Today's Highlights

· In two years, 106 embodied-AI data collection centers have appeared in China, 84 already in operation, with real-robot data priced around 500–1000 yuan per hour

· GPT-6 Astra directly commanded two YAM robotic arms to grab building blocks, succeeding 19 of 20 times; in the same test, Fable 5.1 succeeded 8 times

Causal chain-of-thought written into a world model, CausalWM takes first place on TriWorldBench · world-model

Most world models jump directly from the current frame to a future frame, compressing all the intervening motion, contact, and geometric change inside the model — making it hard to tell whether it truly understands physics or has simply found a shortcut in the training data. Aether AI's CausalWM (16B) has the model write out its reasoning process first: following the order Observation → Physical Motion → Geometry → Future Observation, it first uses optical flow to answer what is moving and in which direction, then uses depth and pointmaps to infer 3D relationships. Each step's result feeds back into the context to inform the next generation step, and a stage-ordered attention mask blocks any path that would let the model infer intermediate quantities by working backward from the answer. After training on roughly 30,000 hours of embodied video, it ranked Top-1 on the robot world-model benchmark TriWorldBench, which evaluates consistency across head-mounted and left/right wrist camera views across six dimensions and 19 metrics; code and weights were released alongside the paper.

Aether AI (founder Huang Biwei) · OpenReview https://openreview.net/pdf?id=3pf4d0EEqm · Analysis: Investment Circle source

Unifying actions as "where pixels fly to" cuts cross-embodiment world model error by 90% · world-model

Every robot speaks its own native language: industrial arms report joint angles, end effectors report spatial coordinates, and human demonstrations have no coordinates at all; the same end-effector command mounted on different structures produces completely different trajectories on camera, so cross-embodiment data can't be pooled together for training. Hydra-0, from NVIDIA together with Brown University and Columbia University, instead represents actions as pixel motion trajectories on the image plane — called action flow. When robot geometry and camera calibration are available, it's computed precisely via kinematic projection; with only raw video, dense point tracking combined with SAM 3 segmentation is used to infer it instead. Replacing the relative 6D end-effector representation used by the Cosmos 2.5 baseline cut gripper motion prediction error from 34.28 pixels to 3.29 pixels, covering seven or eight sources including DROID and EgoDex, totaling over 2,200 hours of video.

Joint team from NVIDIA, Brown, and Columbia · Paper: Hydra-0 · Analysis: TechWalker source

Zhishen Technology | Series B Strategic Round | Hundreds of Millions of Yuan · industrial ⚠️ Company-reported figures

Led by UAE-based Stone Venture, with participation from financial investors Hongshan Capital, Yueke Finance, Wuzhong Rongyue, and Wuzhong Financial Holdings, plus industry participants including Neusoft, Highpower Technology, Farsight Capital, and Riying Electronics. The company makes industrial-grade quadruped and humanoid platforms, with a product line including the Tongchui M1 series, Sibinggeng L2 quadruped series, the explosion-proof quadruped Qianqiu SP1, the humanoid Zhishen Yinyi NE01, and CHAMP joint modules, targeting power, petrochemical, firefighting, and emergency-response scenarios. It disclosed cumulative mass production of over 15,000 units of embodied AI robots as of June 2026, with monthly capacity exceeding 5,000 units and monthly output value of 140 million yuan. Following this round, the company is launching its "Gravitation" program, which in its first phase will provide developers, universities, and industry partners with 1,000 real robots and over 100 million yuan in resources.Source: Autohome source

Moqi Intelligence | Alibaba Takes Equity Stake | Amount Undisclosed · embodied

Shanghai Moqi Wanxiang Intelligent Technology's corporate registration was amended to add Hangzhou Haoyue Enterprise Management, an Alibaba affiliate, as a shareholder. Moqi Intelligence was registered in August 2025 and completed an angel round exceeding 1 billion yuan this July, co-led by Alibaba and Tencent, with over a dozen other institutions including Lanchi Ventures, Legend Capital, and Banyan Capital participating; the round valued the company at 7 billion yuan. The company's goal is a general-purpose home robot, though it has not yet released a product.Source: Guandian.cn source

Tianji Intelligent | Ant Group Takes Equity Stake | Registered Capital Raised to 244 Million Yuan · embodied

Following an undisclosed B++ strategic round from early September, the investor has now surfaced: Guangdong Tianji Intelligent Systems' corporate registration was amended to add Ant Technology Group's wholly owned subsidiary Shanghai Yunyu Enterprise Management Consulting, among others, as shareholders, with registered capital simultaneously raised to 244 million yuan.Source: People's Financial News source

Feitai Manufacturing (Shaoxing) | New Round | Tens of Millions of Yuan · industrial ⚠️ Company-reported figures

Led by Haiyuan Capital and Zheda Holdings, with follow-on participation from Zhongfu Venture Capital, Jinke Tongchuang, the Nantong Jianghai Talent Angel Fund, and individual angel investors. Feitai makes aerospace assembly robots, with a product line covering drilling, riveting, dispensing, sanding, and inspection, using miniaturized collaborative arms in place of traditional large dedicated machines — smaller footprint, more flexible to move, and priced at roughly one-eighth to one-tenth of comparable European and American products. The company states it holds nearly 70% of the Chinese aerospace assembly robot market, with customers including COMAC, AVIC, AECC, and China Aerospace Science and Industry Corporation, and says it became a core equipment supplier for phase-one mass production of the C919. The founder is a Shanghai Jiao Tong University graduate and former technical director at Shanghai Tuopu CNC Technology; over 60% of staff are on the technical team.Source: 36Kr source

Hai Robotics integrates over 1,500 climbing robots into a single system for a major European fashion retailer · industrial ⚠️ Company-reported figures

The automated operations area spans about 30,000 square meters, with roughly 1.2 million double-deep storage locations built, and the system is designed to handle over 24,000 totes per hour. More than 1,500 HaiPick Climb units are connected to a single integrated system running "goods-to-person" fulfillment. The buyer had already been running HaiPick systems at multiple European facilities, and this marks an expansion of an existing partnership. Hai Robotics describes this as its largest climbing-robot warehouse deployment globally to date. The buyer's name, the country where the project is located, the contract value, and the planned go-live date were not disclosed.Source: Weiduwang source

Qiyuan's Q1 and T1 personal robots go on sale starting at 19,999 yuan, shipping by order starting October 1 · humanoid ⚠️ Company-reported figures

The Qiyuan Q1 is priced at 19,999 yuan, with an Explorer Edition offering open SDK and HDK access at 26,999 yuan; the Qiyuan T1 is likewise 19,999 yuan, while the T1 Pro is 29,999 yuan and adds an NVIDIA Orin Nano chip, dual-mode 360° omnidirectional obstacle avoidance, and automatic low-battery return-to-charge. The first batch of authorized experience stores covers seven cities: Shanghai, Guangzhou, Shenzhen, Hangzhou, Xiamen, Wuhan, and Xi'an. The brand is backed by Sci-Tech Innovation Board-listed Shangwei New Materials, whose composite materials are used in the robot body, with self-developed QDD quasi-direct-drive joints about the size of an egg. The company disclosed that it has introduced automotive-industry manufacturing tempo, with a unit rolling off the line every 2.5 minutes and long-term traceable assembly data, and plans to invest 500 million yuan in R&D in 2026. Shangwei New Materials CEO Tian Hua said at the event, "someone has to take the first step," and that Qiyuan is "starting from 'fun' and gradually moving toward 'useful.'" Over 1,000 developers have already joined its skills platform for co-development.Source: Ifeng Technology source

Over 300 AgiBot robots begin work at Chimelong Spaceship Paradise on September 24 · embodied ⚠️ Company-reported figures

Hengqin's Chimelong Spaceship Paradise opens on September 24, with over 300 AgiBot robots from across its lineup entering the park the same day. The park will feature over 100 robot interaction points, spanning entertainment and live shows, science education, wayfinding and guided shopping, service retail stations, smart companionship, and hotel and sports/competition scenarios. Opening day will also see the unveiling of a joint research institute for embodied AI and tourism co-built by AgiBot and Chimelong. The host venue is the world's largest indoor theme park, having welcomed over ten million visitors since opening — a volume that requires the robots to run continuously day after day rather than performing on a scheduled basis.Source: Sina Finance source

Five King Long autonomous shuttle buses deployed at Xiamen Xiang'an International Airport · autonomy

The vehicles have begun scenario-based operation on internal airport roads, running a fairly long route that includes straight sections as well as passage through an underpass tunnel and around the terminal's concourse area. The vehicle model is based on King Long's M-series high-capacity transit bus, retaining its low-floor, fully flat-floor design. Xiang'an Airport has not yet opened for flights; this batch of vehicles is preparing shuttle infrastructure ahead of that opening.Source: Haixi Morning Post source

Figure claims to have measured a scaling law for humanoid robots; two peers push back on the spot · humanoid ⚠️ Company-reported figures

Figure AI rented 30 residential units in the San Francisco Bay Area and had Helix 2.5 — without any on-site data collection or fine-tuning — complete three tasks: making beds, folding towels, and tidying toys. Across 140 trials, the combined success rate was 56%, compared with roughly 9% for a control group without Index pretraining. Figure describes this as the first measured "human-to-robot transfer scaling law" on a humanoid robot. Peers did not offer the customary "great work" response: Sunday Robotics co-founder Tony Zhao pointed out that 237 successes also means the robot is failing nearly half the time; Stealth CEO Nikolai Ensslen was more blunt, saying the data actually proves that the imitation-learning system Figure uses cannot generalize. One caveat worth noting for context: the training data for all three tasks was still prepared by Figure itself — what Helix 2.5 hadn't seen was only the test residences, room layouts, and the specific objects within them.Source: GeekPark source

106 embodied-AI data collection centers built across China in two years, 84 in operation, with pricing bands now taking shape · adjacent ⚠️ Compiled figures

As of August 2026, China had a cumulative total of 106 centralized embodied-AI data collection centers, of which 84 are already operational. Of these, 52 disclosed robot deployment figures, totaling roughly 3,972 units, suggesting that over 5,000 robots are running across data collection centers nationwide. Scale is generally small: 55.8% of centers deploy 50 or fewer units, and the two largest each deploy 300 units — the Yibin High-Speed Rail South-Southwest Embodied AI Training Center and the Kaipuyun Embodied AI Training Ground. By region, Jiangsu leads with 18 centers, followed by Zhejiang with 17, Guangdong with 11, Sichuan with 8, and Beijing with 7. The pace of construction rose from a monthly average of 2.7 new projects between March 2024 and March 2025 to a monthly average of 7.3 between April 2025 and January 2026, peaking in the first quarter of 2026. Teleoperation remains the dominant data collection method, used by 83 centers, or 79.05%. Pricing bands have also taken shape for the first time: real-robot data runs about 500–1,000 yuan per hour, UMI data 300–400 yuan per hour, and simulated/synthetic data 200–500 yuan per hour. Robot platform brands are highly concentrated: AgiBot appears 31 times, LimX Dynamics (Chinese humanoid maker) 15 times, and UBTECH 13 times. This comes after Mech-Mind Robotics CEO Shao Tianlan publicly accused peers of using data collection centers and related-party transactions to fabricate revenue, and Galbot filed a police report denying the claim — the very revenue sources of this batch of centers are at the heart of that dispute.Source: Huxiu

GPT-6 Astra directly issues end-effector poses to a dual-arm robot, succeeding 19 of 20 grasps · embodied ⚠️ Single-source claim

Following the RoboHarm safety evaluation, RoboCurve connected GPT-6 Astra to a pair of real I2RT YAM robotic arms: at each step, the model views three camera feeds — one overhead and one from each wrist — along with the arms' own state, and directly outputs the end effector's x, y, z, yaw, pitch, roll, and gripper state, which lower-level inverse kinematics then converts into joint motions. The task was picking up red building blocks from a table and placing them in a bowl, with 19 of 20 attempts succeeding; in the same test, Fable 5.1 succeeded 8 times. CMU Robotics researcher Wenli Xiao ran a separate experiment: recording a video of a human performing a new task, feeding it into Codex, and then having Astra control the robotic arm to replicate it — succeeding on the first try, which he calls physical ICL. Amazon Alexa chief scientist and Retrocausal co-founder Zeeshan Zia's assessment: "It looks like the OpenAI of robotics might just be OpenAI, rather than Physical Intelligence or Skild." A 20-trial single-task demonstration is a capability showcase, still several orders of magnitude away from the tens-of-thousands-of-trials success-rate statistics used on a production line.Source: JiQiZhiXin source

Dongfeng's humanoid robot enters factories for sorting and quality inspection in October, small-batch trial production by year-end · humanoid ⚠️ Plan-stage figures

Zhang Zhenlin, Dongfeng Motor's chief engineer for intelligent technology, said Dongfeng's humanoid robot will enter factories this October, primarily for sorting and quality inspection in manufacturing; small-batch, limited-site trial production will follow by the end of this year, with the goal of reaching parity with human workers' capability by the end of next year. Its self-developed robot dog is expected to enter scenarios such as dealership sales guidance faster than the humanoid. Zhang's rationale is that automakers developing embodied AI products can reuse work from areas like automotive foundation models. At the same event, Dongfeng unveiled its "Sailing to the Horizon" plan, under which it will invest a cumulative 50 billion yuan overseas, launch more than 50 vehicle models, and build 10 overseas production bases, targeting 1.5 million overseas vehicle sales by 2030.Source: Sina Finance source

Orbbec signs multiple partners at once, expanding on both data collection hardware and robot vision fronts · hardware

Orbbec (688322.SH) launched an all-modal, embodiment-agnostic data collection hardware platform covering four form factors: first-person view, handheld operation, wrist-mounted near-field, and multi-device synchronization. Production-line testing showed continuous recording for over 37 hours with zero dropped frames, multi-camera hardware synchronization error under 1 millisecond, and factory calibration error under 0.3 pixels. Founder Huang Yuanhao explained the difficulty threshold at this year's World Robot Conference main forum: "If visual data and tactile data are off by 15 milliseconds, it's already wrong." Around the same time, the company signed a "vision + tactile" multimodal data collection partnership with Tashan Technology, a bidirectional alignment partnership on 3D vision and physical world models with Lynxi Intelligence, an eye-hand coordination partnership with LinkerBot, and deepened existing collaborations with Horizon Robotics (via its Digua Robotics unit) and Ant Lingbo. Financially, the company reported revenue of 438 million yuan and net profit attributable to shareholders of 41.64 million yuan in H1 2026, with vision-solution revenue for the global robotics market more than doubling year-over-year and overseas revenue nearly doubling as well.Source: Economic Observer source

Zhiqiao Power, a pool-cleaning robot maker backed by Ecovacs, files for its own bankruptcy · adjacent

On September 15, Zhiqiao Power (Suzhou) Technology had a new bankruptcy reorganization case filed, with both the petitioner and respondent being the company itself; the case is currently under review. Founded in September 2021, the pool-cleaning robot company targets the European and American markets, and in 2024 completed a Series A round co-led by Qingshuiwan Fund and Zhixing Yihao Fund, with products spanning the $300–$1,500 price range and cumulative sales approaching 100 million yuan. In August of that year, an Ecovacs subsidiary invested 25 million yuan for an 8.62% stake and appointed a director; in January 2025 it made a further strategic investment. Starting this January the company has faced a series of enforcement actions, with a new 2.51 million yuan enforcement claim added on September 9; a corporate registration amendment on August 20 showed reduced holdings for early shareholders including founder Pu Hanlai, with several senior executives — Jia Lin, Jia Shuiliang, and Bao Qiang among them — exiting. Less than two years have passed since Ecovacs' first investment.Source: Brand Ark source