Today's Highlights

· AgiBot (Chinese humanoid maker)'s 20,000th robot rolled off the line and was delivered at Hengqin Chimelong, taking about 6 months to go from 10,000 to 20,000 units

· Waymo published 271 million miles of driverless data, showing 82% fewer injury-causing crashes than human drivers

· IFR: industrial robots in operation at factories worldwide reached 5 million units, with China accounting for 59% of 2025's new installations

· KAIST's quadruped robot RAIBO2 ran a full marathon on a single charge in 4 hours, 19 minutes, 52 seconds, with the paper published in Nature

· Black Forest Labs open-sourced the weights of its 7B-parameter robot model FLUX 3 Action, less than half the parameter count of Cosmos3-Nano-Policy

· Mech-Mind (Chinese industrial robotics vision company)'s first interim report since listing: new orders of RMB 335.4 million, up 75.3% year-on-year

RAIBO2: a quadruped robot that ran a full marathon on one charge · locomotion

Endurance has long been a hard constraint on deploying legged robots for wilderness rescue and mountain inspection. Starting from a whole-system energy-loss model, the KAIST team modified three things at once: a force-transparent lightweight mechanical structure, a low-resistance motor drive circuit, and a low-dissipation locomotion policy. RAIBO2 completed a full marathon on a single charge in 4 hours, 19 minutes, 52 seconds, with a total cost of transport of 0.25 versus a human baseline of 0.37 — more than three times the per-charge range of existing quadrupeds. Data has been posted to Zenodo, and the locomotion policy training code is public on GitHub.

KAIST team · Nature (2026), DOI 10.1038/s41586-026-11102-5 · Coverage: Nature source; The Register source

MemBodied: giving VLAs a fixed-length episodic memory · vla

HF 8↑. Most VLAs only look at the current frame, so they fail on tasks that depend on history — like "where did I just put that" — while stuffing historical frames into the context slows inference down. MemBodied replaces raw historical frames with a cross-call associative state plus an anchor recording the initial scene. Across five memory tasks on RMBench, average success rate is 7.81x that of stateless policies, while adding only one-tenth the parameters of the strongest memory baseline.

Tej Deep Pala et al. · arXiv 2609.28256 source

LeWAM: a world-action model that drops the video-diffusion backbone · world-model

Most existing world-action models (WAMs — models that jointly predict actions and environment evolution) are built on large video diffusion models, driving up training cost accordingly. The authors found that JEPA encoders' predictive embeddings suit action generation better than VAE-compressed latents, so they predict the future directly in I-JEPA embedding space and refine with offline DPO derived from demonstration-data preferences. Training only 0.4B parameters, it reaches 92.28% average success on RoboTwin 2.0, on par with mainstream VLAs and WAMs.

Xueji Fang et al. · arXiv 2609.27455 source

LADA: driving VLA needs under 5% language-labeled data · autonomy

Vehicle fleets log vast amounts of footage and trajectories daily, but frames paired with language instructions like "turn left at the intersection" are rare, making labeling costly. LADA first learns a discrete vehicle-intent codebook from unlabeled data, then trains a translator on a small labeled subset to map instructions into the codebook. On closed-loop Bench2Drive, it reaches a Driving Score of 87.98 and 70.46% success rate, matching or beating fully supervised baselines.

Alexey Zakharov et al. · arXiv 2609.27747 source

CereVLA: a "cerebellum"-style corrector for frozen VLAs · vla

When action chunks execute without mid-chunk feedback, errors accumulate. CereVLA doesn't retrain the main model — it adds a lightweight residual correction layer, with a state-space model forecasting each correction's outcome and suppressing ones judged harmful. On a real SO-101 robot, it raised frozen SmolVLA's task success rate from 57.5% to 90.0%.

Shuai Zeng et al. · arXiv 2609.27468 source

ChronoFuse: event-camera detection that predicts object position "at output time" · perception

Event detectors take tens of milliseconds to compute, so by the time results are delivered the scene has already changed — a lag standard benchmarks ignore. ChronoFuse directly predicts the target's state at the moment output becomes available, adding only 0.17M parameters and 0.84 ms latency, and recovers 71% of the accuracy otherwise lost to latency on 1Mpx driving data.

Biswadeep Sen et al. · arXiv 2609.26919 source

ForgetMimic: making humanoid policies "forget" specified motions · locomotion

Humanoid policies trained by imitating human motion can absorb poisoned or copyrighted motions, and GDPR's right to erasure can also trigger deletion requests. ForgetMimic, which the authors call the first motion-level unlearning method for real-robot humanoid control, was validated on 12 motions across Unitree G1 and H2, removing targeted motions while leaving the rest of the repertoire unaffected.

Xukun Luan et al. · arXiv 2609.28378 source

AnchorReasoning: a visual-grounding reasoning dataset for long-tail driving scenarios · benchmark

Built on WOD-E2E, containing 416,119 annotated frames and 395,379 decision-critical elements, each frame stitched into a visually grounded chain of thought that goes "identify key objects → infer effects → give a driving rationale → plan a trajectory." After fine-tuning on it, eight backbone models saw 5-second ADE and FDE drop by 7.84 and 11.86 respectively.

Zhipeng Bao et al. · arXiv 2609.28366 source

Shenpu Intelligence (Chinese embodied AI startup) ｜ controlling stake in *ST Meizhi ｜ RMB 810 million · embodied

Shenpu Intelligence, founded less than two years ago, is leading a restructuring investment consortium that plans to subscribe for about 92.86 million shares at RMB 8.725 per share, giving it roughly 27.45% ownership post-restructuring and making it the controlling shareholder; founder Li Xiaofei will become the actual controller. Fourteen other financial investors will separately contribute about RMB 999 million at RMB 10.765 per share, bringing the consortium's total to about RMB 1.81 billion. This is another case, beyond IPOs, of an embodied AI startup taking control of an A-share listed company.Source: Huxiu source; 21st Century Business Herald

Pusi Optoelectronics ｜ Series C ｜ hundreds of millions of RMB · hardware

Makes specialty power chips for extreme environments, rated to withstand ambient temperatures above 250°C, originally used in oil and gas exploration at 8,000-meter well depths. Ten new state-owned and industrial investors joined this round, with funds going toward gallium oxide chip iteration, expanded production of 230°C-class high-temperature chip modules, and development of high-temperature robot joints. 2025 revenue was RMB 230 million.Source: 36Kr source

Realset AI and Flatkey ｜ Series A ｜ $10 million · adjacent

Produces real-world training data, using skilled workers rather than crowdsourced labelers to collect demonstrations in real kitchens and laundry rooms. Plans to release an open household-task benchmark in Q4, testing π0, OpenVLA, GR00T, and Octo on folding laundry, loading dishwashers, sorting, and wiping tasks. Founder Hunter Guo said: "The easy-to-get data is already gone. What's left is the physical world, and the physical world can't be scraped."Source: PR Newswire source

O-ID ｜ Pre-seed ｜ $1.2 million · humanoid

A Tokyo-based company founded in 2025 by Norwegian CEO Stian Jakobsen and Dutch CTO Simon Gormuzov, building fully modular humanoid robots for manufacturing and logistics, where onboard sensors flag part wear ahead of time and worn components are swapped out as whole modules. Led by TAWANI Ventures, with participation from Hustle Fund and Techstars.Source: Tech.eu source; The Next Web source

AgiBot's 20,000th robot rolls off the line; over 300 now working routine shifts at Chimelong · humanoid

On September 24, the 20,000th Yuanzheng A3 Ultra robot rolled off the production line and was delivered to Chimelong Group at Chimelong Spaceship Park in Hengqin. The same day, the theme park the two companies jointly built opened, described by both as the world's first large-scale embodied AI theme park, with over 300 AgiBot robots across its full lineup handling tours, live shows, science education, retail, and hotel services on site. Production pace: passed 1,000 units in January 2025, reached 5,000 by year-end; topped 10,000 in March 2026, hit 15,000 on June 28, and took about 6 months to go from 10,000 to 20,000. The A3 Ultra stands 174 cm tall, runs on a 700 TOPS compute platform, and has a 20-degree-of-freedom tactile dexterous hand. The 20,000-unit figure is a cumulative count across AgiBot's entire lineup; the company did not break out how many are humanoid versus quadruped.Source: Kuai Technology source; Gasgoo source

Waymo's 271 million miles of safety data: 82% fewer injury crashes · autonomy

As of the end of June 2026, Waymo had accumulated 271 million fully driverless miles across Atlanta, Austin, Los Angeles, Phoenix, and San Francisco, about 50 million more than its previous update at end of March. Compared to human drivers, it saw 82% fewer injury-causing crashes and 95% fewer serious-injury-or-worse crashes; 93% fewer injury crashes involving pedestrians and 86% fewer involving cyclists. Waymo converted the reduction into an absolute figure for the first time: 841 fewer injury crashes. Electrek notes that independent research has also found reductions, though smaller than Waymo's self-reported figures; U.S. human drivers log 3 trillion miles a year, so the sample-size gap remains large.Source: Electrek source; Waymo official blog

Volvo's driverless mining trucks have hauled 3 million tonnes cumulatively · autonomy

Volvo Autonomous Solutions has hauled over 3 million tonnes of material driverless across European mines and quarries. At the Brønnøy Kalk quarry in Norway, 9 driverless trucks run a 5-km steep route through a long tunnel, having just passed 1 million tonnes in 2025. The service is billed as "transport as a service," with trucks built at Volvo's Säffle factory in Sweden.Source: Cision source

Kodiak completes first heavy-truck autonomous deliveries under California's new rules · autonomy

Kodiak partnered with produce carrier DTL Transport to deliver from Fresno to a Los Angeles distribution center, a one-way trip of about 225 miles via Highway 99 and Interstate 5, with a safety driver still in the seat. California's DMV only approved autonomous heavy-truck testing on public roads on April 28, and Kodiak received its testing permit on August 13. Its driverless operations in Texas are slated for later this year.Source: Investing.com source; Stock Titan source

Volkswagen's MOIA starts selling robotaxi tickets in Orlando · autonomy

MOIA and Beep are running three fixed routes in the Lake Nona community using ID. Buzz vehicles equipped with Mobileye's system — MOIA's first paid passenger service in the U.S. A human safety monitor remains on board; Finimize describes it as closer to a supervised pilot operation.Source: Finimize source; MarketScreener source

IFR annual report: 5 million robots now in operation at factories worldwide · industrial

IFR's World Robotics report shows the global installed base of industrial robots grew 9% in 2025 to 5 million units, more than double the figure from seven years ago; over 600,000 units were installed for the year, up 11%. China installed 354,000 units, up 20%, accounting for 59% of the global total; Chinese-made robots accounted for 195,000 of those installations, a 55% domestic-brand share within China, down from 57% in 2024. U.S. installations rose 12% to nearly 38,500, overtaking Japan as the second-largest market. Japan's installations fell 19% to 36,219, while Germany installed under 25,000, accounting for 41% of EU installations. IFR forecasts 2026 installations will grow 9% to 655,000.Source: IFR source; Business Wire source

Mech-Mind's first interim report since listing: orders growing faster than revenue · industrial

First-half revenue was RMB 237.3 million, up 54.7% year-on-year, with gross margin rising from 61.4% to 65.0%. New orders totaled RMB 335.4 million, up 75.3%; the order backlog at end of June was RMB 176.5 million, roughly double the year-earlier figure. R&D spending rose 72.9%, and the adjusted loss narrowed from RMB 55.8 million to RMB 52.5 million. Mech-Mind listed on the Hong Kong Stock Exchange on September 1.Source: TradingView / PR Newswire source

Toyota estimate: $6.4 billion a year from 2028, needing about 400,000 robots · industrial

Toyota told investors that factory automation would require roughly 400,000 robots, including both replacement units and new humanoid and non-humanoid robots, with over a third of the investment going to Toyota's own vehicle-manufacturing operations; spending would run about $6.4 billion a year starting in 2028. CNBC notes Toyota has not yet committed to this investment or specified how many years it would run. The demonstration model, ELEY, is a wheeled dual-arm robot taught via teleoperation by workers wearing exoskeletons. CTO Hiroki Nakajima said the robot can learn human movements through cameras and sensors, "so a skilled worker's technique can be transferred to the robot without writing complex programs."⚠️ forward-looking planSource: Repairer Driven News citing Automotive News source; CNBC

Amazon building a fourth robotics factory in Indiana · industrial

The Wall Street Journal reports Amazon will invest over $100 million in a robotics manufacturing facility in Greenwood, Indiana, to come online by 2028 and create 300 jobs. Combined with the Austin, Texas facility announced last month, this doubles Amazon's robotics factory count.Source: WSJ source

Skild AI trains a soccer-playing humanoid via simulated self-play · world-model

Skild AI says its S1 model self-played in NVIDIA Isaac Sim with the sole objective of scoring goals, without hand-designed rewards or demonstrations — equivalent to roughly 140 years of simulated experience — and that the policy transferred directly to real hardware, able to play against humans and other robots. Only a company-released demo video is currently available.⚠️ vendor claimSource: Crypto Briefing source; Zhidongxi source

Light Origins releases Light-O1, pretrained on human video · world-model

The model recovers structured human motion from internet video as a prior, then adapts it to different robots. The company ran experiments across six pretraining scales from 3.75B to 120B tokens, with the largest equivalent to about 100,000 hours of human motion, and says larger scale reduces adaptation error across two humanoid platforms. In its demo, the same model drives both the company's own LightBot and the Unitree G1.⚠️ vendor claimSource: PR Newswire source

AgiBot's two side moves: a new Shenzhen company, and a Malaysia investment MOU · humanoid

Shenzhen Yueying Robotics Technology was established with registered capital of RMB 4 million, with Peng Zhihui (AgiBot co-founder) as legal representative, wholly owned by AgiBot Innovation. Malaysian Prime Minister Anwar Ibrahim said AgiBot intends to invest 1 billion ringgit locally by 2035, creating 5,000 jobs.⚠️ forward-looking planSource: National Business Daily source; Kwong Wah Yit Poh source

Driverless truck warning-device exemption challenged in court; FMCSA responds · autonomy

On July 9, FMCSA approved Aurora's exemption allowing driverless trucks to skip having a driver exit and place warning triangles when stopped, using rooftop warning lights on the cab instead; other operators can apply for the same exemption after filing notice. On September 8, an Illinois van driver running the same route in Texas filed for review with the Seventh Circuit Court of Appeals and sought a stay, and the court ordered FMCSA to respond by September 21; Overdrive reports FMCSA has since filed its rebuttal. This exemption is currently one of the preconditions for driverless trucks operating at night in the U.S., and the court's ruling will directly affect Aurora's and Kodiak's year-end driverless operation timelines.Source: Overdrive source; Land Line

Two quadruped-robot market reports released the same day, differing by 14,000 units · embodied

On September 21, Counterpoint and IDC both released global quadruped robot reports. Counterpoint put first-half shipments at about 35,000 units, with 46,000 for full-year 2025 and industry revenue near $700 million; IDC put first-half shipments at over 49,000 units, up 92.5% year-on-year, with a 2025 market size over $490 million. According to an AIX Finance analysis reposted by Huxiu, the divergence comes down to methodology: Counterpoint's revenue figure includes sensors, software, and maintenance and excludes small companion-type products for children, while IDC counts hardware only and includes consumer-grade units. Both rank Unitree first, with second place going to AgiBot Kuotuo (by shipment volume) or DEEP Robotics (by revenue), respectively. Counterpoint expects 2026 shipments to exceed 90,000 units for the first time, with revenue doubling to $1.5 billion.Source: Huxiu source; ifeng Tech source

Goldman Sachs raises humanoid robot forecast to 890,000 units by 2030 · humanoid

Goldman Sachs's global physical AI report, published August 31, raised its 2026 humanoid shipment baseline from 51,000 to 75,000 units, its 2030 forecast from 256,000 to 890,000, and its 2035 forecast to 6.5 million; average unit price is projected to fall from about $41,800 in 2025 to about $21,300 by 2035. This is a single firm's projection, and differs by an order of magnitude from IFR's reported global humanoid sales of about 7,000 units in 2025.⚠️ projectionSource: Wall Street CN; Tencent News

This week in the supply chain: contract manufacturing and tactile sensing both expanding · hardware

This week's supply-chain signals cluster around two layers. The first is contract manufacturing. Futekang, a joint venture between Foxconn and Altar (Chinese contract manufacturer), each holding 45%, launched on September 21 in Beijing's Yizhuang district, offering turn-key R&D, manufacturing, data training, and sales services to robotics companies. Leadyoung Intelligence's Yizhuang "super factory," visited by 21st Century Business Herald on September 23, plans 10,000 unit-sets in 2026, 20,000 in 2027, and 500,000 by 2030; per-unit assembly time is currently 30–40 hours, with a target of cutting it to 20 hours. The second layer is tactile sensing. Tashan Technology (Chinese tactile-sensor maker) completed a shareholding restructuring, and CEO Ma Yang said there are IPO plans; the company says it ships tens of thousands of tactile sensors monthly, with first-half orders more than four times all of last year's. Joyson Electronics (Chinese auto-parts supplier), one of the investors in Tashan's July Series B, also appeared on the list of suppliers auditing for Tesla's Optimus, reported September 18.⚠️ vendor claimSource: 21st Century Business Herald; Tencent News; Sing Tao Global Network