Today's Highlights

· Tesla engineers begin auditing Chinese suppliers including Sanhua Intelligent Controls, Joyson Electronics, and Topgroup; Optimus mass production will launch at the Fremont plant.

· Waymo names Singapore its next overseas market, with paid Robotaxi service planned for 2028; vehicles will arrive in the coming months.

· Daily rental rates for humanoid robots have fallen from tens of thousands of yuan to the thousand-yuan range; over 90% of rental stalls in Huaqiangbei (Shenzhen electronics market) have shut down.

· Midea's robotics and automation business posted revenue of RMB 16.62 billion in H1; KUKA's China industrial robot shipments exceeded 18,000 units, with its market share rising to 9.7% and second place.

AthenaZero: Bringing a robotic arm's end-effector effective mass close to a human hand's — to play baseball · manipulation

Conventional robotic arms have end-effector effective mass an order of magnitude higher than a human hand's, which is why they can't catch fast-incoming balls. RAI Institute used quasi-direct-drive actuation with a remote transmission to bring AthenaZero's end-effector effective mass down to roughly human levels while preserving torque transparency. Pitch speeds exceeded 30 m/s, with catching and hitting speeds over 14 m/s at a 7.3-meter distance; the system completed rallies both robot-to-robot and human-to-robot.

Andrew S. Morgan et al. (RAI Institute) · arXiv 2609.19194 source

Astronex-World 1.0: a 5B-parameter world model trained on two L20 GPUs, real-time streaming on a single card · world-model

While world models have spent the past year scaling up parameters, this one goes the other way, optimizing for density. Built on the Wan2.2-TI2V-5B prior, a five-stage training pipeline produces a controllable video world model whose causal version generates 832×480 frames in real time at 24 fps on a single NVIDIA L20. It scored 70.0 on WBench Full, ahead of the 13.6B LongCat-Video and 14B Helios, and within one point of the 22B LTX-2.3. All five training stages were completed using just two 48GB L20 GPUs, with action input/output interfaces reserved for post-training on embodied and autonomous-driving tasks.

Xin Zhou, Cong Miao · arXiv 2609.20034 source

VABench: general multimodal models score perfect on object localization but complete only half of full tasks · benchmark

With 7 trending points on Hugging Face, this benchmark tests the full "observe–reason–act–correct" loop: models get no privileged pose information, no demonstration trajectories, and no pretrained action head — they must choose their own camera viewpoint, issue metric-coordinate commands, and revise based on execution feedback. The best model achieved 100% on goal localization and 78.9% on spatial relations, yet the macro-averaged three-round task success rate was only 53.93%. Active camera control versus passive multi-view raised success on a control task set from 27.86% to 57.50%; switching to held-out geometric layouts dropped success by more than 30 percentage points, and no model completed the long-horizon strict tasks.

Zhongbo Zhang et al. · arXiv 2609.19554 source

Modifying only the collision mesh can make a policy that looks fine in simulation fail on real hardware · benchmark

A 3D asset in a simulator carries two geometries: a visual mesh for rendering and a coarser collision mesh to save compute. The two are allowed to diverge by design, and this paper shows that gap is an exploitable poisoning channel. An attacker who alters only the collision mesh — leaving the visual mesh and everything else untouched — can produce a policy that trains and evaluates normally throughout simulation, then degrades, fails, or creates physical risk once deployed on real hardware. Existing asset review processes check for malicious code, copyright, and format compliance, but not visual-collision consistency; the authors tested several defenses, none of which blocked the attack.

Gengyang Xu et al. · arXiv 2609.18122 source

Pruning 63% of a VLA's language backbone drops success to 0.8% — recovered offline in 8 GPU-hours · vla

A multi-billion-parameter language backbone is the main obstacle to running VLAs on robot hardware. Structured pruning that removes 63% of the OpenVLA-OFT backbone dropped LIBERO-Long success from 93.2% to 0.8%; prior fixes required online rollouts burning hundreds of GPU-hours. This work instead uses width pruning to preserve the residual stream's original dimensions, keeping teacher and student hidden-state shapes aligned so they can be matched directly — running the teacher once to build a cache, then distilling offline. In roughly 8 GPU-hours, the student was brought within 3.5 points of the teacher. On a 6-DOF robotic arm, a student at 72% compression achieved 77.5% success, with 2.23x faster onboard inference and 62% less memory usage.

Chiyoung Kim et al. · arXiv 2609.19579 source

Feel-WM: an off-road navigation world model that predicts not just what it will see, but how rough the ride will be · world-model

In urban scenes, predicting future frames is enough; off-road, success depends on how the robot and terrain interact. Feel-WM takes proprioception as input and jointly predicts future body states and failure risk, both learned from the robot's own experience without manual labeling; the planner folds predicted failure risk and goal similarity into a single separable score. Mounted on a Husky running mountain trails, it can anticipate rough ground ahead and route around it, while an end-to-end policy fails on the same course.

E-In Son et al. · arXiv 2609.19863 source

RoM-Nav: G1 does mapless cross-floor navigation with over 10 meters of vertical displacement in a single trial · locomotion

Single-stage RL navigation pipelines stall out on complex terrain interactions like stairs. The authors first train a navigation policy using reduced-order dynamics plus full 3D LiDAR observations, then use that navigation knowledge to warm-start a policy operating on full humanoid dynamics, with the locomotion control policy frozen in the loop. A Poisson safety filter layered on the navigation policy's output lets it recover safely from out-of-distribution obstacles without hurting success rate. On the Unitree G1, the system completed mapless multi-floor navigation with over 10 meters of vertical displacement and path lengths exceeding 100 meters in a single trial.

William D. Compton et al. (Caltech) · arXiv 2609.19272 source

Watney | Series A | $80 million · embodied

Co-led by Valor Atreides AI Fund and Hummingbird Ventures, with Conviction, Abstract, and A participating again, bringing cumulative funding past $100 million. Founded in San Francisco in 2023, the company focuses solely on data-center floor work — swapping cables, restarting servers — and has served hyperscale cloud providers since 2025. According to WIRED, Meta has been testing two dual-arm Watney robots at its Altoona, Iowa campus since last June, with trials also underway at the Prometheus data center in New Albany, Ohio; the robots are still slower than human technicians and require supervision, and one Meta employee told WIRED that a mature cable-swapping system could eventually eliminate as much as 80% of the workload for certain technician roles.Source: WOWTALE source

SynaCore | Pre-A round | nearly RMB 100 million · adjacent

Co-led by CMC Baipu and Kunpeng Fund, with participation from Guangzhou Industrial Investment and Yihe Capital, and existing investor Trend Capital increasing its stake beyond its pro-rata allocation. This is the company's third round in a month. The company runs two lines of business: SynaData converts human behavior into trainable data, and the company says it already serves over 60% of embodied-AI unicorns valued above RMB 10 billion; SynaTac unifies tactile data from different sensors and formats into structured representations for contact-policy training, and this round's funding goes mainly toward the latter.Source: Gasgoo source

Weiyi Intelligent Manufacturing | IPO filing approved for Hong Kong Stock Exchange · industrial

This Changzhou-based industrial embodied-AI company posted revenue of RMB 434 million, RMB 600 million, and RMB 800 million over the past three years, with gross profit rising in step from RMB 184 million to RMB 386 million; however, period profit fell from RMB 15.74 million in 2024 to RMB 5.07 million in 2025, while adjusted net profit over the same period ranged from RMB 44.15 million to RMB 48.73 million. As of the end of 2025 it held RMB 240 million in cash and cash equivalents. Baidu Online was formerly a shareholder and exited entirely across two sales in 2025.Source: Leidian (LeiPhone) source

AMC Robotics | standby equity financing | $50 million · industrial

Signed a $50 million standby equity purchase agreement and issued an additional $3.88 million in convertible notes, with proceeds earmarked for accelerating manufacturing capacity buildout. A standby equity agreement is a drawdown facility rather than a lump-sum disbursement; actual usage will depend on future disclosures.Source: citybiz source

Qingzhi Yuanshi | Series A+ | amount undisclosed · embodied

Participants included 5Y Capital and Hillhouse Ventures.Source: aggregated reports source

Waymo names Singapore its next overseas market, paid service to launch in 2028 · autonomy

Following last week's invite-only launch in Las Vegas, Waymo is now looking beyond the United States. Vehicles will arrive in Singapore "in the coming months," with mapping and autonomous-driving testing under safety drivers beginning in 2027, and a fully driverless robotaxi service open to the public in 2028. Before carrying passengers, the company must still secure approval from the Land Transport Authority; Singapore requires all autonomous vehicles to first pass a safety assessment from CETRAN, with testing protocols set jointly by LTA, CETRAN, and traffic police. Waymo established a local entity in Singapore earlier this year and has joined the country's autonomous vehicle steering committee. Co-CEO Tekedra Mawakana said in a statement that Singapore "has built one of the safest, most efficient, and most forward-looking transportation ecosystems in the world." The company currently operates about 4,000 vehicles across roughly 15 U.S. cities; previously announced overseas launches include London (late 2026), Tokyo (2027), and Munich (late 2027).Source: The Verge source

Zoox's 100-vehicle cap in Nevada expires this month · autonomy

The Nevada Transportation Authority's 100-vehicle permit cap expires on September 25, according to both NTA officials and a Zoox spokesperson who confirmed the change to TechCrunch; the updated permit reflects it as well. Zoox currently operates roughly 100 custom, steering-wheel-and-pedal-free robotaxis across four U.S. cities, with Las Vegas and San Francisco accounting for the bulk of the fleet — Las Vegas being the only city where it currently charges fares. Charging fares since last month was made possible by a temporary federal motor vehicle safety standard exemption from NHTSA, which allows Zoox to deploy up to 2,500 vehicles per year over the next two years. The spokesperson declined to say how many vehicles the company plans to add in Las Vegas, only that the fleet would grow steadily over coming months; Robotaxi Tracker estimates more than 30 vehicles are currently operating in the city. This gap between permit ceilings and deployment capacity has already played out once before, with Tesla.Source: TechCrunch source

Meiluo U completes two months on the job at a Midea washing-machine facility, having tightened over 100,000 screws · humanoid

Midea's flagship industrial humanoid, "Meiluo," has been deployed at washing-machine plants in Jingzhou and Wuxi and in sheet-metal work at a high-end facility, handling material recognition, path planning, grasping, and station feeding; a kitchen-and-water-heater plant in Wuhu is trialing it for standard-box transport and palletizing. The upgraded "Meiluo U" has run stably for over two months at the washing-machine high-end facility, completing more than 100,000 screw-tightening operations. The consumer- and commercial-facing "Meila X2" is reported by the company to achieve a task success rate above 95% in its target scenarios.⚠️ Manufacturer's claimSource: Sina Finance source

CAT autonomous mining trucks have hauled 3.5 million tons at a Virginia quarry · industrial

This marks a data point in autonomous mining trucks spilling over from mine sites into quarry operations. Industry sources say hundred-unit-scale deployments are becoming the new benchmark, with full-year shipments potentially exceeding ten thousand units.⚠️ Industry estimateSource: energytech.com source; finance.biggo.com source

Pudu's BG1 enters scale production, with units now rolling off the Wuzhen plant · industrial

This large commercial cleaning robot launched in late March this year, with the company saying pre-orders surpassed RMB 100 million within a month of launch. Units are now progressively rolling off the company's Tongxiang Wuzhen plant. Pre-orders are not the same as deliveries; actual pace will depend on future shipment disclosures.⚠️ Manufacturer's claimSource: Sina Finance source

AgiBot's A3 Ultra begins on-site work, claiming a 10-second battery swap and 8-hour runtime · humanoid

Both the battery-swap time and runtime figures come from the manufacturer; actual continuous operating duration would need to be verified against real on-site shift records.⚠️ Manufacturer's claimSource: EEFocus (Chinese electronics trade outlet) source

Figure's Helix 2.5 enters 30 unfamiliar homes, completes full tasks 56% of the time · humanoid

Figure 03 robots were sent into 30 homes in the San Francisco Bay Area that were not used in training, performing three task types — tidying a living room, folding towels, and making a bed. Across 420 trials, the full-task success rate was 56%, compared with only about 9% for a control version not pretrained on the Index human-behavior dataset. Industry figures including Sunday Robotics co-founder Tony Zhao have pushed back, arguing that a nearly one-in-two chance of failing a full task doesn't match claims of "already being able to work in real, unfamiliar homes"; the robot is also still performing three specifically trained tasks, not taking a single instruction and handling arbitrary new chores.⚠️ Manufacturer's claimSource: Tech AU source; Sohu source

Toyota plans to invest ¥1 trillion annually starting 2028 to overhaul factories, deploying 400,000 self-developed robots · industrial

Toyota Group plans to deploy 400,000 self-developed robots across its global production bases, with 150,000 at Toyota's own plants and 250,000 at group-company plants. This is paired with a plan to spend ¥1 trillion (about $6.42 billion) annually starting in 2028 to renovate and rebuild factories worldwide. Several media outlets have described the 400,000 units as "humanoid robots," but Nikkei's original reporting describes them as self-developed robots more broadly, with humanoid units — called ELEY — being just one category: a two-fingered, 50 kg unit that moves on wheels rather than walking, capable of drawing both battery and cable power; the name stands for Embodied Learning robot for Enhanced Yield. ELEY is already deployed on some production lines, where workers wear a fixture modeled on ELEY's fingers and perform their normal repetitive tasks while the robot learns by observation. At an investor technical briefing at Toyota's European headquarters in mid-September, one robot folded T-shirts for the audience, reaching near-zero errors after 1,500 repetitions learned over two weeks. Toyota operates 60 plants globally and employs 18,000 veteran skilled workers known as "takumi." Executive Vice President Hiroki Nakajima said: "What we're pursuing is coexistence between robots and people, not replacement."⚠️ Company planSource: Nikkei Asia source

Tesla begins auditing Chinese Optimus suppliers, with mass production based in Fremont · humanoid

Tesla managers and engineers have begun auditing part of its Chinese supply chain. According to two corporate executives with business ties to the companies involved, who spoke to the South China Morning Post, the suppliers under audit include thermal-management component maker Sanhua Intelligent Controls, auto-parts supplier Joyson Electronics, and chassis maker Topgroup — most of which already supply Tesla's electric vehicles and are all based in Zhejiang province. This round of audits corresponds to a new batch of component orders, with Optimus mass production set to launch at the Fremont, California plant. These Chinese suppliers have signed confidentiality agreements with Tesla carrying breach penalties in the millions of dollars. Chen Jinzhu, CEO of Shanghai Mingliang Auto Service Consulting, said: "The audits show these Chinese suppliers are an important part of the Optimus supply chain — they hold cost and technology advantages relative to peers both in China and abroad." In January this year, Musk announced the discontinuation of the Model S and Model X at Fremont, freeing up the plant for robots.Source: South China Morning Post source

Midea's robotics business posts RMB 16.62 billion in H1 revenue; KUKA's China shipments rise to second place · industrial

Midea Group posted H1 revenue of RMB 261.052 billion, up 3.46% year-on-year, with its robotics and automation business bringing in RMB 16.62 billion, up 10.27% year-on-year and accounting for 6.39% of total revenue — a faster growth rate than the smart-home business's 4.27%. KUKA's overall revenue grew more than 10%, with KUKA China contributing over 30% of that. Citing data from MIR DataBank (via Sina Finance and Weike Wang, a robotics industry outlet), KUKA's China industrial robot shipments exceeded 18,000 units in H1 2026 with a 9.7% market share, rising to second place domestically; its share of heavy-duty robots (300kg+) in China surpassed 50%. On core components, precision reducer shipments reached 26,000 units and heavy-duty servo motor shipments reached 58,000 units in H1. As of June, robot usage density at Midea's own factories exceeded 750 units per 10,000 workers.Source: Sina Finance source

Humanoid robot daily rental rates fall from tens of thousands of yuan to the thousand-yuan range as Huaqiangbei stalls empty out · adjacent

Last year, daily rental rates for a humanoid robot ran as high as RMB 30,000, and earning RMB 5 million over three months was considered only middling in the industry; by 2026, daily rates have fallen to the thousand-yuan range. According to Securities Times, over 90% of rental stalls in Huaqiangbei — a major electronics market in Shenzhen and China's hub for robot rentals — shut down in the second half of 2025. Price competition is only the surface issue; underlying it is a lack of stable genuine demand. Platform data shows entertainment performance and commercial marketing orders account for as much as 65% of bookings; IDC data shows that of global humanoid robot shipments in 2025, entertainment/performance use cases accounted for 36.8% and research/education for 24.6%, with the share actually performing routine production work in factories notably low. The asset side is similarly unfavorable, with annual depreciation of 20% to 30%: an engineering prototype bought for RMB 300,000 to 800,000 in 2024 now has a secondhand resale value of only RMB 30,000 to 60,000, and every rental job still requires an accompanying engineer on-site to prevent crashes. Light-industry orders currently make up about 10% of platform bookings.Source: Observer News source

Is robotics having its "ChatGPT moment"? Three different answers on the same day · world-model

Gao Yang (Spirit AI co-founder and chief scientist), gave a timeline: "We expect to reach GPT-3.0-level capability by mid-2027, where you can talk to a robot in natural language and it will execute a logical sequence of physical actions to complete a task." He also said "the brain really is the weakest link in the whole robot system." The company has hired around 1,000 outsourced workers nationwide to collect data — people at its Beijing training center wear sensors while repeatedly opening refrigerators and safes and chopping vegetables; in structured living-room settings, simple-task success rates are around 90%, and dozens of wheeled humanoid Moz1 units are already working on CATL production lines and in the JD.com retail network, with JD also an investor. Founded in 2024, the company has raised over $670 million to date at a valuation of RMB 20 billion. On the same day, Qianjue Technology founder Gao Haichuan took the opposite view: "Customers won't pay a single extra cent because we used an advanced predictive world model — they only care whether the job gets done well, how often it breaks down, and whether it's worth the price." He doesn't believe a single technological breakthrough will make the whole industry take off overnight; the company's first orders come from food service, cleaning, and hospitality robots. A third view came from the 9th China Robot Summit, where Soochow University professor Sun Lining said cerebellum- and limb-level products could reach a good level of maturity within three to five years of launch, but the "brain" is constrained by compute, algorithms, and data simultaneously, and approaching human-level general intelligence could take two to three decades; he cautioned that overly high expectations could erode industry confidence.Source: Reuters source; KrASIA source; 21st Century Business Herald source

Moonshot AI launches embodied-AI talent recruitment drive · embodied

The report also notes the company reportedly remains cautious internally about pursuing an agentic-robot roadmap.⚠️ Unconfirmed reportSource: Ifeng Tech source