Key Takeaways

· China's Ministry of Industry and Information Technology (MIIT): Chinese humanoid robot output in H1 exceeded 40,000 units, and is expected to pass 100,000 for the full year

· GPT-6-Astra scores 100% on LIBERO-Long++, but on real robots it tried plugging in a plug 11 times and never succeeded

· Dyna's napkin-folding robot hit a 99.4% success rate in the lab, but after ten months at Din Tai Fung (Taiwanese restaurant chain) stores, only 75% of its output met the standard

· China's National Bureau of Statistics: industrial robot output rose 29.1% year on year in the first three quarters

· At the current pace, Waymo will import about 5,100 Zeekr-built Ojai vehicles by year-end, more than doubling its fleet

· Boston Dynamics CFO: about 70% of Atlas's bill-of-materials cost goes to actuators, which Hyundai Mobis supplies at cost

Does GPT-6-Astra Count as Embodied Intelligence? A Systematic Evaluation Across 12 Simulation Benchmarks and 8 Real-Robot Configurationsbenchmark

HF 156↑. The evaluation granted two relaxed conditions: in-context learning from training-set demonstrations was allowed, and the action-step limit was relaxed to 10x. Under these two conditions, GPT-6-Astra scored 100% on LIBERO-Long++ and 99.3% on LIBERO-Pro, and on RoboDojo it rose from 27.6% under the standard budget to 81.9%. On the long-horizon household benchmark BEHAVIOR it reached only 37.0%, and extra steps did not help. The gap was larger on real robots: putting a pen into a pen holder took more than 4 hours and cost 55 US dollars; plugging in a plug was cross-validated 11 times across several real-robot setups, and it never succeeded once. The report concludes that the model has broadly crossed the threshold of "embodied capability" but is still far from deployable embodied intelligence. The report attributes the real-robot problems to a few causes: demonstration data is expensive to collect, physical state is hard to verify, and irreversible errors are hard to recover from.

Zhongang Cai et al. (SenseNova joint team: SenseTime, Daxiao Robotics (Chinese embodied-AI company), Ropedia, Nanyang Technological University, University of Tokyo) · arXiv 2610.22471 source · Commentary: Machine Heart (Chinese AI media outlet) source

HexaShell: A Coding Agent Directly Drives a Dual-Arm Robot, and the Bottleneck Is Contact, Not Reasoningmanipulation

HF 104↑. HexaShell does very little: it converts the command line of a coding agent (Codex calling GPT-6-Astra, at medium reasoning effort) into dual-arm actions. After each failure, the agent reads the logs to find the cause, writes or repairs small tools itself, and localizes objects by triangulating across three cameras. Across all 42 RoboDojo tasks, the success rate was 39.7% with a progress score of 48.0, which the authors call a new record on that leaderboard. Pushing tasks reached 92.5%, while tight-fit insertion reached only 8.3%. With the same GPT-6-Astra, the previous report's RoboDojo success rate under the standard step budget was 27.6%; HexaShell's tools, by contrast, were written on 10 layouts per seed, frozen, and only then tested. The tools are open-sourced.

Hongcheng Gao et al. · arXiv 2610.23479 source

RAX: Letting a Coding Agent Explore Robot Programs Online, Without Demonstrationsmanipulation

When a new robot embodiment comes online, a batch of teleoperation data usually has to be accumulated first. RAX takes a different route: a general-purpose coding agent explores online and writes control code directly, with no action demonstrations, no hand-built primitives, and no value-function guidance. Tests covered single-arm, dual-arm, dexterous-hand, mobile, and humanoid embodiments, and in simulation it beat the strongest baseline by an average of 37.4%. The larger the exploration scale, the better it generalizes in unseen environments. The framework is open-sourced, and a new embodiment only needs to connect an API.

Runhan Huang et al. · arXiv 2610.22496 source

Millimeter-Level Assembly Using Only Proprioception: Trained in Simulation, Zero-Shot on Real Robotsmanipulation

The policy sees only joint positions, velocities, and its own issued actions, with no vision, torque, or tactile input. Part poses are unknown and grasp poses are heavily randomized, and all training is done in simulation with reinforcement learning. The peg-insertion policy transfers zero-shot to a real robot, with a 73% success rate across pegs of different shapes and sizes. In simulation comparisons, this blind policy generalizes to unseen parts noticeably better than a vision policy under the same conditions, though its motions are slower. The accompanying MuJoCo assembly generator is open-sourced as well.

Antonia Bronars et al. (MIT, Pulkit Agrawal's group) · arXiv 2610.23849 source

Rotating Baoding Balls Without Looking: Shadow Hand Spins 225 Times in a Rowmanipulation

Using only proprioception and binary touch, the Shadow Dexterous Hand Lite completed 225 180° rotations in a single continuous trial, equal to 112.5 full turns, at about 0.36 turns per second. The authors say this is the best hardware result so far for the Baoding ball task in both number of rotations and speed, counting every sensing modality. The paper also identifies the regions of the hand where tactile information matters most.

Ayush Deshmukh et al. (University of Edinburgh, Sethu Vijayakumar's group) · arXiv 2610.23438 source

FCTS: Generating Training Terrain from Failure Points, Quadruped Parkour Across Construction Siteslocomotion

Most of the authors are from FieldAI. Quadruped parkour policies are stable on standard obstacles but tend to fail on the irregular, combined terrain of construction sites. FCTS uses the location, termination mode, and surrounding terrain of each failure to generate the next batch of training terrain, then uses Monte Carlo tree search to select a curriculum within a fixed budget. On a medium-difficulty course built from real construction-site scans, the success rate rose from 28.4% to 85.8%, and real-robot experiments confirmed that the policy transfers.

Narayanan Palghat Parameswaran et al. (CMU, FieldAI) · arXiv 2610.22619 source

StopWatch: The Weak Point of Object-Goal Navigation Is "Not Recognizing the Target"autonomy

HF 3↑. After a VLM is fine-tuned into a navigation policy, its success rate is still low. The authors diagnose the main cause as target localization: the policy explores well but often fails to notice that the target is already in view, or misjudges how far away it is. With explicit target localization added, it reaches SOTA on six benchmarks, and the authors also corrected annotation errors in those benchmarks. Given ground-truth localization, the policy nearly matches the expert that generated the demonstration data, so the remaining room for improvement lies in perception.

Xinhao Liu et al. (authors include Pieter Abbeel) · arXiv 2610.23841 source

LAW: A VLM as Long-Horizon Scheduler for a World-Action Modelworld-model

World-action models (WAMs, i.e. models that predict both actions and future frames) are good at producing actions, but long-horizon tasks also need language-level planning. LAW uses a 2B VLM to plan the remaining subtasks, scores and fuses the candidate actions proposed by the WAM, and asynchronously judges execution progress. Average success across the 50 RoboTwin tasks is 94.21%, and 95.56% with the OpenWAM-α backbone. On three long-horizon real-robot tasks the success rate is 65%, versus 28.33% for π0.5.

Jiajun Cao et al. · arXiv 2610.23334 source

Xspark AI ｜ New round ｜ Amount undisclosed (cumulative several hundred million yuan)embodied

The round included Liedao Investment, Huashan Capital, Yuanhang Capital, Pudong Venture Capital, Junyuan Capital, Nanshan Strategic Emerging Industry Investment, and Tsinghua-affiliated Shuimu Venture Capital, with Duowei Capital as exclusive financial advisor. The team comes out of Tsinghua University and is led by Ding Wenbo, a tenured associate professor at Tsinghua; it builds a vision-tactile-language-action model with touch built in from the pretraining stage. Spark-0, launched alongside the funding, was published in a paper earlier; the company says it averages a 52% success rate on five contact-intensive real-robot tasks, 2.4 times the strongest baseline. The 1,000-hour HumanTouch tactile dataset (360 contact points per hand) and the SparkArena evaluation benchmark are, the company says, to be open-sourced soon. The money will go toward model iteration and toward deployment and delivery in industrial and commercial-service settings. Source: Gasgoo (Chinese automotive media) source

Demaco Technology (688360.SH) ｜ Convertible bond ｜ Up to 702 million yuanindustrial

Logistics-equipment maker Demaco Technology plans to issue convertible bonds to non-specific investors. The funded projects include building an R&D center for intelligent logistics equipment and embodied-intelligence systems; the rest goes to expanding core-component capacity, industrial data-collection terminals, permanent-magnet motors, and working capital. On the same day, Ant Lingbo (embodied-AI unit of Ant Group) said it had signed with Demaco to jointly tackle high-speed flexible sorting of massive parcel volumes. Source: Eastmoney (relaying STAR Market Daily) source; Sohu source

Deep Medical (Guangzhou) ｜ Series B ｜ Nearly 100 million yuanadjacent

Mifang Capital and Shengshi Investment co-led, with Chenhui Venture Capital, Chengshu Investment, Dongguan Science and Technology Innovation, and existing shareholder Wondfo Chuanggu participating; this is the company's third round in 15 months. Its product FUNZAR is a microscope surgical robot with two arms and 16 degrees of freedom; the company says its positioning error is under 1 micrometer, it has entered China's national "green channel" for innovative medical devices, and a multicenter registration clinical trial is under way. Source: PEdaily (citing Dongmai Net, a Chinese healthcare industry outlet) source

TiMOTION (Taiwan) ｜ Emerging Stock Board listing ｜ NT$168 per sharehardware

Linear-actuator maker TiMOTION listed on Taiwan's Emerging Stock Board on October 21, with SinoPac Securities as lead underwriter. Its 2025 revenue was NT$5.642 billion, with net profit after tax of NT$500 million. Medical applications account for 49.3% of revenue, and the share from industrial applications rose to 16% in the first half of this year. Chairman Tseng Kuan-shu said the company will enter the industrial robot market next year. Source: BigGo Finance source

Dyna's Ten Months at Din Tai Fung: 99.4% Lab Success Rate, 75% Pass Rate in Storesembodied

The company previously said its robots fold napkins every day at multiple Din Tai Fung stores. This time, co-founder Jason Ma went through the numbers for those ten months on the Silicon Valley 101 podcast (Chinese-language tech podcast). The first-generation robot's napkin-folding success rate in the lab was 99.4%, but after running continuously in stores, only 75% met the customer's standard. The customer wants neatness, with tolerances held to between a few millimeters and one centimeter. Speed also fell short: Dyna-1 folded 35 napkins per hour, about 480 a day, while the customer's threshold was 1,500 over an 18-hour shift. Only when Dyna-2 raised speed to 95 per hour, 1,590 a day, did it clear the bar. In his words: "The gap between the lab and real deployment is huge; beyond success rate, there are neatness, language following, and speed." By his assessment, robot data is currently at a stage comparable to the GPT-2-to-GPT-3 range for language models. Source: BigGo Finance source

Waymo to Import About 5,100 Zeekr-Built Ojai Vehicles by Year-End; Washington Worries About Securityautonomy

Tracking data from research firm MoffettNathanson shows that, at the current pace, Waymo will import about 5,100 Ojai vehicles by year-end; its existing fleet is mainly Jaguar I-Pace, so the fleet would more than double once this batch arrives. Autohome (Chinese auto portal) had earlier counted, from customs bills of lading, more than 3,200 of these vehicles cumulatively arrived at port. The US Department of Commerce has ruled that, starting with model year 2027, importing or selling connected vehicles and autonomous-driving software made in China or Russia is banned, and both chambers of Congress are also considering writing the ban into law. Waymo's position is that it buys only the vehicle body from Zeekr (Chinese EV brand under Geely), and that sensors and autonomous-driving software are installed at its Arizona plant. Missy Cummings, a former senior safety adviser at NHTSA, told the New York Post: "I'd have to get in the car myself and look at the physical structure to know whether that is absolutely true." Source: New York Post source

Shuohuang Railway Heavy-Haul Train Autonomous-Driving Trial Passes 7 Million Safe Kilometersautonomy

As of mid-October, the autonomous-driving trial of heavy-haul trains on CHN Energy's Shuohuang Railway had logged a cumulative 7,020,792 km of safe train control and safely operated more than 20,000 trains. This energy-transport line has a main line 598 km long and a maximum gradient of 12‰, and it runs 20,000-tonne heavy-haul trains. The trial began on April 8, 2024; this year it ran 3 hazard-identification exercises and compiled 34 hazard sources. This mileage is still recorded under the "trial" label, and the conditions for putting locomotives into service or disabling them are managed dynamically through a checklist. Source: Cangzhou Fabu (Cangzhou municipal government outlet) source

s.Oliver Picks Nomagic to Pick Shoeboxes with AI in a German Warehouseindustrial

German apparel retailer s.Oliver Group has chosen Nomagic to introduce AI picking at its Dettelbach logistics center. The first system is the Shoebox Picker, built specifically to grab shoeboxes, which have long been a difficult item for conventional robotic picking. The article does not give a go-live date or the number of units. Source: Logistics Manager source

MIIT: H1 Humanoid Robot Output Exceeded 40,000 Units, Expected to Pass 100,000 for the Full Yearhumanoid

Wei Wei, head of the Science and Technology Department at China's Ministry of Industry and Information Technology, said on the 19th at the State Council Information Office press conference on science and technology for commerce that in the first half of this year, complete humanoid robot products in China exceeded 400 models and output topped 40,000 units, with the full-year figure expected to pass 100,000. The same day, Guandian Net (Chinese financial news site) cited a Frost & Sullivan report: global humanoid robot sales in H1 were 36,000 units, exceeding last year's full-year 35,000, with revenue of 540 million US dollars, and all of the top five by revenue are Chinese companies. China's official output figure is more than 4,000 units higher than this global sales figure. One counts output and the other counts sales, and the model boundaries may not be the same either. Source: Yicai (Chinese financial media) source; Guandian Net source

Industrial Robot Output Up 29.1% Year on Year in the First Three Quartersindustrial

Mao Shengyong, deputy head of China's National Bureau of Statistics, announced on the 19th at the State Council Information Office press conference that industrial robot output grew 29.1% year on year in the first three quarters, broadly in line with the 29% cumulative growth for January to August. Also announced at the same session: output of 3D printing equipment and lithium-ion batteries grew 62.5% and 39.5% respectively; China has cumulatively built more than 9,000 advanced-level and more than 500 excellence-level smart factories, and the adoption rate of key AI application scenarios in above-scale manufacturing exceeds one third. Source: Cailian Press (Chinese financial news agency) source

Nine Chinese Ministries Including the Ministry of Commerce Issue Document Pushing Embodied-AI Results to Marketadjacent

China's Ministry of Commerce and 8 other departments issued an opinion on accelerating science and technology for commerce, calling for faster market translation of frontier results such as brain-computer interfaces and embodied intelligence, and supporting commerce and trade enterprises in smart warehousing and digital supply chains in applying for "specialized, refined, distinctive and innovative" (SRDI) designation. The opinion also calls for cultivating and building commercial technology application innovation centers. Source: Guandian Net source

Boston Dynamics CFO: About 70% of Atlas's Bill-of-Materials Cost Is Actuators, Supplied by Hyundai Mobis at Costhumanoid

Amanda McMaster served six years as Boston Dynamics CFO and has been acting CEO since February, until Rohit Prasad took over on October 7. Speaking at the Fortune Most Powerful Women Summit, she said about 70% of Atlas's bill-of-materials cost is actuators, produced by Hyundai Mobis at cost. Parent company Hyundai has committed to deploying about 50,000 robots in Hyundai and Kia plants, which gives the company what she called "planning certainty" and "cost-target certainty." Her approach is "control cost first, then reach scale as fast as possible." The Spot base platform costs about 100,000 US dollars and up to 300,000 US dollars with payload, and Intel is one of the largest customers. Hyundai's earlier public figure was deploying 25,000 Atlas units in Hyundai and Kia plants; this 50,000 refers to "robot deployment," and she did not elaborate on whether the two use the same definition. Separately, according to Digitimes, Hyundai Motor Group has set up an RX business strategy team to coordinate its global robotics strategy and speed up Atlas mass production and sales. Source: Fortune source; Digitimes source

Daxiao Robotics Releases 9B Embodied Foundation Model ACE-Brain-0.6, Claims It Beats GPT-6 Astra on Six Spatial-Grounding Evaluationsworld-model ⚠️ Vendor claim

ACE-Brain-0.6 is supervised fine-tuned from Qwen3.5-9B on about one million spatial-grounding samples, focusing on object recognition, spatial localization, and interaction-region judgment. Daxiao says it beats GPT-6 Astra on six of eight grounding evaluations: an average of 74.53 across seven public benchmarks versus 73.68 for Astra, and 94.13 on the real-robot-scenario evaluation RobotGrounding versus 91.10 for Astra. Daxiao was founded by SenseTime co-founder Wang Xiaogang and is also one of the collaborators on the GPT-6-Astra embodied evaluation report mentioned above. Source: IT Home (Chinese tech news site; republished by Sina Tech) source

Lenovo's "Mother Factory" Pilot-Production Base Expands to 11,000 Square Meters, Trial-Producing and Mass-Producing for Robot Companiesadjacent

Lenovo Group announced on October 16 that the "manufacturing as a service" workshop at its Shenzhen Southern Intelligent Manufacturing Base has grown from 4,000 square meters to 11,000 square meters, a fourfold increase in capacity, and that partner companies have grown from an initial 4 to more than 20. There are already a few cases: rehabilitation exoskeleton company Maibu Robotics assembled, tested, and delivered its first batch of 50 units in under a week with materials ready; Daimon Robotics' visuo-tactile sensors completed a round of trial production in two weeks and met the conditions for mass production in three months; and Benmo Technology's TITA wheeled-legged robot is already in mass production here. Six companies including Zhejiang Humanoid signed on to join the same day. Source: Shenzhen Business Daily source

Ant Lingbo Signs Partnerships with Longsheng Weirui, Demaco Technology, and Deyi Xiangwangworld-model

Each of the three partnerships targets one scenario: a joint embodied-intelligence robot innovation center with Longsheng Weirui, a majority-owned subsidiary of Longsheng Technology, starting with industrial loading and unloading; high-speed flexible parcel sorting with Demaco Technology; and automotive manufacturing workshops with Deyi Xiangwang, a unit of Deer Corp. (Deer Co., Ltd.). None of the three discloses an amount or a deployment count. At the China International Industry Fair, two robot bodies, from Taihu and Stardust Intelligence, both carried Lingbo's LingBot-VLA 2.0, demonstrating loading, unloading, and sorting. Source: Sohu source

Shenyi Launches Truck-Loading/Unloading Robot Sculptor, Rated at 400 to 800 Boxes per Hourindustrial ⚠️ Vendor claim

Shenyi Robotics unveiled Sculptor at the China International Industry Fair, powered by its in-house end-to-end embodied model SE-1 and aimed at mixed depalletizing/palletizing and truck loading and unloading. Rated specifications: maximum payload of 35 kg, arm working radius of 1.8 meters, a cycle rate of 400 to 800 boxes per hour, and 8 hours of continuous operation. The company also notes that actual performance depends on cargo specifications and on-site conditions. Source: Juti Zhijia (Chinese embodied-AI media outlet) source