-pedal-free vehicle had to negotiate its own individual approval; going forward, applicants can at least plan around a timeline and a defined review standard. Combined with NHTSA's move this month toward proposing to eliminate the mandatory brake-pedal requirement (previously reported), the US federal government is systematically dismantling the regulatory ceiling on purpose-built robotaxis. This stands in contrast to the FCC's tightening on the robotics import side: the same administration is loosening rules for autonomous driving while tightening them for the robotics supply chain, both pointing to the same underlying industrial-protection logic.Source: TechCrunch source
Research report discrepancy: was H1 embodied-AI funding RMB 43.7 billion or RMB 90 billion? Depends how you count · adjacent
Two conflicting figures circulated this week: one report states total H1 2026 funding for China's embodied-AI sector at RMB 43.766 billion, of which the "brain faction" — companies whose core business is the embodied AI model — raised RMB 22.253 billion, or 50.8% of the total, while integrated hardware/body companies raised only RMB 5.6 billion, or 12.8%; another report, citing IT Juzi data, puts total H1 domestic embodied-AI funding at over RMB 90 billion, up 5x year-over-year, across more than 300 deals. The two figures differ by more than double, driven by differences in sector scope (whether components, data services, and autonomous driving are included) and deal counting methodology (whether pending or strategic investments are included) — either figure must be cited together with its methodology. The genuinely stable signal isn't the absolute number but the structure — capital is shifting from "body" to "brain," and within "brain," further splitting into roughly 42% VLA, 27% world models, 20% data infrastructure, and 11% edge chips.⚠️ Single-source claimSource: 机器最前线 (Robot Frontline, Chinese industry outlet) source (WeChat, CN), 具身宝宝研习社 (Embodied Baby Study Society, Chinese industry outlet) source (WeChat, CN)
Stanford AI Index: humanoid robots complete 1,000 real-home chores at only 12.4% success rate · humanoid
According to Stanford's 2026 AI Index Report, humanoid robots achieved only a 12.4% success rate on 1,000 real-home chore tasks, compared with 89.4% on the same category of tasks in simulation. This nearly sevenfold gap is the backdrop for several bankruptcy and cooldown events this week: usable robots are too expensive, and cheap enough ones don't have much use — the vast majority of robots currently sold go to labs, university departments, science museums, and data collection centers, with genuine production-line work still handled by a minority. Industrial-grade units start at several hundred thousand RMB, and enterprise payback periods often run four to five years. This also explains why capital is shifting toward the "brain": marginal improvements to the physical body alone can no longer lift this curve.Source: 首席财经社 (Chief Finance Society) source (WeChat, CN)
Weekly supply chain roundup: harmonic reducer localization rate in China hits 75%, the bottleneck is high-end consistency, not capacity · hardware
This week's supply chain signals form a single narrative: the volume problem is basically solved; the quality problem is just beginning. PwC's July white paper puts China's harmonic reducer localization rate at over 75%, up from around 30% two years ago; Leaderdrive's revenue from humanoid-specific reducers grew 215% year-over-year in H1, with orders booked through 2027, monthly capacity rising from under 50,000 units at the start of the year to 80,000 in June, with plans to reach 120,000 by year-end, while unit average selling price fell over the same period from around RMB 2,000 to the RMB 500 range — and the company has already entered Tesla Optimus's supply chain. Zhuoshi Consulting forecasts China's robotic harmonic reducer shipments will grow from approximately 1.4 million units in 2025 to 20.7 million units by 2030.⚠️ Research report claim
But the half-sentence in the white paper noting "clear advantage at low-to-mid end, breakthrough still needed at the high end" deserves more attention. Humanoid joints require repeatability precision an order of magnitude higher than industrial robots, and Chinese products still lag Harmonic Drive (Japanese manufacturer) by 15% to 20% in backlash and transmission error; design lifespan is around 8,000 hours versus over 10,000 hours for the Japanese benchmark, and batch-to-batch consistency variance also propagates into final assembly yield. Producing 80,000 units a month means shipping over 2,600 units a day while maintaining the same precision level — this is a manufacturing management problem, not a technical one, with no shortcut.
Two other sub-threads are moving in parallel: on the chip side, MediaTek is entering the 50TOPS tier with a 3nm platform, joining AMD's earlier Kria modules in eroding Nvidia Jetson's exclusivity, while Chinese vendors commonly "mix and match" domestic and Nvidia solutions across low- and high-compute boards; on the actuator side, South Korean firms are pushing forward — LG Electronics has moved up mass production of its in-house actuator Axium, and Samsung Electro-Mechanics has begun mass production of a humanoid-specific camera, while Nidec is repositioning its precision reducers toward humanoids — Japanese and Korean manufacturers are attempting to hold the high end and automotive-grade reliability segment beyond Chinese manufacturers' price advantage. This week's sharpest point of tension isn't capacity in any single segment, but batch-to-batch consistency at the high-precision end: it will determine whether China's supply chain keeps competing on price at the low-to-mid tier, or can cross into full market-wide competition within the two-to-three-year window.Source: 具身笔纪 (Embodied Notes, Chinese industry outlet) source (WeChat, CN), 武汉贝莱德咨询 (Wuhan BlackRock Consulting) source (WeChat, CN)
Top comments (0)