Today's Highlights

· Einride and Lidl put a cabless L4 truck into service on public roads in Germany, receiving Germany's first permit of its kind

· Agility unveiled Digit 5: charges in 9 minutes for 90 minutes of runtime, designed to work alongside humans without guardrails

· National Bureau of Statistics: August industrial robot production reached 96,174 units, up 34.6% year-on-year

· TC260 released AI Safety Governance Framework 3.0, incorporating embodied intelligence risks into the framework

· A quadruped robot reached 10.74 m/s, with the paper claiming it as the first legged robot to exceed 10 m/s

· GXO deployed 127 Exotec robots for Guess in the Netherlands, processing 40,000 to 70,000 items daily

Drawing on the Huygens coupled pendulum principle, a quadruped robot breaks 10 m/s · locomotion

A quadruped robot reached 10.74 m/s, covering 100 meters in 12.2 seconds; the authors say this is the first legged robot to exceed 10 m/s in speed. The speed gain didn't come from bigger motors. The team drew on the Huygens coupled pendulum, using inertial coupling between limbs to redistribute energy over the gait cycle, reducing the peak joint torque needed for high-speed running. Hardware parameters were also folded into the optimization, with the body and controller co-designed.

Yucheng Tao et al. · arXiv 2609.13290 source

PhysBrain 1.5 technical report: embodied pretraining supervision drawn entirely from human interaction videos · vla

HF 168↑, today's most-discussed paper. Following DeepCybo's (Chinese embodied-AI startup) open-sourcing of 2B and 8B weights on September 13, the technical report went live today. All embodied supervision in the pretraining stage comes from human interaction videos; language responses, end-effector trajectories, and dense visual goals are encoded into a single discrete sequence and jointly predicted autoregressively. The 8B model averages 72.5 across 28 embodied understanding benchmarks, which the authors call the best among open-source models, claiming performance on par with GPT-6-Astra and Gemini 3.6 Flash.

DeepCybo Team (深度机智) · arXiv 2609.14973 source

Visible Touch: rendering contact signals into the images the policy already sees · perception

Most VLAs use only vision and proprioception, with no way to incorporate touch. This paper renders contact signals into the same image space the policy already observes, so any image-input policy can use it without architectural changes. The authors also open-sourced a low-cost magnetic contact sensor built from off-the-shelf parts. Across four real-world contact-rich tasks, π0.5's success rate improved by 30 percentage points.

Metin Alp Dogan et al. · arXiv 2609.14156 source

Using MPC as scaffolding, real-robot reinforcement learning teaches a dexterous hand to rotate objects in 7 minutes · manipulation

The costliest part of real-robot reinforcement learning is inefficient early exploration and dropped objects. The authors first pretrained on hardware using 12 minutes of sampling-based MPC trajectory collection—20 trajectories total. During online training, MPC intervenes intermittently, gradually handing control to the learned policy, with no human demonstrations required throughout. On a 16-DOF Allegro hand performing continuous in-hand rotation, after 7 minutes of online training, the policy succeeded in all 5 independent evaluation trials, with an average of about 3 drops during training. By 20 minutes of training, rotation speed exceeded that of the MPC controller by more than 5x, and the hand rotated continuously for 1,000 revolutions over more than 110 minutes without a drop.

Emek Barış Küçüktabak et al. · arXiv 2609.14878 source

WLA³: using changes in world state as unified action supervision · vla

First-person human videos are abundant, but ones with high-quality hand motion annotations are scarce. WLA³ instead learns compact latent actions from observed changes in world state, with the same representation feeding semantic, dynamic, and kinematic components alike. Across six real-robot tasks, average success rate reached 81.9%, versus 66.2% for π0.5.

Peidong Liu et al. · arXiv 2609.15870 source

The biggest lever for efficient VLAs is action-head initialization · vla

The authors fixed the SigLIP2 and Qwen2.5 backbones and adjusted action-head design item by item, pairing each configuration with measured on-device latency. The single biggest gain came from copying the language backbone's final layers into the action head, at no added latency cost. Flow matching and heavier decoders only helped when the action head was misaligned—once aligned, they actually hurt performance; returns from scaling up the action head diminish quickly near the sizes currently used in the π series. Based on these findings, the resulting EffVLA matches the strongest open-source VLA on LIBERO with lower latency. The authors note this is only the explanation best fitting the current measurements—causality has not been established.

Luoyang Sun et al. · arXiv 2609.13984 source

X-WBC: a cross-embodiment foundation model for humanoid whole-body control · locomotion

Humanoid whole-body control typically trains one policy per robot, with locomotion experience unable to transfer across different embodiments. X-WBC uses human-centric instruction tokens to align human motion, reference robot motion, and sparse VR observations. A shared causal Transformer learns temporal structure from multi-robot data, with a lightweight module per robot mapping outputs to its own joints. Across 9 simulated embodiments and 4 real robots, joint training improved motion-tracking performance in every case.

Juntong Zhang et al. · arXiv 2609.15213 source

A world model picks scoop locations for a wheel loader, already running on a full-size machine · world-model

The system proposes multiple scooping plans each cycle, first ruling out geometrically infeasible ones, then uses a world model to predict terrain changes and load volume, executes the scoop predicted to load the most, and replans. Across 32 simulated test episodes, the average number of scoops dropped from 651.8 to 540.6, a 17.1% reduction. Evaluating 5 candidate plans on a Jetson AGX Orin took 72.4 milliseconds. The authors frame the real-machine portion as a feasibility demonstration.

Ailing Zhang et al. · arXiv 2609.15382 source

Exein (Italy) | New round | $270 million | $1.7 billion valuation · adjacent

Led by Headline. According to FT reporting, the valuation is more than double the $770 million from a round nine months ago. Exein builds embedded cybersecurity, writing security functions into firmware during device production; it is also developing a foundation model for physical AI security, planned for launch next year. Founder and CEO Gianni Cuozzo said: "The technology shift in robots, drones, and autonomous vehicles will only truly happen once people feel safe." Source: 디지털투데이 (citing FT) source

Paxini (Chinese tactile-sensing startup) | Series B++ | Several hundred million RMB · hardware ⚠️ single-source claim

This round brought in new investors including Guangxin Fund, Sichuan Provincial Sci-Tech Innovation Investment Group, Shandong Guohui Fund, and CITIC Capital, with existing investors Jingming Capital and Hexin Fangce continuing to invest. A 36Kr article states that Paxini closed a Series B of over 1 billion RMB in March and a 1 billion RMB strategic round in August, totaling more than 4 billion RMB across three rounds in half a year. The headline says "Samsung takes stake in Paxini," but the body only says Samsung made a "strategic bet," without disclosing the timing or amount of the investment. BYD made a strategic investment in Paxini in 2025, and the two signed a strategic cooperation agreement again on August 25. Source: 36Kr source

Yincheng Intelligence | Series A | Nearly 100 million RMB · industrial ⚠️ single-source claim

Co-led by Huamin Capital and Shanghai Semiconductor Industry Investment, with existing investor Detong Capital following on—the company's second round after a tens-of-millions-USD angel round in May. The company also says it secured a hundred-million-RMB order from SF Express, with robots first entering the parcel-sorting stage at transfer centers, with plans to deploy thousands of units within two years. Its Starborn sorting robot is rated at up to 1,800 items/hour, and the company says a single unit can replace 2 to 3 human workers. The order amount and deployment pace currently rest on the company's own account alone. Source: 智东西 source

Zeno AI (芝诺机器人) | Seed round | Several hundred million RMB · embodied

Investors include Sunny Optical (Chinese optics manufacturer), Metax, Wangxin Group, Saiyi Industrial Fund, and Heli Capital. The company simultaneously released its multi-robot collaboration foundation model, Zeno-1: each robot runs the same model independently, with no central controller and no access to other robots' internal states. The company says the same policy can continuously execute multiple subtasks for over 10 minutes without a mid-task reset. Source: 极客公园 source

RobotPlusPlus (ROBOT++) | Series C | Several hundred million RMB · industrial ⚠️ vendor claim

Led by Qianggang Capital Fund, with Chuangyi Capital Group (GIG Capital Group), Sealand Innovation, and Juntong Capital participating, and Fosun RZ Capital following on. The company makes specialized robots for high-altitude hazardous work such as hull derusting, and says its derusting robots have entered over 100 shipyards with over 70% market share, having served more than 10,000 cargo ships. Source: GlobeNewswire source

Hangcha Group (603298) | Convertible bonds | Up to 2.259 billion RMB · industrial

Proceeds will fund a smart factory for forklift robots, expansion of new-energy forklift production, and R&D for forklift and logistics robots, plus working capital; the plan still requires shareholder approval. The company's H1 revenue was 10.1 billion RMB, of which smart logistics integration and robotics-related business brought in 796 million RMB, up 30.81% year-on-year. Source: Securities Times source

Huanchuang Technology | Hong Kong IPO | Passed hearing September 13 · hardware

Sponsors are CICC and Guosen Securities (Hong Kong). The company makes spatial-sensing components such as lidar for robot vacuums; over the past several years, more than 98% of its revenue has come from the robot vacuum industry, with customers including Xiaomi and Roborock (Chinese robot vacuum maker). Its September 2025 Series D valued it post-money at about 3.04 billion RMB. Source: PEdaily source

Gongsheng Wujie | Angel round and Angel+ round | Tens of millions RMB · embodied

Funded by Lihe Venture Capital and Shenzhen High-Tech Investment. Its core product, GS-Hub, is a pluggable edge-side intelligent terminal that gives different robot embodiments spatial understanding and multi-robot collaboration capabilities. The company says industrial embodied robots equipped with the basic version of the system entered mass production in September. Source: 智东西 source

Einride's cabless L4 truck runs daily operations on public roads in Germany · autonomy

The truck has neither a driver nor a safety operator on board, hauling goods daily between one of Lidl's warehouse distribution centers and a store. It is a cabless SAE L4 vehicle, and the operating permit comes from Germany's Federal Motor Transport Authority (KBA)—the country's first such permit, issued after completing safety validation with the KBA. Einride CEO Roozbeh Charli said: "Germany's approval process is among the strictest in the world, and getting a permit there gives us the confidence to scale." Currently there is only one warehouse-to-store route. The two companies plan to expand it into a multi-stop delivery loop, with discussions underway for further deployments with other companies under the Schwarz Group. Source: Tech.eu source

GXO's Venlo, Netherlands warehouse deploys 127 Exotec robots for Guess · industrial

The Exotec Skypod system includes 127 robots, 60,000 storage locations, 8 goods-to-person picking stations, and 200 meters of conveyor. It processes 40,000 to 70,000 items daily, with a peak of 2,200 order lines per hour. Exotec served as the sole system integrator, delivering the full solution. The warehouse handles Guess's outbound shipments to EMEA and Asia stores, and is also preparing to handle e-commerce orders for the Benelux region. Source: securities.io source

Waymo partners with GO and Nihon Kotsu, plans robotaxi launch in Tokyo by 2027 · autonomy ⚠️ plan-stage claim

The three companies' goal is to launch Japan's first fully driverless taxi service by 2027, starting small and gradually scaling to about 100 vehicles. Since 2025, the three have conducted public-road testing across 7 wards of Tokyo, with a Nihon Kotsu driver still on board during current testing. The exact launch timeline depends on safety validation and approvals under road traffic law and other regulations. Source: Kyodo News source; 디지털투데이 source

Agility unveils Digit 5, designed to work alongside humans without guardrails · humanoid

Digit 5 charges in 9 minutes for 90 minutes of runtime, and Agility says a single unit can work over 20 hours per day; the previous-generation Digit 4 took 1 to 2 hours to charge with about 4 hours of runtime. The new model's repetitive lifting payload rises to 22.7 kg, 40% higher than its predecessor, standing 1.81 meters tall and weighing 129 kg. On safety, multiple sensors plus AI recognize nearby people and cause the robot to avoid, stop, or sit down, with audio and light alerts; an independent safety controller also monitors human-robot distance. Agility is the first launch partner for Nvidia's Halos for Robotics platform. CEO Peggy Johnson said market response has produced "over $300 million in multi-year orders." Early customer trials begin in the first half of next year, with full availability by the end of 2027, marking the first sales outside North America, starting with the EU and UK. The previous OSHA management-standard on-site assessments applied to Digit 4; the new safety architecture will be tested on customer production lines after delivery.

Source: PR Newswire source; The Decoder source; The AI Insider source

August industrial robot production up 34.6% year-on-year · industrial

National Bureau of Statistics data shows China's industrial robot production reached 96,174 units in August, up 34.6% year-on-year; cumulative production from January to August totaled 729,352 units, up 29% year-on-year. Service robot production in August was 1,656,132 units, up 5% year-on-year. The single-month growth rate exceeded the cumulative growth rate for the first eight months. Source: Mysteel, citing National Bureau of Statistics data source

TC260 releases AI Safety Governance Framework 3.0, incorporating embodied intelligence risks · adjacent

China's National Information Security Standardization Technical Committee (TC260) released version 3.0 of the "AI Safety Governance Framework" on September 14, during the opening of National Cybersecurity Publicity Week in Jinan—the third version in three years. The new version incorporates risks from computing infrastructure, AI agents, and embodied intelligence into the framework, focusing on scenarios such as autonomous systems escaping control and impacts on the physical world. Source: MLex source

Deep Blue Robotics and Zhicheng Power both enter bankruptcy proceedings · adjacent

Deep Blue Robotics' (Chinese robotics manufacturer) bankruptcy liquidation case was accepted by the Shanghai Pudong New Area People's Court on May 7. On September 7, the administrator disclosed the fifth batch of employee claims, involving over a hundred people with a total exceeding 23 million RMB. The company had once announced orders for tens of thousands of units and built a facility with annual capacity exceeding 20,000 units. Pool-cleaning robot company Zhicheng Power (Chinese robotics startup) had bankruptcy reorganization filed against it by creditors in early September, and the court has accepted the case for review; its investors include Qingshuiwan Fund and Ecovacs (Chinese robot vacuum maker), and it had claimed cumulative sales of nearly 100 million RMB in early 2025. The same report cites IT Juzi data: as of August 20 this year, there were 466 financing events in China's embodied intelligence sector, with disclosed totals reaching 124.51 billion RMB. Source: 智东西 source

Hesai and JD.com form strategic partnership, supplying lidar for JD's driverless vehicles and robots · hardware

JD's sixth-generation Plus-version delivery vehicle has adopted Hesai's AT series lidar. Hesai's robot-oriented JT series has a 360°×189° field of view. JD Logistics says it plans to procure 3 million robots, 1 million driverless vehicles, and 100,000 drones over the next five years. ⚠️ plan-stage claimSource: Gasgoo source

UBTECH grants H-share incentives at 1 RMB exercise price to 409 employees · humanoid

An announcement on September 14 said the grant covers 3.62447 million shares, or 0.72% of total share capital. At the closing price of HK$78.9 that day, book value is about HK$286 million. Chairman and CEO Zhou Jian (UBTECH founder) received 400,000 shares, with the remaining 405 employees collectively receiving 3.16357 million shares. The final three vesting tranches require meeting 2026 sales targets. UBTECH's cumulative losses from 2022 to 2025 exceeded 4.2 billion RMB. Source: Sing Tao source

SuperPanther publishes heavy-truck "Truck Brain" roadmap, signs cooperation MOU with DHL · autonomy ⚠️ plan-stage claim

The eTopas 600 electric heavy truck has completed its first batch of European customer deliveries, produced locally in Austria by Steyr Automotive. Truck Brain's first engineering goal is to deploy a 7-billion-parameter specialized model on-vehicle, tailored to a next-generation onboard platform of about 1,000 TOPS. The press release summary refers to "five European countries" for daily operations, but the body text lists six: Germany, the Netherlands, Austria, Switzerland, Poland, and Czechia. Source: SuperPanther press release source

Dongfeng's humanoid robot "Xiaodong" begins factory training in late October · humanoid ⚠️ plan-stage claim

Zhang Zhenlin, chief engineer for intelligent technology at Dongfeng Motor, said: "Xiaodong will enter factory training in late October, join trial production by the end of this year, and achieve small-batch mass production next year." It will primarily handle sorting, material handling, and quality inspection; the "cerebellum" component is being developed by Huazhong University of Science and Technology. Source: Changjiang Net source

"Embodied intelligence robot application technician" added as a new occupation in China · adjacent

China's Ministry of Human Resources and Social Security and other departments released 11 new occupations and 23 new job types, with embodied intelligence robot application technician among them. The ministry says it will next accelerate development of standards for the new occupation. Source: Sohu source