Today's Highlights

· The first 6 Chinese national standards for humanoid robots have been released, taking effect on April 1 next year, requiring a success rate of at least 90% when grasping rigid objects

· South Korea has used a regulatory sandbox to clear 20 bipedal humanoid robots to work alongside workers on site at four companies in semiconductors, automotive, shipbuilding and logistics

· Cologne police seized a robot accused of sexually assaulting a 16-year-old girl, and are investigating whether it was remotely controlled in real time by two young men

· Frontier LLMs controlling a real robot arm take about half an hour per task on average, versus about 20 seconds for direct human operation

· Boston Dynamics has disclosed for the first time that assembling one Atlas now takes about one day, with a target of cutting it to one hour by 2027

· EACON (Chinese autonomous mining-truck developer) has removed safety drivers from 7 90-tonne mining trucks at a coal mine, which have run around the clock since the end of September

· After FoveaVLA removed the wrist camera, real-robot success rate rose from 52.1% to 77.1%

FoveaVLA: no wrist camera, the policy decides where to look next · vla

Wrist cameras see clearly but have a narrow field of view; external cameras see everything, but the image is compressed to low resolution before it reaches the policy. Built on π0.5, FoveaVLA predicts, with each chunk of actions, the interaction region to focus on next, then crops that region from the original-resolution image and feeds it back to the policy. Across four real-robot tasks, the single external RGB-D camera version reached a 70.8% success rate and the two-camera version 77.1%, while π0.5 with a wrist camera managed only 52.1%. The gains were concentrated in tasks with distractors, cloth manipulation, and distant objects.

Bingao Chen et al. (including C. Karen Liu, Guy Tevet) · arXiv 2610.20064 source

λ0: 500 hours of egocentric human data to teach humanoids to work while walking · locomotion

Previous egocentric human data mostly did not record walking or body posture. HumanVerse collected 500 hours of human data in open environments, with video synchronized to whole-body and hand motion. The team used it to do mid-training (a training phase between pretraining and fine-tuning) on the humanoid foundation model λ0. The Unitree G1 can transfer zero-shot without robot demonstrations of the target task. λ0 reaches a 90.0% success rate on SIMPLE and 62.5% on real-robot mobile manipulation tasks; the more whole-body human data, the higher the downstream task progress.

humanoid Team · arXiv 2610.18950 source

FACT: adding "force" to human demonstrations, fine-tuning dexterous manipulation with just 10 robot trajectories · manipulation

Most human demonstration data contains no force measurements. The team collected EgoForce, a human tactile dataset, and pretrained the dexterous manipulation foundation model FACT together with robot demonstrations. The model first predicts actions, future video and visual trajectories to make a coarse plan, then predicts them together with future contact forces to refine the actions. The tests cover ten force-feedback-dependent tasks, including tool use, playing musical instruments and delicate lab operations. Fine-tuned with only 10 robot trajectories plus 90 human trajectories, it beats T-Rex by up to 60 percentage points in success rate.

Zhi Cao et al. (including Yejin Choi, Li Fei-Fei) · arXiv 2610.19155 source

Contact-UAN: 6.5 minutes of real-robot data to fix force mismatch in humanoid sim-to-real transfer · locomotion

When a humanoid policy trained in simulation moves to a real robot, the motions look right but the forces do not match, and the problem often lies in the actuators. Contact-UAN keeps the simulator's existing physics priors and learns only a residual model to correct actuator dynamics. It also uses proprioception to estimate the stance-foot pose, so that replay starts from consistent contact forces. The model was trained on only 6.5 minutes of real-robot walking data; after transfer, the policy's running torque drops by 20% and force-tracking error by 43%.

Nolan Fey et al. (including Pulkit Agrawal, MIT) · arXiv 2610.18991 source

Discrete actions for continuous control, with success rates no worse than continuous actions · vla

Almost all robot foundation models today output continuous action vectors. This paper runs controlled comparisons across simulation and real robots, and single-task and multi-task VLA training: the plainest uniform-bin discretization can match or even exceed continuous actions in success rate. The authors argue that the weakness of discrete policies lies not in precision but in the factorization error that arises when multiple timesteps and action dimensions are predicted at once as a product distribution. With iterative discrete flow, or with classifier-free guidance, the discrete policy actually adapts faster within an action chunk.

Zeji Yi et al. (including Koushil Sreenath) · arXiv 2610.20429 source

80% success per manipulation step, only 19% once chained into long mobile-manipulation tasks · manipulation

The team sorted failures into three types: task planning, navigation and base positioning, and local manipulation. For each type, they added interventions to a fixed sub-task VLA, such as demonstration-based planning guidance, visual memory and base-reachability adjustment. Success rate rose from 19% to 47% on RoboCasa365 and from 14% to 37% on BEHAVIOR-1K, and real-robot tasks went from 5/20 to 13/20. The code is planned to be open-sourced.

Zixuan Wang et al. (Peking University) · arXiv 2610.19600 source

Plugging world models into VLAs: which way of doing it actually helps · world-model

Existing world-model + VLA approaches differ in representation, integration method and backbone all at once, making it hard to tell which factor is doing the work. This paper groups existing approaches into four paradigms and compares them one by one in a unified framework. 2D semantic representations give stable gains across paradigms; the decoupled MoT future expert is the most balanced overall and the most stable in out-of-distribution scenarios. World modeling helps most on long-horizon and dynamic tasks.

Qisen Ma et al. · arXiv 2610.19825 source

LidarVerse: one model for lidar scene generation, prediction and re-simulation · world-model

Lidar simulation has so far split into two types: reconstruction-based methods preserve geometry but need per-scene optimization; generative methods generalize but become inconsistent over long rollouts. LidarVerse merges the two into a single conditional generative model; by changing the input conditions, it can generate new scenes, predict future point clouds, or re-simulate already-captured scenes. After distillation it runs in real time on a single H100 at 11 FPS, compatible with 32-, 64- and 128-beam lidars.

Henry Che et al. (including Sanja Fidler, Shenlong Wang) · arXiv 2610.19201 source

TacClaw-Infinity: a vision-tactile-language-action dataset of over 10,000 hours · benchmark

Tactile manipulation data has been hard to collect at scale. TacClaw is a UMI-style handheld collector with fingertip tactile sensing that records vision, touch and end-effector trajectories synchronously without needing the target robot. The team collected data separately in homes, labs, factories, retail stores and outdoors; the dataset exceeds 10,000 hours, 1 million episodes and 1 billion frames, covering over 1,000 task types. On three different robots, it transfers without robot-specific demonstrations for each.

Fucai Zhu et al. · arXiv 2610.20320 source

Icon (US) ｜ Series D ｜ $200 million ｜ valuation undisclosed · industrial

Funds managed by Baillie Gifford and BlackRock co-led, with Tiger Global, Norwest and others participating. The Austin company builds 3D-printed construction systems, and the new funds are mainly for mass production and deployment of Titan, its multi-story building system. The company says it has built over 270 homes and other buildings across 5 US states and 2 countries, and that revenue this year is expected to reach 4 times last year's, but gave no figure. Source: AI Insider source

Yuehongyuan A (Chinese listed company) plans to acquire Jinwangda ｜ cash purchase of 100% equity ｜ up to RMB 550 million · hardware ⚠️ Planned figures

Jinwangda's production base is in Dongguan, and its products include precision planetary roller screws, ball screws, linear guides and reducers. With July 31, 2026 as the valuation reference date, the price is a premium of about 286% over the target's net assets attributable to parent shareholders, and goodwill of about RMB 400 million is expected to arise. The performance-commitment parties guarantee net profit of no less than RMB 36 million, RMB 41.5 million, RMB 47.5 million and RMB 55 million for 2026 to 2029 respectively, a cumulative total of no less than RMB 180 million over the four years. Yuehongyuan A's main businesses are real estate development, recycling of lead-containing scrap resources, and factory-building leasing. Only a letter of intent has been signed so far. Source: Duchuang Finance source

Ruoyu Technology ｜ Pre-A round ｜ hundreds of millions of yuan ｜ valuation undisclosed · embodied

Sunway Communication (Chinese electronic components maker) led, with Jinmu Capital, Qindao Capital, Huanghai Capital and Zhongchuang Capital participating. Founded in 2023 on the basis of Harbin Institute of Technology (Shenzhen), the company builds the "Ruoyu Jiutian" robot brain and in-house task hardware, starting from high-risk scenarios. It has obtained explosion-proof certification and is piloting at a gas station in Foshan. Cumulative funding has reached several hundred million yuan, with earlier investors including Dongshan Precision and Oriental Precision. Source: PEdaily source

Hurotics (South Korea) ｜ Series A ｜ 25 billion won ｜ valuation undisclosed · adjacent

Atinum Investment led, with Daewoong Group, Naver and Lotte joining as strategic investors, bringing cumulative funding to over 33.5 billion won. The company makes soft exoskeletons driven by cables and tendons and worn like clothing. Its main product H-Medi has received Class 2 medical device certification from South Korea's Ministry of Food and Drug Safety and has entered several university hospitals. The funds are mainly for overseas registration and for mass production of H-Care, an everyday walking-assist product. Source: Venture Square source

Origin ｜ first round ｜ $20 million ｜ valuation undisclosed · industrial

Accel and Zacua Ventures led, with some customers also investing. The company was founded in January 2025, and its flagship robot OG-1 currently does only wall sanding, priming and top-coat painting. 6 units are running in the US and Europe, with a plan to build 150 by the end of next year. CEO Yogesh Ghaturle says a 5-person crew at a US contractor costs $50,000 a month in labor, and replacing it with Origin's robots saves 20% of the cost from day one. Source: Livemint source

EACON removes safety drivers from 7 90-tonne mining trucks at a coal mine, moving to round-the-clock operation · autonomy

At a coal mine in northern China, 7 TR100A rigid mining trucks from Inner Mongolia North Heavy Truck (Chinese mining-truck maker) are already running without safety drivers, with the system supplied by EACON. Installing the driverless system took only three shifts, with the drive-by-wire chassis supplied by the OEM. This was followed by a month of control and system-safety validation, then about three months of full-process production testing with safety drivers on board. After safety drivers were removed at the end of August, the trucks ran day shifts first, then switched to around-the-clock operation at the end of September. Operating conditions here are not simple: the trucks may mix with human-driven vehicles, must climb a slope to the crusher when fully loaded, and face ice and snow in winter and muddy roads during the thaw. The company says it has adapted over 100 models from more than 10 OEMs. Source: International Mining source

Waabi signs a second Texas carrier to haul general freight on the Dallas–Houston lane · autonomy

Nine days ago, Waabi and Volvo began hauling freight for Warp. This time Waabi has signed carrier Warrior Logistics to move general freight such as food and beverages and consumer goods on the same lane using Waabi Driver trucks. The two began commercial operations earlier this year, and Waabi says its on-time rate so far is 100% with zero safety incidents, but it did not disclose vehicle counts or shipment volumes. Warrior owns over 200 trucks. CEO Scott McFarlane says the driverless trucks are used to fill lanes drivers do not like to run, alongside professional drivers. These trucks still have a person in the driver's seat for now. Source: FreightWaves source

Dyna says its robots fold thousands of napkins a day at multiple Din Tai Fung locations · embodied

Jason Ma, co-founder and chief scientist of Dyna Robotics, said on the Latent Space podcast that the company's robots fold thousands of napkins a day at multiple Din Tai Fung restaurants in the US, which is now routine operation rather than a pilot. He said that, to his knowledge, no other general-purpose robotics company has had a customer roll robots out across multiple locations. He also said that learning to cut vegetables on top of Dyna-2 took only 45 minutes of robot data. Neither the number of locations nor the number of robots was disclosed. ⚠️ Single-source claimSource: The Paper (Pengpai) source

Apollo Go (Baidu's robotaxi service) holds its first free public rides in Tung Chung, Hong Kong from October 21 · autonomy

The ride trial runs for 14 days, ending November 3, and is expected to serve over 1,000 people. Each trip takes about 15 minutes and covers 5.1 km, departing from Car Park 1 at Yat Tung Estate, passing the Sheraton Tung Chung Hotel and returning. Each vehicle carries two passengers, and the front passenger seat is not open. Residents must take part with physical tickets distributed by the Home Affairs Department from October 16. Apollo Go is the first company in Hong Kong to obtain an autonomous driving pilot licence, and its test mileage in Hong Kong has exceeded 280,000 km. Source: DotDotNews source

First 6 Chinese national standards for humanoid robots released, with numbers for grasping, payload and disturbance rejection · humanoid

The State Administration for Market Regulation has approved and released the first 6 Chinese national standards for humanoid robots, taking effect on April 1 next year. Five belong to the "Technical Requirements for Humanoid Robots" series, specifying body composition, system architecture, and functional indicators such as environment perception, decision-making and planning, motion control, and task operation. The other is the "Technical Specification for Humanoid Robot Simulation Test Platforms". The task-operation standard gives concrete figures: a grasp success rate of over 90% for rigid objects, over 85% for easily deformed or fragile objects, and force operation of over 15 kg. Motion-control stability is tested like this: push the robot with a rated external force, and it must return to a stable working state within 1 second. Drafter Quan Qi gave an example: decision-making and planning require command accuracy of over 95%, with a single action completed within 1 second. Wang Xingxing, founder of Unitree, is also a drafter. He said that products from different makers currently differ widely in size, performance, battery voltage, communication and usage interfaces. Two earlier national guiding technical documents on embodied AI standardization covered training data; this time it is the turn of the whole machine. Source: CCTV News source

South Korea's regulatory sandbox clears 20 bipedal humanoid robots to work alongside workers · humanoid

South Korea's Ministry of Trade, Industry and Energy announced on October 13 that, through the industrial convergence regulatory sandbox, it approved robotics company Robros to deploy 20 bipedal humanoid robots at four companies, working in the same or adjacent areas as workers. The tasks are cleaning semiconductor packaging parts at Nepes, moving and positioning car-body parts at Sungwoo Hitech, stacking scaffold connectors in shipbuilding at HD Hyundai Samho, and handling cold-chain packaging at Delivery Lab. Under the original rules, fixed industrial robots had to be separated from workers by fences; this is relaxed only for the pilot projects. Companies must still conduct risk assessments as required by the Ministry of Employment and Labor. The government plans to use pilot data to revise safety standards for mobile robots. The Hyundai Motor union has previously warned that large-scale deployment of humanoid robots would hit employment, and demanded that the company consult the union before introducing them. Source: cnBeta source

Cologne police seize a robot accused of sexually assaulting a girl · adjacent

According to the Associated Press, police in Cologne say that from last Saturday night into Sunday, two young men operated a robot to touch the buttocks of a 16-year-old girl and make thrusting motions at her. The suspected owner is a 16-year-old boy and his companion is 19; the two are accused of filming the incident and livestreaming commentary online. Police found the robot in a car trunk near a fast-food restaurant and confiscated it along with both men's phones. Neither has been charged. The investigation will determine whether the robot's movements were pre-programmed or remotely controlled by someone in real time. A police spokesperson said they had not seen this kind of case before, and that it was "new territory for us". Source: Associated Press source

Boston Dynamics discloses for the first time how long it takes to assemble one Atlas: about a day · humanoid

At its prototype shop in Waltham, Massachusetts, Boston Dynamics now needs about one day to assemble an Atlas, versus several days a year ago. In a video of about 6 minutes released by the company, senior technical program manager Katie Conner says the goal is to cut it to one hour by 2027 and ultimately to a matter of minutes. This is the first time the company has publicly given a per-unit assembly time. Gears are cut in the company's own shop, and a change an engineer proposes on Monday can be installed on a complete machine that same week. Tech Times cautions that this figure is the time to build one machine from parts into a complete unit. Figure said in April that its BotQ factory is "producing one per hour", which refers to production-line takt time, so the two cannot be compared directly. Source: Tech Times source

XPENG (Chinese EV maker) formally begins selling complete vehicles and physical AI technology externally in Paris · autonomy

In September, Reuters reported that XPENG planned to sell technology to foreign automakers. On October 13, XPENG formally launched its smart-vehicle solutions business XPENG AI ENGINEERING at the Paris Motor Show. The scope of cooperation extends from one-stop solutions covering vehicle development, production and after-sales to robots, robotaxis, flying cars, the Turing AI chip and a physical-world foundation model. This is the first time the Turing chip has been explicitly listed for external cooperation. According to Netcoo, XPENG is in talks with a leading European mobility platform about robotaxis. Source: Netcoo source

Inovance shows a humanoid robot for the first time at the China International Industry Fair · humanoid

The robot from Inovance (Chinese industrial automation maker) is a wheeled base with a seven-axis arm and force control. On site it worked with a vision system and an arc-welding robot to load and unload workpieces, reset them, and assist arc-welding operations. Pandaily reports that the model is the InoBot W1, which is liquid-cooled and aimed at factory shift work. Source: Shanghai Securities News; Pandaily source

Executives at humanoid robot companies leave one after another, most staying in embodied AI · adjacent

Liao Hongtao, former co-CTO of UBTech (Chinese humanoid robot maker), has reportedly joined Tencent's Futian Lab to lead embodied AI R&D; Red Star Capital Bureau (Chinese financial media) asked Tencent to confirm and received no definitive answer. Red Star Capital Bureau traced where executives at humanoid-body companies have gone over the past year. Xu Huazhe, co-founder and chief scientist of Galaxea (Chinese embodied-AI startup), left in February and founded Pokeqiao Robotics, with his former employer taking part in the angel round. Zhang Li, co-founder and COO of LimX Dynamics (Chinese humanoid and legged-robot maker), left in February and joined embodied-brain company BeingBeyond in April. Da Xingye, head of motion control for XPENG's robots, went back to Nvidia. Destinations cluster in home-scenario startups and "brain"-layer companies, and some have returned to large companies and universities. Source: Red Star Capital Bureau source

microagi claims $100 million ARR in half a year, on an annualized extrapolation basis · adjacent

Munich's microagi neither builds robots nor trains foundation models; it does only factory deployment. It records data on production lines with dedicated equipment, fine-tunes off-the-shelf models after simulation augmentation, then sends engineers on site. CEO Bercan Kilic said on X on October 9 that after launching its consumer business Shift, it went from 0 to $100 million ARR in 6 months. Humanoids Daily pointed out that the footnote on the company's chart reads "annualized run rate", meaning an extrapolation from recent revenue rather than actual revenue over 6 months. Shift offers free cleaning and private-chef services in exchange for workers wearing recording devices while they work. ⚠️ Vendor claimSource: Wowtale source

Dobot launches PULSA collaborative robots, sixth axis up to 790°/second · industrial

Dobot (Chinese collaborative robot maker) says this is a speed record for industrial collaborative robots. The PULSA 7 and PULSA 12 have payloads of 7 kg and 12 kg, repeatability of ±0.02 mm and ±0.03 mm, and a functional safety rating of PLd Cat.3. ⚠️ Vendor claimSource: Shenzhen News Network source

① China's MIIT-sector AI safety evaluation framework released, with embodied AI as its own category · adjacent

On October 15, the "AI Safety Evaluation Framework for the Industry and Information Technology Sector (2026 Edition)" was released in Jinan, under the guidance of the Cybersecurity Administration of China's Ministry of Industry and Information Technology and compiled under the lead of the AI Industry Alliance of China. The framework has three layers: general safety, application safety and business safety. Within application safety, agents and embodied AI are listed separately, with evaluation requirements set for permissions, behavior, memory and tool calls. Version 3.0 of the "AI Safety Governance Framework" released by TC260 (China's national cybersecurity standards committee) in September already listed embodied AI as a risk category; this time the MIIT system supplies the evaluation method. Source: Qilu Evening News source

① St. Paul joins Twin Cities robotaxi regulation; Minneapolis votes Thursday on whether to override a veto · autonomy

On Wednesday local time, the St. Paul City Council passed a resolution directing the planning commission to study autonomous vehicles and propose possible rules on safety, data sharing and worker protection. Minneapolis has gone further. Seven of the city council's 13 members voted to require robotaxis to carry human safety drivers, and Mayor Jacob Frey vetoed the ordinance; the vote on whether to override the veto is set for Thursday. St. Paul council president Rebecca Noecker said safety drivers might be one approach but that it is too early to draw conclusions. Council member Hwa Jeong Kim, citing the livelihoods of ride-hail drivers, opposed "siding with billionaire tech CEOs". Waymo has been testing in Minneapolis for nearly a year. Source: MPR News source

③ Screws: orders from a major customer are shrinking, while outside capital still pays premiums to acquire · hardware

This week, planetary roller screws showed two signals pointing in opposite directions. On the demand side, Xinjian Transmission (Chinese screw maker) disclosed in its ChiNext inquiry reply that its screw sales to top customer "Company A" fell from RMB 16.5875 million in 2025 to RMB 1.5248 million in the first half of this year, and only RMB 224,400 in July and August. The company explains that the customer's R&D demand has fallen as it enters early mass production. On the capital side, Yuehongyuan A, whose main businesses are real estate and scrap resource recycling, plans to pay up to RMB 550 million in cash to acquire screw maker Jinwangda, a premium of about 286% over net assets, with the other side committing to cumulative net profit of no less than RMB 180 million for 2026 to 2029. On one side, suppliers are still waiting for major customers' mass-production orders; on the other, capital has already paid a high premium against four-year profit commitments. Source: Xinjian Transmission inquiry reply; Duchuang Finance source