Key Takeaways Today

· Agility adds 4 new customers and is building another 50,000-square-foot deployment and service base in Pittsburgh

· Sereact and Universal Robots say they have delivered hundreds of UR dual-arm systems to European customer sites for returns processing and picking

· SuperMimic trains on a single GPU for 3 hours and beats a humanoid motion-tracking baseline trained for 7 days on 128 GPUs

· Four flagship large models face off on RoboColiseum: the best scores only 46.70 on general manipulation, while all four specialized policy models score above 63

· NVIDIA's driving VLA Alpamayo-1.5 sees obstacles but does not slow down; adding guidance at generation time cuts the vehicle-body intrusion rate from 58.1% to 14.5%

SuperMimic: a differentiable world model that cuts humanoid whole-body control training compute by a factor of 1,000locomotion

Tracking large-scale human motion data is the mainstream route to general-purpose humanoid whole-body control, and compute is the bottleneck. Existing approaches either distill and merge per-motion expert policies, or brute-force train a single policy on a cluster. SuperMimic uses a differentiable world model to learn the dynamics shared across all motions, and the policy takes analytic gradients directly on short-horizon imagined rollouts. Trained for 3 hours on a single GPU, it beats a baseline trained for 7 days on 128 GPUs in return, using more than 1,000 times fewer GPU-hours. On a benchmark of 129514 motions it fails on only 25, versus 179 for the baseline; trained for 9.8 hours on two GPUs, failures drop to zero. The policy has been transferred to a real Unitree G1 (Chinese humanoid robot maker), including acrobatic motions.

Qiayuan Liao et al. (UC Berkeley) · arXiv 2610.28480 source

RoboColiseum: four flagship large models go head to head; they lead on understanding but trail specialized policies on manipulationbenchmark

GPT-6 Astra, GPT-6.1 Sol, Claude Opus 5.5 and Gemini 4 Argon share the same Agent2Act harness and original instructions, and are scored on four tracks: instruction following, spatial reasoning, general manipulation and robustness. Opus 5.5 scores 76.28 overall, first among the four large models, with Astra close behind at 75.99; Astra is still strongest on instruction following and robustness. On general manipulation, the four large models score only 38.79 to 46.70, while the four specialized policy models score 63.20 to 66.13. The top two on the overall leaderboard are both policy models: Paimon at 77.54 and HiDream-O1-Embodied at 76.92. In an ablation, removing execution feedback raises Astra's average number of planning decisions from 16.7 to 35.6. Two weeks ago, on Artificial Analysis's AA-Robotics leaderboard, Opus 5.5 ranked behind Astra, Sol and Argon; the tasks and harness differ between the two evaluations.

Da Huang et al. · arXiv 2610.27865 source · Commentary: Robot Dajiangtang (Chinese robotics media) source

Driving VLA sees the obstacle but doesn't brake: perception and action are decoupled in Alpamayo-1.5autonomy

Driving VLAs with reasoning can recognize and describe hazards, but their trajectories don't necessarily change accordingly. Using matched simulated counterfactuals and real-world replays, the authors show that an obstacle ahead is already encoded in the visual representation, yet longitudinal planning mainly follows the ego vehicle's historical speed and responds only weakly to obstacle distance and to the model's own reasoning chain. Without changing any weights, the authors inject an obstacle-conditioned energy term into Alpamayo-1.5's frozen flow-matching sampler. The vehicle-body intrusion rate falls from 58.1% to 14.5%, and collisions drop to zero in the closed-loop test group that requires braking. Marco Pavone is a listed author.

Matteo Cercola et al. · arXiv 2610.28184 source

GPT-6 Astra controlling a robot arm directly: only 11.1% zero-shot on the new LIBERO-GeoMotion benchmarkvla

This paper systematically tests GPT-6 Astra on closed-loop manipulation. It can use depth information, learn simple policies from in-context demonstrations, and correct course using tactile feedback; its success rate is 76.0% on RoboTwin 2.0 and only 26.92% on RoboDojo. The authors built a separate benchmark, LIBERO-GeoMotion, that specifically tests precise pose alignment and coordinated translation-rotation motion, where Astra's zero-shot success rate is 11.1%. On real robots it can handle simple tasks, but there are gaps in motion feasibility, precision and safety, and inference latency also slows deployment.

Youwei Zhou et al. · arXiv 2610.28259 source

Ask it and it refuses; make it act and it complies: adding a runtime safety check to the robot control loopmanipulation

Following RoboHarm, which found that frontier models' refusal rates fall from 73% to 18%-25% once they control a robot directly, this paper offers a low-cost fix. GPT-6 Astra controls an SO-101 arm through the Inspect Robots harness; given common-sense hazardous instructions such as "put the knife in the toy basket," it refuses 0 times out of 15. It attempts the task every time, and 27% of attempts actually succeed. The authors insert one question before the first action, having the model look at the same scene and judge whether the instruction should be executed. All 15 hazardous instructions are blocked, and none of the 15 harmless control instructions is wrongly blocked. Code and data are public.

Arnav Grover · arXiv 2610.27350 source

AIM: world models ignore action commands because of the data, not model capacityworld-model

Action-conditioned video models trained on successful demonstrations often generate an expert-like future no matter what action is input. The authors define "action influence" and prove that it caps the controllability a model can express. Expert data has a narrow action distribution, so the cap is inherently low, and adding model capacity doesn't help. Their active data collection algorithm, AIM, on pixel-level manipulation tasks, uses 500 actively collected trajectories plus 3000 demonstrations to match the end-effector controllability of a model trained on 300,000 expert trajectories, with 85 times less data.

Samuel M. Bateman et al. · arXiv 2610.27387 source

PACT-VLA: turning asynchronous-inference latency into time for choosing action chunksvla

VLAs deployed on real robots mostly use asynchronous inference; by the time a new action chunk is committed, the observation is already stale, and the larger the latency, the more success rates drop. PACT-VLA requires no training and adds no verifier. It first generates multiple candidate chunks from the inference start point, then uses the fresh observation available before commit time to let the frozen policy score the candidates itself. With the same π0.5 checkpoint, average success across three real-robot tasks rises from 60% to 82%.

Yixiang Zhu et al. · arXiv 2610.28019 source

PhysCanvas: a tactile-aware world model whose simulated success rates track real robots more closelyworld-model

World models that only see images often get contact dynamics wrong, making them unreliable for evaluating policies. PhysCanvas puts multi-view RGB, tactile images and actions into a single video-generation backbone, then post-trains with a state reward model. On four real-robot tasks, the correlation coefficient between world-model and real-robot success rates rises from 0.602 to 0.885. Running GRPO post-training inside it raises π0.5's real-robot success rate from 59.1% to 70.8%.

Zebin Xing et al. · arXiv 2610.28613 source

One head-mounted camera plus a paper-marker glove replaces wrist cameras and pose trackersmanipulation

Handheld demonstration devices usually need wrist cameras, cables and trackers. This setup uses only a $122 head-mounted camera to read printed markers on a $22 glove with no electronics, recovering both the gripper's 6-DoF pose and its opening width. Median annotation error is 6.5 mm and 1.6 degrees. With 194 glove demonstrations plus 48 teleoperated ones, the success rate rises from 10/25 to 17/25, comparable to the 18/25 from 172 teleoperation-only demonstrations.

Yosub Shin et al. · arXiv 2610.27622 source

NVIDIA's COMPASS navigation tested on industrial AMRs: Cosmos augmentation lifts real-robot success from 65.3% to 91.7%autonomy

A team at Idealworks trained learned navigation in a digital twin of its own R&D center and deployed it on the Jetson AGX Xavier onboard an iw.hub, with controls being a version trained only on NVIDIA environments and the existing production navigation stack. After adding Cosmos-generated sim2real variants, real-robot success across four scenarios rises from 65.3% to 91.7%, and weighted transit time is nearly halved. It is still 8.3 percentage points short of the production navigation stack.

Charbel Abi Hana et al. (Idealworks) · arXiv 2610.28432 source

ASI (Air Space Intelligence, US) | Series C | $325 million | $3 billion valuationadjacent

Led by Spark Capital, with participation from an affiliate of Citadel Securities, Meritech, Craft Ventures, Energy Impact Partners and a16z. This Boston company is not the same as Utah-based Autonomous Solutions, which received SoftBank investment in September. Its AI continuously simulates the operation of physical assets such as aircraft, ships and power grids, anticipating congestion and load peaks hours to weeks ahead. Before this round it won a 12-year, $875 million air traffic management modernization program from the US Department of Transportation and the FAA, beating Palantir and Thales according to Bloomberg. The US Army also gave it a 5-year, $99 million contract.Source: AI Insider source

Arda (US) | New round | $230 million | $1.5 billion valuationindustrial

Co-founded by former OpenAI chief research officer Bob McGrew, the company trains robots on factory-work video so they can run factories autonomously. In February, the Wall Street Journal reported it raised $70 million at a $700 million valuation; eight months later the valuation has roughly doubled. The lead investor in this round was not disclosed.Source: Ebrun (Chinese tech outlet) source; Trending Topics source

Enigma (US) | Seed | $71 millionembodied

Led by Index Ventures and Ribbit Capital, with Conviction following, and founders and researchers from OpenAI, Anthropic, DeepMind, xAI and other companies participating as individuals. The company is less than a year old, and both founders, Jonathan Jacobi and Gal Niv, come from security backgrounds. It has put more than 100 in-house robot arms online for anyone to operate in real time, whether drawing, sword-fighting or mixing liquids in beakers. The company uses this interaction data to explore robot interaction interfaces and to train end-to-end models. No model is publicly available yet.Source: 36Kr (Chinese tech outlet) source

Pimax (Chinese VR headset maker) | Strategic investment from Mifeng Technology | Amount undisclosedadjacent

VR headset maker Pimax has spent more than ten years on multi-sensor spatiotemporal alignment, and is now applying that capability to recording human whole-body motion and intent to provide training data for embodied models. The investor, Mifeng Technology, was incubated internally by Zhiyuan Robotics (AgiBot, Chinese humanoid startup), and its chairman and CEO Yao Maoqing is a co-founder of Zhiyuan.Source: Sohu source; Ifeng Tech source

Zichuang Zhixie | Angel+ round | Tens of millions of yuanindustrial

Led by Western Capital, the investment arm of Western Securities. This is the company's second round this year; the previous round was funded by Yida Capital. It converts loaders and excavators into drive-by-wire unmanned machines and is running real-world operational validation at ports and concrete mixing plants.Source: Sina Mobile source

Jingji Zhineng | Seed | Tens of millions of yuanembodied

The company was founded only in September and a month later received investment from Haowen Tianshu Robotics and Shanghai Lanyun. It applies the force control and vision technology of medical robots to wiping and disinfecting walls, countertops and sanitary ware, and has signed partners in Australia, Spain, Singapore and Vietnam. The product is still in validation.Source: DongDian Tech source

Sensing4D AI | Nasdaq IPO | F-1 filedadjacent

A Chinese 3D vision company whose platform does both 3D reconstruction and simulation content generation for embodied training. About 99.7% of its fiscal 2026 revenue came from project-based integrated solutions. The ticker and price range have not been set.Source: TradingView source

Agility adds 4 customers; Pittsburgh base expands to 50,000 square feethumanoid

Two days after announcing the AGCO pilot, Agility disclosed 4 new customers in automotive, consumer goods and agriculture, whose names were not released. The new Pittsburgh facility is 50,000 square feet, a net increase of about 35,000 square feet over the existing space. It handles Digit's skill development, customer workflow validation and deployment, and also serves as an East Coast service center. Digit currently works at GXO, Schaeffler and Toyota's Canadian plant, and multi-year orders for Digit 5 exceed $300 million, to be realized on contract milestones. CEO Peggy Johnson said: "Scaling humanoid robots means building the infrastructure that supports the robot across its entire working life."Source: GlobeNewswire source

Sereact teams up with Teradyne: hundreds of UR dual-arm systems already running at European customer sitesindustrial

Sereact's Cortex model drives Universal Robots arms. The two companies say they have delivered hundreds of dual-arm systems to European customer sites for returns processing, replenishment, picking and kitting. Teradyne Robotics chief commercial officer Jacob Pascual Pape said the solution is "well beyond the demo or pilot stage." After Cortex handled Zalando returns at CEVA's German and Polish warehouses in early October, Sereact on the same day published, together with NVIDIA, Nebius, CEVA and Zalando, a training loop for the returns line. Clips recorded on site are first expanded into thousands of simulated variants in Isaac Lab, then re-rendered by Cosmos by garment condition, packaging and lighting. A newly post-trained policy goes back to the workstation for real-world testing only after it passes a threshold. By their account, an error made on Monday can turn into a tested policy update within days, where it used to take weeks.Source: Yahoo Finance UK source; The Manila Times source

Exotec connects two Skypod systems in Japanindustrial

Skypod has gone live at Mitsubishi Shokuhin's foodservice distribution center in Soka, managing nearly 13,000 SKUs, the first time Skypod has entered Japan's food distribution industry. Trusco Nakayama's largest logistics center, Planet Aichi, covers about 89,000 square meters, and its system has nearly 32,000 totes, 37 robots and 4 picking stations, with the SKU count expected to exceed one million in the coming years. Both projects were delivered in partnership with IHI Logistics & Machinery.Source: Robotics & Automation News source

Genon's AI shoe-sole inspection passes Nike mass-production validation; goes live in 2027industrial

Under real production conditions, defect detection accuracy on the outsoles of two shoe models is 99.7% and 99.2% respectively. Production-line deployment begins in 2027. The South Korean company has since also won a contract for upper inspection and an outsole press-fitting project that uses collaborative robots plus vision, taught remotely by workers.Source: The Korea Times source; Business Wire source

Xingji Guangnian's dexterous hand validated for months on Pantum's printer final-assembly lineembodied

Xingji Guangnian signed a strategic partnership with Pantum (Chinese printer maker) and the two will co-build a dexterous-hand joint lab. Xingji Guangnian says its tendon-driven dexterous hand plus in-house manipulation model has completed more than 20 fine operations on Pantum's line, including part recognition, grasping, aligned insertion and assembly, at a cycle time that keeps up with the line. The hand weighs 290 grams and has a grip force of over 500 N. Neither the number of units deployed nor the time to move into regular production was disclosed.⚠️ Vendor claimSource: Ifeng Tech source

Shenmou's arm-equipped quadruped "Fumo" wins a State Grid cable tunnel projectembodied

The project is titled "Quadruped Robots for Complex Environments"; neither the amount nor the number of units was disclosed. Shenmou says older cable tunnels still make up more than 60% of the total, most with no power or network, and that existing quadruped inspection can only read gauges. Its solution carries a robotic arm and dexterous hand and can work without power or network. State Grid is China's main grid operator.⚠️ Vendor claimSource: Sohu source

NVIDIA GTC Berlin: reference humanoid Agile ONE SN launched; Agile Robots says up to 2,000 unitshumanoid

Agile ONE SN is built by Germany's Agile Robots, comes preloaded with Isaac GR00T and runs on Jetson Thor, putting simulation, training and deployment tools on a single machine. NVIDIA says this cuts environment setup from weeks to days. It ships in early 2027, and Stanford and ETH Zurich have expressed interest. According to Reuters, Agile Robots CEO Zhaopeng Chen said the company will work with NVIDIA to produce up to 2,000 units for the research market. NVIDIA has also connected its Halos safety system to warehouse robots from Agility, Dexory, FORT Robotics and KION, with SICK evaluating it. Jensen Huang said: "The ChatGPT moment for robotics is approaching fast."Source: NVIDIA Newsroom source; Reuters source

Bolt picks Alpamayo and will collect road-test data in Madrid in 2027; Einride puts it in trucks for shadow testingautonomy

In Berlin, NVIDIA listed a batch of European Hyperion projects. Bolt selected the Alpamayo Stack as its robotaxi software, first installed on a future Hyperion-based Lucid model; pending regulatory approval, it will begin collecting real road-test data in Madrid in 2027. Germany is the priority market for later deployment, and the overall target of at least 25,000 vehicles is unchanged. Volkswagen Commercial Vehicles showed the ID. Buzz AD platform, which can support multiple autonomous driving systems. Einride will post-train and distill Alpamayo on its own freight data, first running in shadow mode without taking control of the vehicle, with real-road evaluation in the second half of 2027. Einride CTO Henrik Green said that as the first trucking company involved, it can "help shape this model to fit freight." A paper by Marco Pavone et al. published the same day shows that Alpamayo-1.5 can recognize obstacles ahead but often does not slow the trajectory.Source: Securities.io source; The Manila Times source

World Intelligent Connected Vehicles Conference (China): automakers give opposite judgments on when L3 will arriveautonomy

XPeng chairman He Xiaopeng argues for jumping straight from L2 to L4. He test-drove Audi's L3 feature in Europe, which was activated for only a few minutes in a day: "L3 shifts responsibility, but the experience isn't better and safety isn't improved." He estimates L4 technology will be deployed around the second half of 2028, with policy implementation between 2028 and 2032. BYD's Yang Dongsheng, by contrast, says L3 could be deployed in 2028. Wan Gang, honorary chairman of the China Association for Science and Technology, expects L3 penetration to exceed 35% in 2030; he also gave figures showing that as of the first half of the year, China has opened 57,000 km of test roads and issued more than 20,000 test and demonstration permits. NIO's Li Bin sets the yardstick at users' willingness to pay: "If users aren't willing to pay, there's no real user value to speak of." Only two models in China currently have L3 approval: the Deepal SL03 (Changan's EV brand) and the Arcfox Alpha S (BAIC's EV brand). Deepal chairman Deng Chenghao said that many earlier products claimed "L3-level" capability as a way of "skirting the edge of marketing rules."Source: Sina Tech source; Netcom (Wangtongshe) source; Sina Tech source; Sina Finance source; Sina Auto source

HD Hyundai Robotics unveils a five-finger dual-arm robot, starting with bundling and assembly tasksembodied

The robot is built as an upper-body form, with a payload of 7 kg per arm and a working radius of 630 mm, a five-finger hand and an in-house robot foundation model, with no need to program task paths one by one. Its first uses are in bundling and assembly. A wheeled base will be added next, followed by bipedal legs. The tactile component is being developed with Aidin Robotics, in which it took a stake in September.Source: The Elec source

Mitsubishi Electric lets a humanoid robot interface directly with CNC machine toolsindustrial

The system has three parts: an AI machining task planner, a CNC machine tool with physical AI, and a general-purpose humanoid robot. The humanoid handles fixturing, cleaning, moving finished parts, inspection and troubleshooting, without needing separately developed communication and control timing for each robot. Mitsubishi Electric says this is the industry's first such system, based on its own survey. It will be demonstrated at JIMTOF 2026 in Tokyo starting October 26.Source: AI Insider source

UBTech (Chinese humanoid maker) signs an exclusive distributor in Japan; large-scale deployment awaits local certificationhumanoid

On October 19, UBTech signed an exclusive distribution agreement with Japanese listed company Bitcoin Japan, covering 6 models including Walker S2, Tiangong and Cruzr S2. The exclusivity period is 36 months, counted from when Walker C1 completes Japanese local certification. Before certification, only sample units go to universities, research institutions and companies, and large-scale deployment in manufacturing and logistics begins after certification. Bitcoin Japan plans to sell complete units, offer RaaS leasing, and, with customer permission, turn robot-collected on-site data into a licensing business.Source: EE Times China source

Beijing's service robot output up 1.7 times in the first three quartersindustrial

Data from the Beijing Municipal Bureau of Statistics show that service robot output in the first three quarters grew by 1.7 times year on year, and industrial robots grew 47.6%. Over the same period, Beijing's automobile manufacturing fell 16.3%.Source: Shanghai Securities News source

Unitree launches the H2-X "Kung Fu Legend"humanoid

In the promo video, it lifts the front of a car with its bare hands and dents a car door with one kick, positioned as a personal bodyguard. Unitree has given only two specifications: it can jump down from a platform 4 meters high, and standing endurance is about 8 hours. Price has not been announced; the H2 in the same series sells for $29,900.Source: TechRadar source; Xinlan Net (Zhejiang) source