Today's Highlights
· China's Ministry of Industry and Information Technology and eight other departments issued the "15th Five-Year Plan" for intelligent connected new-energy vehicles, targeting large-scale deployment of vehicles with autonomous driving capability by 2030
· Skild AI reaches $100 million in annualized revenue just 10 months after its first commercial deployment
· A Fanuc-Google AI welding agent can weld just by scanning drawings with a tablet, shipping starting December
· UBTECH (Chinese humanoid robot maker) lands over 50 million yuan in overseas orders, with buyers in Europe, Japan, and South Korea
· Estun (Chinese industrial robot maker) completes consolidation of Estun Cool Drive after a 487 million yuan acquisition, with original shareholders committing to cumulative revenue of at least 620 million yuan over four years
· HuRo uses "roboticized" human videos for pretraining, lifting real-robot task completion rate from 51.5% to 80.3%
Research Progress
HuRo: Converting Human Videos into Robot Data for VLA Pretraining · vla
Five human video sources were converted into roughly 630,000 robot-aligned episodes totaling 142 million frames, forming the HuRo dataset. The authors set out to test whether roboticized human video can support VLA pretraining and scale favorably. Across four real-robot tasks, as pretraining scale increased, overall completion rate rose from 51.5% to 80.3%, while out-of-distribution scenarios with varied spatial and visual conditions rose from 34.9% to 72.2%. Code and data have been released.
Jinho Jeong et al. · arXiv 2609.10706 source
IMLE-VLA: Single-Step Action Generation Lifts π0.5 Inference Rate to 55 Hz · vla
π0.5's flow-matching action head requires 10 steps of Euler sampling each time, causing jerky robot motion. IMLE-VLA replaces this with a single-step generator trained using conditional Implicit Maximum Likelihood Estimation (cIMLE), raising the inference rate from 15 Hz to 55 Hz. It achieves a 98.0% average success rate across 40 LIBERO tasks and outperforms π0.5 on all four real-robot tasks with a Franka arm.
Kian Hosseinkhani et al. (Simon Fraser University / University of Pennsylvania) · arXiv 2609.10915 source
MaP-WAM: Moving Memory to the Planner, Fixing the Executor's Context Length · world-model
HF↑7. Long-horizon manipulation requires memory, and the common approach is to feed the entire history to the executor. MaP-WAM instead compresses completed segments into a language plan for "what to do next" plus visual prompts, so the executor only sees the current plan. It achieves 83.3% success rate on RMBench and 78.0% on real robots, with inference latency remaining largely flat as history length grows.
Sizhe Zhao et al. · arXiv 2609.11561 source
2AM: Memory Goes to the Agent, the Action Model Just Executes · vla
The best prior baseline on LIBERO-Mem achieved only a 14.8% average completion rate; 2AM reaches 76.3% without using depth, online geometry, or a planner. The approach lets a multimodal agent solely own task memory, compiling interaction history into subtask language and 2D pick-and-place prompts, which are then handed to an action model that only sees RGB and carries no state to execute. Strict success rate remains only 11.8%.
Yutong Hu et al. · arXiv 2609.11308 source
ReactHuman: A Multimodal LLM as Humanoid Brain Gets Roughly One in Three Sudden Hazards Wrong · benchmark
Sudden hazards like a plate sliding off or a knife falling demand on-the-spot reactions. ReactHuman built 17 event categories and over 1,000 scenarios reproducible down to the position using 240 Hz rigid-body simulation, and included objects whose appearance and physical properties conflict, such as a foam anvil and a steel apple. The seven MLLMs tested got roughly one in three hazard responses wrong, and errors did not decrease with larger model scale. The dataset has been released on Hugging Face.
Yizhan Li et al. · arXiv 2609.10895 source
An Anthropomorphic Dexterous Hand Writes with a Pen After an 18-Second Startup on a Laptop CPU · manipulation
After roughly 18 seconds of initialization, an anthropomorphic dexterous hand writes with a pen using purely in-hand motion, achieving an average planar precision of 0.6 mm. The method estimates and corrects the task Jacobian matrix of the hand-and-pen system in real time on the physical robot, without relying on contact models, simulation training, or demonstration data. The authors say this is the first time an anthropomorphic hand has written an arbitrary single-stroke trajectory using purely in-hand motion.
Kai Stewart et al. · arXiv 2609.11775 source
FARM: Reading Failure Signals from Frozen World-Model States · world-model
The entire method trains only a 33,985-parameter readout head, attached to frozen VLA-JEPA predicted states, to output a failure score at each step. Across seven source tasks, it achieves a pooled AUROC of 85.68 and transfers to three real robots: PIPER X, SO-101, and Franka. Once the frozen state is obtained, the added latency is about 0.23 milliseconds.
Haoran Pei et al. · arXiv 2609.11445 source
ActSafeGuard: Writing Hard Constraints Directly into Flow-Matching Policy Training · vla
Actions generated by VLA models can violate physical hard constraints, and existing approaches mostly bolt on corrections at inference time. ActSafeGuard instead turns feasibility constraints into a differentiable layer during training, using the same constraints for both training and execution. Across two backbones, π0.5 and Fast-WAM, step-wise safety rate reaches 100% with no drop in task success rate.
Jianming Ma et al. · arXiv 2609.11697 source
Other papers today: SEED-UMI (humans and robots wear the same exoskeleton to collect dexterous manipulation demonstrations, 3x the collection efficiency of exoskeleton teleoperation, arXiv 2609.11753 source); CAP (a humanoid locomotion policy that stays stable even when depth perception is intermittently unreliable, validated on Unitree G1, arXiv 2609.11553 source); GRADE (estimating dense depth in smoke using mmWave radar plus generative priors, MAE of 0.313 m under smoke, arXiv 2609.10756 source); UniMPA (unifying memory, future prediction, and action into a single VLA framework, arXiv 2609.11875 source); EgoGenEval (evaluating physical consistency of video generation models under camera ego-motion, none of 16 models tested met the bar on both dimensions, arXiv 2609.11172 source); ORCH (orchestrating up to 50 heterogeneous agents using organizational theory, raising average wildfire rescue task scores by 63.97%, arXiv 2609.11737 source); using formal verification to test end-to-end steering networks under unseen weather conditions (arXiv 2609.10951 source); SwarmNxt (an open-source hardware/software platform for high-speed drone swarms, arXiv 2609.11382 source); a human motion retargeting pipeline for the Galaxea R1 Pro wheeled humanoid (arXiv 2609.11357 source)
Open Source · Tools · Benchmarks
· Puffin-World: Dahua Robotics (Chinese robotics company), together with Nanyang Technological University's S-Lab, has open-sourced a unified multimodal world model that places physics, geometry, and appearance into a single framework as three types of 3D world state. Code, model, and the Puffin-16M dataset were released together, with the dataset containing 15 million vision-language-camera triplets. The team says it achieves SOTA on four public benchmarks including Stanford2D3D source
· π0 fine-tuning tutorial: AWS released a tutorial for fine-tuning π0 for robotic arm manipulation on SageMaker HyperPod EKS source
Funding & Deals
Motive (USA) | Growth Funding | Over $1.3 billion | Valuation Undisclosed · adjacent
General Catalyst's Customer Value Fund contributed over $1.3 billion. After securing the funding, Motive withdrew its S-1 filing, shelving its planned NYSE listing under the ticker "MTVE." The company builds fleet telematics, AI dashcams, and maintenance management software, and was last publicly valued at $2.85 billion in 2022. Founder Shoaib Makani said: "We're building the intelligence layer for the physical economy." Direct competitor Samsara, already public, is valued at roughly $20 billion with ARR exceeding $1 billion.Source: Tech Funding News source
Kinetix AI | Angel+ Series | Over 500 million yuan · humanoid
Temasek-owned Vertex Ventures led the round, with Fangguang Capital, Wanshi Capital, and others following. The company was founded in September 2025, with a founding team from Huawei's autonomous driving unit and the University of Hong Kong. Its products include the full-size humanoid platform KAI Bot, with 117 degrees of freedom across its body and tactile skin covering nearly 80% of its surface area with about 18,000 contact points, as well as a 37-DOF dexterous hand and a first-person-view data-collection headset. The company says it has accumulated over 100,000 hours of first-person video, and that the same algorithms behind its table-tennis system have been migrated onto both the Unitree G1 and the Zhiyuan Robotics (Chinese humanoid startup) Yuanzheng A3.Source: PEdaily source
Aidin Robotics (South Korea) | Strategic Investment | HD Hyundai Robotics, KRW 13 billion · hardware
Aidin makes six-axis force/torque sensors, exported to 10 countries including Germany, France, China, and Japan. The two companies will jointly develop a five-fingered dexterous hand for shipbuilding and heavy industry, to be mounted on HD Hyundai's humanoid robots; they will also start with shipyard automation for grinding and polishing, later targeting U.S. shipyards. The same day, Samsung Ventures separately invested KRW 3 billion, with the resulting equity stake undisclosed at the investee's request.Source: Seoul Economic Daily source
Swarmer (NASDAQ: SWMR) Acquires Ratel Robotics | Cash + Stock | Up to $224 million · adjacent
Swarmer's core business is drone-swarm coordination software; this acquisition of Ukrainian unmanned ground vehicle (UGV) maker Ratel extends its business from air to ground. Ratel has already signed $86 million in contracts this year, with its vehicles used for battlefield logistics, casualty evacuation, reconnaissance, drone launch, and mine clearance. The $224 million figure is the ceiling if all earnout milestones are met, with the deal expected to close in Q4 2026.Source: The Robot Report source
Moxin Technology | New Round (About to Close) | Could Reach 1 billion yuan | Valuation Near 10 billion yuan · world-model ⚠️ Rumored terms
This is Moxin's fifth funding round this year, coming about 20 days after its previous 500-million-yuan-class round was disclosed, after which its valuation stood at nearly 4 billion yuan. The company started out in consumer 3D printing, pivoted to world models in late 2024, and in July released MoWorld, a 28B-parameter MoE model that it says can generate 3D scenes in real time at 50 frames per second, running entirely on Huawei Ascend NPUs.Source: 36Kr · Zhihui Yongxian source
Westlake Robotics | Series A1 | Several Hundred Million Yuan · embodied
The company was incubated by Westlake University; its previous Pre-A++ round was funded solely by a Henan state-owned investment fund.Source: Dahe Caili Cube source
Commercialization & Deployment
Skild AI Reaches $100 Million Annualized Revenue 10 Months After First Commercial Deployment · world-model
$100 million is the run rate — an annualized figure extrapolated from recent revenue, disclosed by the company to Bloomberg. Skild's robot foundation model is now used by more than 60 companies, running on hundreds of robots, up from just 8 earlier this year. Customers include a Japanese coffee shop and a U.S. warehouse. Co-founder and CEO Deepak Pathak said this marks robotics moving "from the demo era into the deployment era." The S1 model, released last month, uses in-context learning — following a demonstration video without weight updates — and can learn multi-step procedures over 10 minutes long from a single human video. The company has not disclosed a customer list, so the annualized figure cannot currently be reconciled with the deployment count.Source: Bloomberg source
UBTECH Lands Over 50 Million Yuan in Overseas Orders · humanoid
The orders cover the commercial service humanoid Walker C1 and the ultra-realistic humanoid UWorld U1, with buyers based in Europe, Japan, and South Korea. UBTECH sold 921 full-size humanoids in the first half of the year (previously reported). In South Korea, partner Deoksan JMR signed with auto parts maker AJIN Industrial, with plans to validate the Cruzr Y1 and Walker S2 at AJIN's plant starting in August.Source: Gasgoo source
Neolix's L4 Autonomous Delivery Vehicles Begin Testing in Tokyo · autonomy
The testing is taking place at the Tokyo Ryutsu Center (TRC) in Heiwajima, as a closed-facility trial. The first vehicles arrived on August 31, with the fleet subsequently expanding in batches to 10 vehicles, using the 520-kg-payload Xolra3 and the 1,000-1,130-kg-payload Xolra6. TRC operates around the clock, with daily truck traffic exceeding 10,000 vehicles. The company says it has deployed roughly 27,000 delivery vehicles worldwide.Source: LeiPhone source
Comau Tests Collaborative Robots for Picking and Palletizing at a Decathlon E-Commerce Warehouse · industrial
The system combines MyCo collaborative robots, ROS2 software, and modular grippers to autonomously perform picking, transport, and palletizing. This is part of the EU's MASTERLY project, currently in the testing and validation stage, with no deployment scale disclosed yet.Source: The Robot Report source
Industry Developments
China's Nine Ministries Issue "15th Five-Year Plan" for Intelligent Connected New-Energy Vehicles, Targeting Large-Scale Autonomous Driving by 2030 · autonomy
China's Ministry of Industry and Information Technology, together with the National Development and Reform Commission, the Ministry of Public Security, the Ministry of Transport, and nine ministries total, jointly issued the "15th Five-Year Plan for the Development of the Intelligent Connected New-Energy Vehicle Industry." The overall goal for 2030 states that vehicles with autonomous driving capability should achieve large-scale deployment, with highly automated driving realized on highways, urban expressways, and some city roads — a scenario that Autohome has categorized as L4. The plan also requires that the safety performance of vehicles equipped with autonomous driving systems substantially exceed that of human drivers. Under the same set of targets, new-energy vehicles should account for 70% of new passenger car sales and 40% of new commercial vehicle sales. Supporting measures include 17 key tasks across 5 areas, 4 major engineering initiatives, and 3 key actions, including an "AI+ Automotive" development action and a high-value general-purpose autonomous driving scenario library. A week earlier, China's draft amendment to the Road Traffic Safety Law added a dedicated chapter on autonomous driving (previously reported) — industry targets and legal liability rules have each advanced a step within the same month.Source: Cailian Press source
Fanuc and Google Launch AI Welding Agent That Reads Drawings, Shipping in December · industrial
By scanning a paper drawing with a tablet, the system can identify weld locations along with requirements such as voltage, current, and speed, then hand the job to a Fanuc collaborative robotic arm to autonomously perform arc welding. It can also recognize handwritten annotations on drawings. The system runs on Google's Gemini Enterprise, requires no programming from workers, and allows manual intervention when needed. The product ships in December, priced on a subscription basis (pricing undisclosed), targeting small and mid-sized businesses primarily in industries such as electrical components and furniture. The developers say this is the first system of its kind globally.Source: Nikkei Asia source
Estun Consolidates Estun Cool Drive into Its Books with a 487 Million Yuan Deal · industrial
In an announcement on the evening of September 7, Estun (Chinese industrial robot maker) said it had completed the acquisition of 100% of Estun Cool Drive's equity through two wholly owned subsidiaries, in a deal totaling 487 million yuan. The target company was previously held by Estun's controlling shareholder, Pailesite, and its core business covers collaborative robots, embodied AI robots, and related core components. Cool Drive posted revenue of 50.1679 million yuan in 2025, with a net loss of 52.9985 million yuan. The original shareholders committed to cumulative revenue of at least 620 million yuan from May 2026 through 2029, including at least 120 million yuan in 2027 alone — about 2.4 times its 2025 revenue. Humanoid-related revenue was not disclosed separately; a staff member from Estun's securities department said it accounts for a relatively small share.Source: DoNews source
Embodied AI Forum at the Bund Summit: Speakers Disagree on Whether Compute or Data Is the Bottleneck · embodied
Jia Peng, co-founder and CEO of Zhijian Power (Chinese embodied AI startup), argued that looking purely at the models, embodied AI has already shown some in-context learning ability — but for it to genuinely become productive, "it may still take three to four, or even four to five, more years." He identified GPU compute as the biggest bottleneck, since many factories can't connect to the network, meaning insufficient on-device compute directly limits the models. Wang Xiaogang, founder and chairman of Dahua Robotics (Chinese robotics company), placed the biggest bottleneck on data instead: "If you always need to retrain for new tasks, there's no way to achieve scale." On the "bubble" debate, Jia attributed it to a mismatch between the field's stage of development and its valuations, while Wang said "we're still in the early dawn."Source: PEdaily source
Teradyne Sues Jaka at the Unified Patent Court in Europe Over 3 UR Patents · industrial
Teradyne Robotics, the parent of Universal Robots, filed suit against Jaka's German subsidiary at the Unified Patent Court's (UPC) Copenhagen local division on August 24. The three patents concern robot interfaces, safety brakes, and joints, all with a priority date of March 2006. Any ruling can take effect across the UPC's 17 member states, and the UK and Spain are also included in the scope of the suit. Jaka has applied to the court for an injunction, saying Teradyne's public statements about the case are "false, misleading, and damaging," and cites two prior freedom-to-operate (FTO) analyses that found no infringement.Source: The Robot Report source
JD.com Launches "Physical AI Acceleration Plan" · adjacent
Cao Peng, chairman of JD Group's Technology Committee and president of JD Cloud, said at the JDD conference on September 9 that a large share of the technology "remains stuck at the demo stage." Under the plan, JD aims to collect over 10 million hours of real-world human video within two years; JD Logistics plans to procure 3 million robots, 1 million unmanned vehicles, and 100,000 drones within five years.⚠️ Company's stated planSource: TMTPost source, Gasgoo source
Baidu Sets Up South Korean Entity, Apollo Go Follows Waymo · autonomy
Baidu registered "Baidu Apollo Korea" in Seoul's Gangnam district on April 30, with Yang Nan — vice president responsible for Apollo Go's overseas business — serving as director; the registered scope covers autonomous driving R&D, ride-hailing services, and high-precision mapping data. Waymo established its own South Korean entity just in July. As of Q2, Apollo Go has been operating in 28 cities outside China.Source: Herald Business source
Tesla Showcases Cybercab in Japan · autonomy
Tesla's Japan head said the company hopes to launch a robotaxi business in Japan "if the timing is right," starting with rolling out advanced driver-assistance features on existing models. The Cybercab was unveiled in Japan this week.Source: Sina Finance source
Galbot Responds to Shao Tianlan: "Not Engaging in a War of Words" · embodied
After being called out by name by Shao Tianlan, CEO of Mech-Mind Robotics (previously reported), Galbot issued a statement on the evening of September 10: "We will not engage in pointless wars of words and will not be distracted by short-term noise." The statement did not address the specific allegations concerning its data collection center and related-party transaction revenue.Source: Ruicaijing source
Hardware · Supply Chain
· Fengli Intelligent (Chinese precision reducer maker): the company said at an earnings briefing that its harmonic, planetary, and small micro-reducer products for dexterous-hand joints can all be adapted for humanoid robots, that it already has batch-delivery capability, and that it is using proceeds from a private placement to expand production source ⚠️ Company's own statement
· Junding Da (Chinese tendon-cable supplier): sales revenue from tendon cables and tendon-cable sheaths was under 1 million yuan in the first half of the year; dexterous-hand glove protective materials are still in R&D, with no mass production and no formal contracts signed yet source
Top comments (0)