Today's Highlights

· Li Auto is pushing a distilled version of Mach VLA 2.0 to already-delivered AD Max vehicles, covering nearly a million older Orin-X and Thor-U cars.

· News that regulators plan to raise the IPO bar for humanoid robot companies sent Unitree Robotics (Chinese humanoid robotics maker) below RMB 500 intraday, wiping out over RMB 240 billion from its peak market cap on listing day.

· Nine days after listing, Mech-Mind (Chinese industrial robotics/3D vision company) CEO Shao Tianlan publicly accused peers of using data-collection centers and related-party transactions to fabricate revenue, naming Galbot (Chinese embodied-AI robotics startup) in the comments.

· Pony.ai and Verne launched Europe's first fully driverless Robotaxi passenger pilot in Zagreb, on a 22-kilometer route.

· WeRide (Chinese autonomous driving company), Uber and AVOMO secured Spain's first L4 passenger vehicle operating permit; the company's Q2 overseas L4 revenue grew 164% year over year.

· Maven Robotics emerged from stealth with $100 million, with eight mixed-case palletizing robots already working 16 hours a day in a customer warehouse.

Show-Harness: give a VLM a semantic interface and it can drive a robot directly · vla

Turning frontier closed-source VLMs into zero-shot robot controllers doesn't require a new model or embodiment-specific pretraining — just an interface layer. Show-Harness exposes a set of discrete semantic action units for the VLM to reason over, while an embodiment-side interpreter deterministically turns them into local actions, leaving fine-grained physical decisions to the VLM; the same interface spawns GUMI, which lets people collect demonstration data on a graphical interface without teleoperating hardware. Small open-source VLMs need only a few GPU-hours of fine-tuning to deploy, and outperform agentic and VLA baselines across tasks, embodiments, and environments. HF trending score 60+.

Yanzhe Chen et al. · arXiv 2609.10522 source

First agile humanoid locomotion on real granular sand · locomotion

Granular terrain is hard because foot-ground interaction can't be modeled cleanly — prior approaches either ignore granular contact or bolt on an ad-hoc tangential component to the normal force. This work turns 3D Resistive Force Theory (3D RFT) into a physically correct granular contact solver for reinforcement learning training, exposing real intrusion and tangential drag during training, then uses a teacher-student setup with a variational autoencoder to compress terrain information into a latent for online adaptation. Hardware was validated on basalt, dry sand, and beach sand; the authors say this is the first time agile humanoid locomotion has run on real granular terrain.

Junnosuke Kamohara et al. (incl. Daniel I. Goldman) · arXiv 2609.10286 source

FolDeX: a 2,000-hour real-robot laundry-folding benchmark that accepts external policy submissions · benchmark

Policies that work well in simulation degrade on real hardware, and long-horizon deformable-object manipulation degrades the most, while most existing real-robot benchmarks stay confined to short-horizon rigid-body tasks. FolDeX is built entirely from real-robot data, centered on laundry folding, covering over 2,000 hours of data, 20-plus tasks, and 10-plus embodiments, organized around four axes: in-deployment human intervention and recovery data, cross-task and rigid-to-deformable transfer, cross-scene reuse, and cross-embodiment transfer. The accompanying evaluation platform, hosted on Midea's (Chinese appliance maker) domain, accepts externally submitted policies under a unified initialization and execution protocol.

Chenhuan Liu et al. · arXiv 2609.10243 source

No free judge: a survey of 150 robot-policy verifiers · benchmark

VLAs use verifiers both for evaluation and for training; this paper lays out roughly 150 verifiers along two axes — accessibility (how cheap, early, and dense the judgments are) and trustworthiness (how much a high score actually means). Verifiers are grouped by who issues the judgment: human, rule-based/formal, learned/pretrained, and model-intrinsic — and the finding is that all four groups show the same trend: as accessibility rises, trustworthiness falls. The paper closes with nine metrics for making verifier claims verifiable.

Yang Wan et al. · arXiv 2609.09250 source

One hand assembles two parts together · manipulation

The hand holding a pen naturally snaps the cap on with the same hand — this kind of finger division of labor is a hallmark of human dexterity. The authors use a reinforcement learning framework driven by the relative pose of two target parts to unify in-hand assembly without a second hand or a fixture; finger coordination is shaped by a functional assistive reward and regularized toward a single human reference pose, while domain randomization combined with historical fusion of proprioceptive and object observations resists estimation noise from occlusion. The same recipe solves three assembly tasks — bottle cap, syringe, and marker — trained purely in simulation and transferred zero-shot to real hardware with just one camera.

Liuao Pei et al. · arXiv 2609.10137 source

Generated video as motion reference, 1,500 clips fed into simulation · manipulation

Generated hand-object interaction videos are a convenient source of motion plans, but tracking on the simulation side hasn't scaled, bottlenecked by the lack of reliable reference motion. This method mass-produces references using near-automatic hand-object interaction reconstruction, feeding over 1,500 generated clips into simulation, boosting training-phase success rate by more than 25 percentage points over the baseline. In real-robot closed-loop tests it achieved functional grasping, non-prehensile manipulation, and post-grasp pose tracking.

Tianyue Wu et al. · arXiv 2609.10050 source

JEPA Policy: imitation learning without diffusion, only 0.29 ms extra latency · manipulation

Behavior cloning only supervises actions, without constraining the future representation paired with each action chunk. JEPA Policy treats the action chunk and its observed future representation as a paired training target, letting the two interact inside a shared Transformer and refine over two forward passes, so future prediction in turn shapes the representation used to generate actions. Across nine simulated tasks it achieves higher average success rate than the MIP baseline and outperforms Diffusion Policy, at the cost of adding just 0.29 ms of latency over MIP; five real-robot tasks with 630 episodes show the same ranking.

Jie Xu et al. · arXiv 2609.09630 source

SwingBot: a humanoid robot swings continuously across a bar using wrist hooks · locomotion

When ground passage is blocked, primates go up high — porting that motion to a high-degree-of-freedom humanoid is hard because the controller must discover the long-horizon sequence of releasing, swinging, and grasping on its own, without reliable measurements of inter-segment displacement or hook contact. SwingBot uses bio-inspired keyframes to make these rare transitions reachable early in exploration, then uses recurrent privileged state estimation to produce compact position and contact latents for deployment. Hardware experiments achieved continuous bar-crossing, remaining stable under payload, external disturbances, and varying bar spacing.

Yujie Xiong et al. · arXiv 2609.10283 source

Maven Robotics | Seed, out of stealth | $100 million · industrial

Led by RoboStrategy, with participation from LocalGlobe, Vine Ventures, and XTX Markets Ventures. The company works on mixed-case palletizing: full pallets arriving from different factories reach a distribution center, and robots reassemble them into mixed pallets by store order — work currently done entirely by hand, item by item, in warehouses. CEO Hamza Derbas said that in 2024 the company had "a cartoon drawing of a robot and a team of people," and won business away from four competitors that already had working hardware by visiting customers' factories and warehouses first to study the workflow before pitching a solution. Two years later, eight robots are working 16 hours a day at customer sites with over 99% uptime; the robots' wheeled base moves at 10 mph, and each of its two arms can lift 30 kg. This round funds building 250 third-generation machines and starting fourth-generation design. Source: TechCrunch source

Inspiren | Series C | $70 million | valued at over $500 million · adjacent

Led by NewView Capital, with participation from Insight Partners, Primary, Scale Venture Partners, and others, bringing total funding to $225 million. The product links environmental sensing in a room with residents' longitudinal data, letting care teams know the moment a fall or medical emergency happens, rather than after the fact. North America's over-85 population is its fastest-growing age group, and this round bets on the real-world caregiver-shortage curve. Source: PR Newswire source

Algomatic Dynamics (Japan) | Seed | $32.5 million · embodied

Funded by DMM.com; the company officially launched on September 9, reorganized out of DMM's Algomatic unit, with Yuki Nanri as CEO. Its three business lines are motion data and robot learning, AI robotic hands, and AI entertainment robots; a multi-fingered hand is planned for launch in Japan by end of 2026, with tactile sensing hardware and a full platform spanning task data collection, model training, and deployment. Its video analysis aims to capture what skilled workers can't articulate: motion timing, force magnitude, and where the eyes look. Source: AI Insider source

Vecna Robotics | Series D | $31 million · industrial

Led by Unless, with participation from Drive Capital, Tiger Global, Highland, and Tectonic. This Waltham, Massachusetts-based maker of autonomous forklifts and pallet trucks ties this round directly to demand following the FCC's July ban, which blocked internationally made AMRs and AGVs commonly used in logistics, shifting procurement attention toward US-made capacity. CEO Karl Iagnemma's long-term framing: North America has about 15 billion square feet of warehouse space, the vast majority still without material-handling automation. Source: DC Velocity source

Qianjue Robotics | new round | tens of millions of RMB · hardware

Strategic investment from an embodied-AI industry fund and Chigo Home Appliances (Chinese appliance maker). The company was founded in May 2024 by Ma Daolin, an associate professor at Shanghai Jiao Tong University and winner of the ICRA 2021 Best Paper Award; its product line spans multimodal tactile sensors, data collection systems, and a VTLA model that folds touch into vision-language-action, and it has served over 300 leading industry clients. Its full-hand solution extends sensing from fingertip to palm and finger pad, with fingertip marker density of 125 per square centimeter and force resolution of 0.03 newtons. Source: Robot Lecture Hall source

Qingyan Technology | Series A | over RMB 100 million · world-model

Led by Legend Capital, with participation from Ningbo Yishi and Suzhou Venture Capital Group. The company was co-incubated by Tsinghua University and the Yanqi Lake Beijing Institute of Mathematical Sciences and Applications, and for four years has quietly worked on data infrastructure and the intersection of mathematics and AI, entering world models through differential geometry and mathematical physics. Chairman Sun Mingming and board member Tang Ke are both researchers at the Yanqi Lake institute, with Tang Ke also a Tsinghua professor. As the term "world model" is currently used simultaneously by video generation, VLA, and autonomous driving companies, this team of mathematicians wants to redefine it more narrowly from first principles. Source: 36Kr source

Li Auto pushes distilled Mach VLA 2.0 to nearly a million older vehicles · autonomy

Li Xiang (Li Auto founder and CEO) announced on the evening of September 10 that, after months of model distillation, platform adaptation, and multiple rounds of validation, Mach VLA 2.0 has begun rolling out to vehicles that have already been delivered with AD Max, covering nearly a million owners. He credited the feasibility to long-term investment in self-developed chips, self-developed models, the self-developed Xingxuan OS, Harness engineering, and compilers, and said "the nearly one million owners of Orin and Thor platforms are our most valuable asset." The same OTA 8.6 update also brings command-parking for the Li Auto AI glasses Livis. It applies to the 2024-2025 MEGA, i-series, and 2022-2025 L-series, with some features limited to the AD Max version. Compressing a new model into two generations of older compute platforms and covering a million-plus installed base at once has no prior precedent at this scale.Source: Ifeng Technology source

UBTECH signs overseas orders worth over RMB 50 million, including Walker C1 and Youshijie U1 · humanoid ⚠️ Company's own figures

Following the first delivery of 13,361 pre-ordered U1 units scheduled for September 16, the company says it has recently signed a series of deals with customers in Europe, Japan, and South Korea, with order products including the commercial service humanoid robot Walker C1 and the hyper-realistic humanoid Youshijie (UBTECH's humanoid brand) U1. What can be verified is the half-year report: revenue of RMB 1.27 billion in H1 2026, up 104.2% year over year; total humanoid robot sales of 16,123 units, up 268.3%; of which full-size embodied-AI humanoid robots brought in RMB 590 million in revenue on sales of 921 units. The order value and client list come from the company's WeChat account, with no third-party disclosure found. Source: China Securities Journal source

Pony.ai and Verne launch Europe's first fully driverless Robotaxi passenger pilot in Zagreb · autonomy

The test route runs 22 kilometers, connecting Verne's headquarters with Zagreb's main business district and Franjo Tuđman Airport, with plans to gradually expand across the full operating area. This comes just over five months after the three parties (Pony.ai, Verne, Uber) launched Europe's first commercial Robotaxi service — that fleet has logged over 200,000 kilometers and completed thousands of rides since going live in April, with an average passenger rating of 4.7. The vehicles are Pony.ai's seventh-generation Robotaxi, with a domain controller built in-house on the NVIDIA DRIVE AGX platform, running safety-certified DriveOS. Verne CEO Marko Pejkovic called it a key step from safety-operator-supervised operation to fully driverless. The company's stated plan — over 4,000 vehicles deployed overseas/internationally and a global fleet of over 3,500 covering 20-plus cities by the end of 2026 — has not yet materialized. Source: Gasgoo source

Ant Group's LingBot robot sorts medicine at a real Shanghai Guoda Pharmacy store · embodied

In a retail pharmacy scenario set up at Inclusion Conference, a robot powered by LingBot's general-purpose brain autonomously handled order receipt, medicine identification, sorting, and delivery in an 80-centimeter-wide narrow aisle, without requiring changes to the store's existing layout or medicine placement. Beyond the demo booth, the robot has already begun working in real Shanghai Guoda Pharmacy stores, handling medicine sorting, with more pronounced value during night shifts. The same brain is also used for random package sorting in logistics warehouses and industrial loading/unloading.Source: Sina News source

WeRide, Uber, and AVOMO secure Spain's first L4 passenger vehicle operating permit · autonomy

The three parties, with support from the Madrid government, obtained a permit from Spain's General Directorate of Traffic under the ES-AV framework, allowing deployment of vehicles on public roads. AVOMO is the autonomous driving business unit under the Moove Cars group. WeRide already ran an autonomous shuttle pilot with Renault in central Barcelona in March 2025. Per its Q2 earnings, the company has now obtained autonomous-driving-related permits in 8 countries, with overseas/international L4 revenue up 164% year over year for the quarter.Source: Guandian.cn source

Zoox built a factory with annual capacity for 10,000 units — regulations only let it deploy 2,500 a year · autonomy

On July 31, NHTSA issued its first Part 555 commercial exemption for a purpose-built Robotaxi model, waiving parts of eight federal motor vehicle safety standards including windshield wipers, defrosting, and rear visibility, while capping volume at 2,500 vehicles in any 12-month period. The Hayward factory built for these orders has rated annual capacity of over 10,000 units. On August 10, Zoox began charging fares in Las Vegas, priced to match comfort-tier ride-hailing; in California it still only holds a pilot permit, not a deployment permit, and under pilot rules cannot charge a cent. By comparison, Waymo has nearly 4,000 vehicles across 15 cities handling over 500,000 rides a week; Zoox has about 100 vehicles handling about 10,000 rides a week. Waymo's constraint is capital and permitting — in February it raised $16 billion at a $126 billion valuation; Zoox's constraint is written into motor vehicle safety law passed by Congress, which no amount of Amazon money can move.Source: Business Model Analyst source

Reports of tighter humanoid IPO bar; Unitree falls below RMB 500 intraday · humanoid ⚠️ Unconfirmed reporting

According to The Information, citing people familiar with the matter, regulators have recently given informal window guidance to some investment banks and investment institutions, requiring humanoid robot companies planning to go public to demonstrate they can generate sustained revenue, are narrowing losses, or hold genuinely significant technological innovation. The China Securities Regulatory Commission has not publicly responded. The secondary market reacted immediately: Unitree Robotics fell intraday to RMB 498.7, its first drop below RMB 500 since listing, with the intraday low corresponding to a market cap of about RMB 202.2 billion — down over RMB 240 billion, or nearly 55%, from its RMB 444.9 billion peak market cap at the opening-day high. The company's IPO price was RMB 150.8, with an 85.28% turnover rate on listing day. H1 revenue was RMB 1.152 billion, up 48.5% year over year, with net profit attributable to shareholders of RMB 274 million; non-GAAP net profit was RMB 244 million, down 19.34% year over year. The same RMB 500 level represents a roughly 50% loss for anyone who bought at RMB 1,100 on listing day, but still a return of hundreds of times for early-stage investors who entered a decade ago.Source: Huxiu source, Sina Technology source

Shao Tianlan accuses peers of fabricating revenue via data-collection centers, names Galbot in comments · embodied

Mech-Mind (Chinese industrial robotics/3D vision company) founder Shao Tianlan posted two consecutive statements on WeChat Moments, saying many embodied-AI companies are using so-called "data collection centers" and related-party transactions with local governments, investors, and suppliers to manufacture fake, unsustainable revenue, calling the practice "illegal, unethical, and not even smart." His reasoning: a company that manufactures such revenue this year must not only repeat it next year but do more of it, falling into a spiral of continued fabrication and cash burn. He also said "the secondary market is not a 'venture capital' market, and it shouldn't become one." In the comments he directly named Galbot, asking, "Could any Beijing municipal leader step forward and say they support such a company going public?" Galbot had not responded as of press time. Shao's own company listed on the Hong Kong Stock Exchange on September 1 at HK$101.70, with public offering subscription during the IPO reaching about 3,835 times oversubscribed; the stock broke below its offer price on its first trading day and closed at HK$82.25 on September 9, roughly 19% below the IPO price. The company posted cumulative net losses of over RMB 1 billion from 2023 to 2025, including a RMB 360 million net loss in 2025; excluding non-cash items such as changes in redeemable liabilities, adjusted net loss narrowed from RMB 334 million to RMB 109 million. When the Shanghai Stock Exchange revised its STAR Market listing rules in June, it named robotics and embodied AI as a priority support area while also stating it would "strictly tighten the listing access gate."Source: 36Kr source

Unitree unveils 6-billion-parameter embodied foundation model UnifoLM-WLA-1.0, one model handling 64 tasks · world-model ⚠️ Company's own figures

The model was trained on about 2,500 hours of high-quality real-robot data, with a single set of weights covering 10 whole-body manipulation tasks and 54 tabletop manipulation tasks; the embodied-reasoning side, UnifoLM-ER-1-4B, is built on Qwen3-VL-4B, and among 16 published multimodal benchmarks it scored highest among participating open-source models on 7. These results were self-reported by Unitree; the project's homepage still shows "Coming soon" under Code, Models, and Datasets, so external verification of generalization beyond the 64 tasks isn't yet possible.Source: ZDNet China source

ShengShu Technology unveils Motus2, folding action evaluation into a video-action model · world-model

At Inclusion Conference, ShengShu Technology (Chinese video-generation AI company) co-founder and CEO Luo Yihang unveiled Motus2, a self-evolving general-purpose world model for robotic dexterous manipulation. It handles policy, world simulation, and value evaluation together within a unified video-action model: proposing an action, predicting its outcome, then judging whether the current action moves closer to the goal, using that feedback to refine subsequent policy.Source: The Paper source

After Tiangong Ultra beat Bolt, the team shifted focus to reliability · humanoid

At the World Humanoid Robot Games in August, Tiangong Ultra ran the 100-meter dash in 8.64 seconds, nearly a second faster than Usain Bolt's world record, the result of all-night testing and repairing a machine worn down between two races. Jack Guo, the 34-year-old engineer leading development of Ultra and Omni, told Reuters: "We want to turn high performance into high reliability, high availability." His organization, the Beijing Humanoid Robot Innovation Center (X-Humanoid), was founded in November 2023 with backing from UBTECH, Xiaomi, and Baidu; it both supplies hardware, motion-control software, and AI to the industry and sells its own robots. The center has not disclosed sales or commercial deployment figures for its own robots, with customers mainly universities, research institutions, and companies exploring industrial applications.Source: Reuters source

COMAC plans to bring humanoid robots into large-aircraft production workshops · industrial ⚠️ Stated plan

He Dongfeng, Party Secretary and Chairman of COMAC (Commercial Aircraft Corporation of China), said at the 2026 China Unicom Partner Conference that the company will build AI training grounds and construct a cloud-edge-device coordinated final-test environment to bring humanoid robots into large-aircraft production workshops. No timeline, aircraft model, or scale was given.Source: Guandian.cn source

CoreWeave launches Physical AI Field Engineering · adjacent

The service pairs customer teams with domain experts, covering the full engineering cycle from R&D to field operations, drawing on the team and methods from CoreWeave's acquisition of Monolith AI, running on its own platform and already-integrated engineering AI solutions. GPU cloud providers are beginning to sell services into the physical-world data side of the business.Source: CoreWeave source

FCC adds foreign-made robots to Covered List; industry group asks for a narrower scope · industrial

On July 28, the FCC added foreign-made connected inverters and advanced robotics equipment (certain mobile robots) to its Covered List, meaning new models cannot obtain equipment certification — effectively banning import, marketing, and sale in the US unless a waiver is granted. "Foreign" is determined using the "domestic end product" definition in the Buy American Act, 48 CFR § 25.101(a). The ban is prospective only: previously certified models and already-purchased equipment remain usable; companies can apply for conditional approval, reviewed by the Department of War and Department of Homeland Security. The rule is based on two national security determinations that a White House-convened interagency executive branch group submitted to the FCC on July 27. This week the tech industry group ITI publicly called on the FCC to take a targeted, risk-based approach; policy and general counsel John Miller said: "Security and innovation are not competing goals. Technology shouldn't be treated as inherently risky just because it's manufactured outside the United States." ITI opposes blanket mandatory hardware and software bill-of-materials disclosure requirements, and opposes mandatory registration for all equipment following the supplier's declaration-of-conformity path. The rule's commercial fallout is already visible: Vecna Robotics' $31 million Series D this week, which the company itself tied to demand shifting after the ban.Source: Supply Chain Digest source

National standard for dexterous hands expected to be published for comment by year-end · hardware

Multiple dexterous-hand companies have already participated in drafting the standard, which is expected to take shape and be published for public comment by end of 2026. Earlier, the China Academy of Industrial Internet and other bodies had already released a first batch of group standards, including "General Technical Requirements for Humanoid Robot Dexterous Hands" and "Humanoid Robot Motor Capability Grading." A finalized national standard would directly affect procurement in this segment — until now, degrees of freedom, payload, and lifespan have all been self-reported by manufacturers, making apples-to-apples comparison impossible.Source: Jiemian News

Interact Analysis cuts robotic picking market forecast: $4.6 billion by 2030 · industrial

A report released August 19 puts global robotic picking revenue at $1.7 billion in 2025, rising to $4.6 billion by 2030, an average annual growth rate just under 22%. The structural shift matters more than the total: palletizing and depalletizing accounted for roughly 83% of the share combined in 2025, falling to about 58% by 2030; bin-picking rises from $125 million in 2025 to $577 million in 2030, growing about 36% annually. The report attributes long-term expansion to tightly coupled solutions making deployment easier, and notes that robot makers and warehouse automation vendors are putting money into perception and motion software, which is becoming the differentiator. This $4.6 billion figure is lower than the firm's previous $6.8 billion forecast.Source: Interact Analysis

Samsung SDS sets a 10 trillion won investment plan through 2031 · adjacent

About $7.4 billion, broken down as roughly 5 trillion won for AI data centers, roughly 4 trillion won for M&A, and roughly 1 trillion won for AI services and platforms. Funding comes from a new 1.2 trillion won partnership with KKR plus 6.6 trillion won in cash and cash equivalents. At REAL Summit 2026, the company said it is reviewing several investments, including taking a stake in a robotics foundation-model developer, and named the United States as a primary market, citing manufacturing reshoring there.Source: Korea JoongAng Daily

Supply chain this week: money and capacity pushing into tactile sensing and actuators at once · hardware

Tactile sensing has seen an unusually large shift this year among components. Total 2026 funding in tactile sensing has already surpassed RMB 7 billion, and the tactile-sensor penetration rate in commercial humanoid dexterous hands has risen from 20% at the start of the year to over 60%; Qianjue Robotics' tens-of-millions-RMB round this week is just one point on that curve. What's driving it is application need, not technology preference: repetitive, low-precision work doesn't need touch, but precision assembly and compliant grasping — where force control and slip detection can't be done by vision alone — do. The technical approach is still fragmented — MEMS, flexible electronics, and visuotactile sensing each go their own way, with no single winner. On the actuator side, Robotis has defined its next-stage task as building up capacity and price competitiveness before scaling into mass production, taking a capacity-ahead-of-demand approach while also expanding from actuator hardware into humanoid software. Price pressure comes from another direction: a finished humanoid robot retails for hundreds of thousands to over a million RMB, but core components typically account for under 30% of that value, and still face price pressure from integrators. Suppliers of dexterous hands, joint modules, and force-controlled arms are responding by bringing their own complete robots to trade shows — Rokae's line at the show was that component specs are only meaningful in the lab, and once installed on a real wheeled base, ground unevenness and sudden load changes amplify the deviation; without a complete robot as a testbed, force-controlled joint iteration is just talking to a spec sheet. Aowei Technology's dexterous-hand orders surpassed 10,000 units in H1 2026, with its product line now expanding from biomimetic hands to an open-source dual-arm platform, industrial actuators, and lower-limb exoskeletons.Source: Robot Lecture Hall source, LanJinger source, Seoul Economic Daily