Today's Highlights

· Waymo opens invitation-only Robotaxi service in Las Vegas, launching with dozens of Ojai vehicles in its 15th market

· Two WeRide Robobuses at Zurich Airport remove in-vehicle safety operators, now monitored remotely from a control center

· Hai Robotics updates its Hong Kong IPO prospectus, reporting first-half revenue of RMB 1.118 billion and an adjusted net loss of RMB 208 million

· An 11.5-ton excavator uses a reinforcement learning policy to autonomously build a 42-meter embankment in 45 minutes, with 201 scoops and zero failures

· Universal Robots launches its Gen 7 collaborative robot platform, with gigabit Ethernet built into the flange and torque sensing built into the arm

· Unitree releases the G1+, with the standard version priced at RMB 95,000 including tax

· AgiBot's (Chinese humanoid robotics startup) first-half humanoid shipments: Counterpoint records 9,700 units, Smart Analytics Global records 8,400 units

One Set of Weights Across Machines: Reinforcement Learning for Material-State-Aware Excavator Earthmoving · manipulation

Existing autonomous earthmoving systems can basically only dig and dump, whereas human operators use every face of the bucket to push and compact soil. The authors train an RL policy inside a GPU-parallelized Material Point Method particle simulator, conditioning on soil shape and compaction state and outputting actions in a normalized end-effector space. The same set of weights is deployed on both an 11.5-ton hydraulic excavator and a 500-gram tabletop robot. On the real machine, it builds a 42-meter-long, 2.1-meter-high embankment in 45 minutes, with 201 scoop motions, no failures, no retries, and no human intervention, matching the pace of an expert operator.

Lennart Werner et al. · arXiv 2609.12677 source

Breaking the Vision-to-Action Shortcut: Latent-Interface Training for Generalizable Robot Foundation Models · vla

HF 61↑. Robot foundation models lose accuracy when the camera position or lighting changes, partly because the model latches onto irrelevant visual cues that happen to correlate with actions in the training distribution. LIT trains in two steps: first, an action expert is trained without images, using only language, robot state, and end-effector goal poses; then a latent interface, tasked solely with reconstructing the goal pose, is added as the sole channel through which vision reaches the action expert. The method is architecture-agnostic, improving LIBERO-Plus success rates by 3.87-10.70 percentage points across four VLA and world-action architectures including Pi0.5 and MolmoAct2, with real-robot gains of 13.30-16.70 percentage points under unseen camera positions, lighting, and distractors.

Jianman Lin et al. · arXiv 2609.12641 source

Pelican-Sim 1.0: A General-Purpose World Model Simulator for Embodied AI · world-model

Using a world model as a simulator is hard because action controllability doesn't transfer well across different robot embodiments. Pelican-Sim uses a unified 28-dimensional action space covering mainstream embodiments, injects action videos rendered from URDFs into the model paired with a sparse MoE, and then distills it into 4-step autoregressive generation, 5.67x faster than a 35-step version. Training data comprises roughly 1 million real and simulated trajectories. On RoboTwin, adding 500 generated trajectories to 50 demonstrations per task raises policy success rate from 70% to 93%.

Shilong Zou et al. · arXiv 2609.12036 source

FoldNet++: A Large-Scale Synthetic Dataset for Robotic T-Shirt Folding and Unfolding · benchmark

Large clothing deformations have long made folding policies hard to generalize. The team generates physically simulatable T-shirts and synthesizes demonstrations keyed to semantic keypoints for 6 embodiments, totaling 1,000 T-shirts, 1,000 environment assets, and 120,000 episodes. A policy trained only on this synthetic data, deployed directly to unseen real environments and T-shirts, achieves an end-to-end success rate above 90%.

Yuxing Chen et al. (Peking University) · arXiv 2609.12433 source

DATAFARM: Making Planner-Generated Data Actually Usable for VLA Fine-Tuning · vla

Task and motion planning (TAMP — symbolic task planning combined with motion planning to automatically produce trajectories) makes generating demonstrations cheap. Yet the authors find raw TAMP trajectories barely help when fine-tuning a pretrained VLA, because planner actions are misaligned with the pretraining data distribution. DATAFARM aligns joint configurations, motion style, and timing during generation itself, raising the average success rate from 8.3% with raw TAMP to 56.7%, close to the 61.7% achieved with human teleoperation data. Code is publicly available.

Samrat Sahoo et al. · arXiv 2609.12316 source

IMPLY: Using Physical Parameters to Check the Consistency of World Model Rollouts · world-model

Current metrics for checking world-action model consistency only look at whether several rollouts agree with each other, without touching physics. IMPLY inverts the simulator to read out the mass and friction implied by each rollout, then scores it against two calibration pushes the model has actually observed. On V-JEPA 2-AC, the self-consistency metric picks the correct evidence only 52% of the time, no better than random; after anchoring, this rises to 73%, with correlation to rollout error reaching 0.92-0.99.

Aman Mehta et al. · arXiv 2609.12441 source

STAR: Sparse Tactile Representation Learning for Vision-Tactile-Language-Action Models · manipulation

Dexterous hands lack real tactile data, and tactile signals are themselves sparse and hard to use. The team built its own bimanual dexterous manipulation platform, collecting 200 hours of synchronized vision, tactile, and language data across 65 tasks and 10,576 trajectories. Training combines joint vision-tactile pretraining, tokens that mix sparse and global tactile representations, and future tactile prediction. With 100 post-training trajectories per task, average success rate across four real-robot tasks reaches 61%.

Xiangcheng Liu et al. · arXiv 2609.12549 source

ArtManip: Category-Level In-Hand Manipulation of Articulated Objects · manipulation

Manipulating objects with internal joints in-hand requires simultaneously controlling the object's joints and maintaining a stable grasp; the authors call this the first category-level method to generalize across instances and initial grasp poses. An automated pipeline procedurally generates articulated objects and synthesizes functional grasps, with the policy trained in two stages while randomizing articulation physical parameters during training. Experiments cover four object categories, with zero-shot real-robot transfer to 12 objects of varying shapes and joint structures.

Yang Yang et al. · arXiv 2609.12498 source

Hai Robotics (Chinese warehouse robotics maker) ｜ Hong Kong IPO ｜ Updated Prospectus ｜ First-Half Revenue RMB 1.118 Billion · industrial

First-half 2026 revenue was RMB 1.118 billion, with a gross margin of 34.4%, a net loss for the period of RMB 430 million, and an adjusted net loss of RMB 208 million. Orders in the same period totaled RMB 2.353 billion, about 2.1 times revenue; 20 key customers accounted for 87.1% of that total. Hai Robotics makes case-handling autonomous case-moving robot (ACR) systems for warehouses; revenue rose from RMB 800 million to RMB 2 billion between 2023 and 2025, while the period loss narrowed from RMB 1 billion to RMB 828 million. Cash and cash equivalents stood at RMB 1.417 billion as of end-June. Wuyuan Capital (Chinese VC) holds a combined 14.53% stake, General Atlantic holds 12.06%, and Sequoia (China) holds a combined 4.56%. Source: Ledian.com source

Inx Robotics (Nanjing) ｜ Series B ｜ Over RMB 300 Million · hardware

CITIC Jinshi, Nice Group, Suzhou Venture Capital, Huarui Investment, Nanjing Communications Holding, and Huarui Venture Capital are new investors, while existing shareholders including Fosun RZ Capital, Shenzhen Capital Group, and Huakong Fund also followed on. Inx Robotics, founded in 2022, makes joint modules for embodied robots; in June it released a second-generation quick-release joint module, which the company says cuts maintenance time to the minute level. Source: Sohu source

Lingchayun Control (Shenzhen) ｜ Series C+++ ｜ Amount Undisclosed · hardware

According to Tianyancha, Xichen Capital and Huakong Fund participated in this round, with DoNews reporting Xichen Capital as lead investor. Lingchayun Control makes robot joints, hollow absolute encoders, and high-power miniature servo drives. Huakong Fund is also an existing shareholder of Inx Robotics; both companies make joint modules. Source: Stcn.com source

Boston Dynamics ｜ IPO Timeline ｜ Unlikely Within 2027 · humanoid ⚠️ Single-source account

Reuters, citing an anonymous Hyundai Motor Group executive, reported that Boston Dynamics is unlikely to go public within 2027, citing that Atlas has not yet entered large-scale deployment and the company remains unprofitable. Hyundai Motor Group had previously set a target of building annual production capacity for 30,000 units of Atlas by 2028, with plans to deploy 25,000 units across Hyundai and Kia plants. Source: IT之家 source

Harmoni (United States) ｜ Series A ｜ $10 Million · industrial

Bessemer Venture Partners led the round; valuation undisclosed. Harmoni also launched a shop-floor AI assistant called HAL the same day, which workers can ask directly why a machine has slowed down or use to submit a repair ticket. The company currently has 40 customers and a 19-person team; 36Kr reports its target market as roughly 280,000 factories across the United States still relying on traditional equipment and manual labor. Source: 36Kr source

Waymo Opens Robotaxi Service in Las Vegas, Launching with Dozens of Ojai Vehicles · autonomy

Waymo began issuing invitation codes to the public in Las Vegas in batches starting Monday, launching with "dozens" of new Ojai vans covering the Strip area south of Highway 589, extending to Boulder Junction. This is Waymo's 15th market, following invitation-only launches earlier this month in Denver, San Diego, and Tampa. Nevada's transportation authority has set a one-year cap for Clark County of up to 1,000 vehicles for Waymo, up to 5,000 for Tesla, and up to 1,000 for Uber (via Motional and Zoox), totaling 7,000 vehicles. Waymo's overall fleet currently numbers over 4,000 vehicles across a dozen-plus cities. Tesla Cybercab chief engineer Eric Early told a state regulatory meeting in August that the company would be "very happy and satisfied" if it could reach "2,500 vehicles, or a bit more" within a year. The cap Tesla was granted is double that self-stated target. Source: TechCrunch source

Two WeRide Robobuses at Zurich Airport Remove In-Vehicle Safety Operators · autonomy

Starting earlier this month, Zurich Airport has been operating two electric WeRide (Chinese autonomous driving company) Robobuses without a driver or in-vehicle safety operator, monitored in real time by the ZRH Innovation Hub's remote control center. Vehicles stop automatically when encountering obstacles, and remote operators then decide whether to intervene. The project began in March 2025, with the safety operator first moving from the driver's seat to the back row before transitioning to primarily remote monitoring; the two vehicles have logged over 15,000 kilometers combined. A roughly four-week passenger-free test is currently underway on a route between employee access Gate 101 and the maintenance area, not crossing any taxiways. Senior Innovation Manager Raphaël Glaesener expects staff to be able to resume riding once testing concludes. The pilot runs through the end of the year, and the airport is also studying whether a single remote operator could monitor multiple vehicles simultaneously. Source: Digital Today (citing Electrek and Zurich Airport) source

Swisslog to Provide Warehouse Automation for New Aircraft Maintenance Facility in Saudi Arabia · industrial

Gulf-region logistics provider Starlinks has selected Swisslog to deliver warehouse automation for a new aircraft maintenance, repair, and overhaul (MRO) facility in Saudi Arabia. The solution uses SynQ software to unify scheduling of a high-density AutoStore tote system and a Vectura stacker crane pallet warehouse; deal value and the airline client involved were not disclosed. Source: EIN Presswire source

Universal Robots Launches Seventh-Generation Collaborative Robot Platform, Gen 7 · industrial

Teradyne subsidiary Universal Robots launched Gen 7 at the IMTS trade show in Chicago, with a completely redesigned control cabinet, teach pendant, and tool flange, running on the PolyScope X operating system. Three new g-Series arms were introduced simultaneously: the UR10g-1750, UR17g-1300, and UR18g-950. Product pages list payloads of 8, 15, and 18 kilograms respectively (with extended payloads of 10 and 17.5 kg for the first two), and reach of 1,750, 1,300, and 950 millimeters. The flange directly provides 1Gb Ethernet and 24/48V power, eliminating separate wiring for cameras and high-bandwidth sensors; the arm has built-in torque sensing and impedance control, with joints rated IP65. The new CB7 Core controller carries inconsistent figures: the company states a 40% increase in compute power and 30% reduction in size, while the product page states a 20% size reduction. Teradyne Robotics President Jean-Pierre Hathout said "robots for the AI era need to be advanced, not complex," and stated UR has over 100,000 industrial deployments to date. Inbolt CEO Rudy Cohen said his company's vision software now runs directly on the UR controller, "no external vision PC needed, and no extra cabling."Source: Business Wire source, Unite.AI

Humanoid Robot IPO Review: Investment Bankers Say "The Standard Hasn't Changed" · humanoid

Following last week's news that regulators plan to raise the bar for IPOs, Yicai interviewed several investment bankers. One executive at a North China securities firm said "it's not really a tightening — the standard has always been this, strict scrutiny of IPO quality." A South China investment banker said the stricter review "isn't specifically targeting embodied robot companies, but applies to unprofitable companies generally," with regulators making a substantive judgment on a company's profitability outlook as early as the application-acceptance stage. Among embodied AI companies currently under review, Dobot (Chinese robotics maker) has not submitted its registration since passing its listing hearing on July 22; Deep Robotics and LimX Dynamics have seen no progress since entering the inquiry stage in May, with only Deep Robotics profitable in 2025 among the three. Unitree's stock touched a new low of RMB 465.58 intraday on September 14, closing at RMB 470, down 57.27% from its first-day opening price of RMB 1,100. Source: Yicai source

AgiBot's First-Half Humanoid Shipments: Two Research Firms Differ by 1,300 Units · humanoid

VnExpress, citing Smart Analytics Global data, reported that Chinese companies accounted for over 97% of global humanoid robot shipments in the first half of the year, with AgiBot leading at 8,400 units, up 562% year-on-year, followed by Unitree at 5,900 units. Fortune, citing Counterpoint Research data from late August, put AgiBot's first-half shipments at 9,700 units. Beyond shipment figures, there's another estimate of where these robots actually end up: Georg Stieler, a robotics analyst who consults for industrial companies in China, told Reuters that 50%-70% of humanoid robots produced this year may have gone to "data factories" rather than performing productive work. AgiBot co-founder Yao Maoqing said at the Fortune conference in Macau on September 8: "For humanoid robot companies today, it's genuinely difficult to build robots that can run reliably in factories." Source: VnExpress source; Fortune source

Unitree Releases G1+, Standard Version Priced at RMB 95,000 Including Tax · humanoid

Unitree released an upgraded version of its G1 humanoid, the G1+, with the standard version priced at RMB 95,000 including tax. Six changes include: a new 2-degree-of-freedom neck, motor peak torque increased by up to 110%, fused dual-vision plus LiDAR perception, a 6-microphone far-field array, a rear external power port, and optimized battery life. The full unit weighs about 35 kilograms with 25 degrees of freedom in the body, and rated battery life of 2 hours. The research-oriented EDU version expands to 70 degrees of freedom, with an optional high-compute module. Source: DoNews source

Pony.ai and GAC Lingcheng Unveil Fourth-Generation L4 Heavy Truck, Say Europe and Middle East Entry Within Two Years · autonomy

Pony.ai unveiled an L4 autonomous heavy truck developed with GAC Lingcheng (GAC's heavy-truck subsidiary) at the IAA Transportation show in Hanover, built on the T9 pure-electric platform with 9 LiDAR units, 3 millimeter-wave radars, and 13 cameras, with plans for scaled production within the year. Robotruck business head and vice president He Xing told Reuters the company "has already been in contact with several ports and some European government agencies" to discuss testing. Pony.ai currently operates roughly 200 L4 trucks on routes in China. ⚠️ Stated as planSource: Guandian.cn source; Reuters source

Tesla AI VP: Robotaxi to Run Around the Clock After Switching to FSD v15 · autonomy

Tesla's Robotaxi service currently operates on fixed hours in every market where it has launched. AI VP Ashok Elluswamy, responding to a user on X, said the next-generation FSD architecture, launching next month, will enable 24-hour passenger service. ⚠️ Stated as planSource: Sina source

Lanchong Embodied Releases Modular Humanoid "Xiaobai," Starting at RMB 9,800 · humanoid

Lanchong Embodied, a spinoff company from Xi'an Jiaotong University, released the Mantis Standard "Xiaobai," with quick-mount modular chassis, arms, waist, and head, priced starting at RMB 9,800. The unit stands 1.4 meters tall with 22 degrees of freedom, a single-arm payload of 7 kilograms, omnidirectional movement speed of 3 meters per second, and 3-hour battery life, targeting embodied AI research and training use. Source: DoNews source

Foxconn and AIT Form Robotics Joint Venture in Beijing · industrial

According to Tianyancha, Beijing Futekang Intelligent Robotics Technology Co., Ltd. was recently established with registered capital of RMB 50 million, with shareholders including AIT Automotive Technology, Foxconn Technology Group, and a limited partnership entity; its registered business scope includes industrial robot sales and specialized-task robot manufacturing. Source: PANews source