Key Takeaways Today

· Japan plans to certify L4 and L5 vehicles without steering wheels or pedals under a single standard, starting with vehicles produced in December

· BMW's four plants have used Embotech's factory autonomous driving to move vehicles more than 1 million times in total

· Agility's merger won SPAC shareholder approval and it lists on November 5; net sales in the first half were only about $174,000

· South Korea's first bipedal humanoid data factory has opened; after a budget cut it has only 12 robots

· 5,699 blind matches across 124 sites: no policy completed a task on the G1 humanoid

· Tutor Intelligence closed a $75 million Series B; the number of facilities using its robots nearly quadrupled in a year

RoboColosseum: distributed real-robot blind testing; frontier agents tie with VLAs, humanoids all failbenchmark

Across 124 sites and 5,699 pairwise blind matches, RoboColosseum rolls out a unified policy-serving interface and standard procedure onto independently operated robots. It tested 5 VLA policies and 3 frontier AI agents on four embodiments: SO-101, Franka, bimanual YAM and G1. Frontier agents with no task adaptation already tie with specialized VLAs. After fine-tuning on 100 demonstrations per task, GR00T N1.7 leads on SO-101, while π0.5 leads on Franka and bimanual YAM; rankings are almost inconsistent across embodiments (Kendall W = 0.24). On the G1 humanoid's 5 tasks, no policy completed any, with average progress of only 13.3% to 23.2%.

Tan Sang Nguyen et al. · arXiv 2611.02497 source

Auriga: adding an execution layer to VLMs, topping RoboDojomanipulation

On RoboDojo, GPT-6 Astra used directly as the policy succeeds 22.5% of the time; removing the robot-arm execution step lifts that to 76.2%, so the gap comes from execution. Auriga is a task-agnostic agent harness: the model makes all task decisions, while the harness provides calibrated 3D measurements, previews the trajectory before execution, checks tolerances during execution, and halts and reports the deviation when it drifts. The frozen version tops the RoboDojo leaderboard with a 57.7% success rate, more than twice what the same model achieves with the reference harness. The code is open-sourced.

Chengcheng Wang et al. · arXiv 2611.01112 source

uni-robot-agent: an open-weight model as a robot-use agent, outscoring GPT-6 Astramanipulation

HF 38↑. Robot agents have long relied on closed-source frontier models, leaving researchers unable to keep training them; this unified harness makes open-weight models usable too. With no privileged inputs, an out-of-the-box Qwen3.8-Flash-Next scores 31.39 on RoboDojo, above the published 28.97 for GPT-6 Astra-RoboProbe, and is also competitive on 19 real-robot tasks.

Jiarui Yuan et al. · arXiv 2611.01560 source

Xiaomi Auto World Model 2.0: a streaming causal world model shared by driving and embodied AIworld-model

On the same day Baidu released Apollo World, Xiaomi Auto (Chinese EV maker) published the technical report for World Model 2.0. WorldRec 2.0 reconstructs static and dynamic 3D Gaussians from multi-view, multi-timestep observations in a single forward pass; the rendered output serves as explicit scene context that preserves geometry and appearance in long-horizon generation. The temporal layers switch to Gated Linear Attention, cutting compute and KV-cache use in multi-view rollouts. The model is already used for interactive simulation, policy training and world action models (WAM). The abstract gives no quantitative results.

Zhenxin Zhu et al. (Xiaomi Auto) · arXiv 2611.01677 source

Pixel: a mobile-manipulation full stack with hardware, data collection, control and learning co-designedmanipulation

Deepak Pathak's team at CMU puts four things that are often done separately into one system. The platform is a low-cost wheeled humanoid with human-aligned kinematics and a compliant gripper; data can be collected by teleoperation or directly by humans without a robot; there is a single whole-body controller, and human and robot demonstrations go through the same policy-learning pipeline. The companion Pixel-UMI (arXiv 2611.03647 source) learns environment-aware whole-body manipulation from human demonstrations with sensors.

Tony Tao et al. (CMU) · arXiv 2611.03648 source

Anchor 1: a precise humanoid controller for whole-body manipulationlocomotion

Archon Robotics replaces online joint-space retargeting with calibrated sparse motion references and separates error feedback from motion encoding. Under the strictest real-robot global scoring, tracking error is 30% to 80% lower than existing controllers, with less look-ahead input required; the authors also release a whole-body manipulation benchmark scored on precision, sustained stability and reachability.

Archon Robotics (Caojun Wang et al.) · arXiv 2611.00193 source

Validation loss cannot predict real-robot successbenchmark

On SO-101, the authors trained ACT, Diffusion Policy, SmolVLA and a π0.5 action expert, and compared 82 checkpoints against 930 bench trials. Across policies the ranking is completely inverted (ρ = +0.90): the policy with the lowest held-out error succeeded only 3 times in 110. The most common failure is the arm staying frozen at its initial pose, which a held-out set made of demonstrations can never reveal.

Simon Cowling · arXiv 2611.00283 source

Measuring post-training potential of pretrained policies with pass@kmanipulation

Shuran Song's team evaluates pretrained policies with closed-loop pass@k, the probability that at least one of k rollouts from the same initial condition succeeds. Checkpoints with similar pass@1 but different pass@k at large k also differ in post-training results; from this the authors propose an offline checkpoint-selection criterion that needs no real-robot evaluation, plus a generative data-augmentation method.

Zhanyi Sun et al. (Shuran Song's team) · arXiv 2611.01565 source

Roz: latents in world models do not remember physicsworld-model

Using 1,734 scenes to test mainstream latent-dynamics world models, the authors find they do well on tasks reducible to image similarity but fail to retain object mass, deformation, friction and elasticity. Roz only requires that a model can ingest still frames or video and expose its latents; the full test suite runs in a single forward pass.

Kenneth Ge et al. · arXiv 2611.00612 source

With only an onboard monocular camera, G1 learns to dodge ballslocomotion

Marco Hutter's team at ETH has a Unitree (Chinese robot maker) G1 estimate the incoming ball's position and threat from a sequence of detection boxes, then switch between two policies: dodge-and-recover and search-and-track. On the real robot, single-throw dodge success exceeds 85% (15+ trials); in simulation, for balls at 10 m/s from 2.5 m away, success falls to 33.2% to 40.8%.

Feng Tian et al. (ETH Zurich) · arXiv 2611.01197 source

Tutor Intelligence (US) | Series B | $75 million | valuation undisclosedindustrial

Led by a16z, with Series A lead Union Square Ventures and seed lead Neo participating, bringing total funding to $117 million. Tutor's robots Cassie and Sonny work in warehouses and manufacturing sites, charging by the hour or per picked item; remote operators step in when needed, and the operational data feeds back into training the company's own foundation model. SiliconANGLE reports that the number of facilities using its robots has nearly quadrupled in a year; Cassie uses vacuum grippers to depalletize and pack boxes, handling up to 18 items a minute and lifting loads of more than 50 pounds. CEO Josh Gruenstein said: "The demand is there, and we can barely build robots fast enough."Sources: The Robot Report source; SiliconANGLE source

Agility Robotics | SPAC merger (shareholders approved) | about $465 million | $2.5 billion pre-money equity valuationhumanoid

After the amount raised was finalized on October 30, Churchill Capital Corp XI shareholders approved the merger on November 2; if closing happens as scheduled on November 4, Agility will list on Nasdaq on November 5 under the ticker AGLT. Citing filings, 36Kr (Chinese tech media) reports that Agility's net sales in the first half of this year were about $174,400, of which about $171,700 came from deployment and professional services; cash used in operating activities over the period was about $86.2 million, and cash and cash equivalents at end of June were about $14.65 million. The company says it has secured deployment commitments covering nine customer facilities. Zhitong Finance (Chinese financial media) noted the timing of the listing: the Philadelphia Semiconductor Index fell 15.36% in October.Sources: The Robot Report source; Stockstar (via Zhitong Finance) source; 36Kr source

Sazhi Intelligent (Shanghai, China) | Pre-B | several hundred million yuan | valuation undisclosedindustrial

Investors include Jiantou Investment, Nantong Investment Management and Yunqi Partners (Chinese VC); existing shareholders include Shenzhen Capital Group (Chinese VC) and Shunwei Capital (Chinese VC). This multi-arm industrial robot company, founded in 2017, has a six-arm model, "Xuanjia", with up to 111 degrees of freedom. According to Hard Krypton, the company says it has already matched last year's full-year business volume in the first half of this year, expects 100% to 120% growth for the full year, and has overseas letters of intent worth over RMB 500 million; letters of intent are not confirmed orders.Sources: Sina (via OFweek Robotics) source; CNFOL source

Yicheng Innovation (China) | Series C | RMB 200 million | valuation undisclosedindustrial

Co-invested by Yueke Group, Shanda Investment, Anhui Guotou and Foshan Venture Capital. Yicheng (Chinese autonomous-vehicle company) makes L4 driverless short-haul vehicles for factories and campuses; the funds will go toward expanding delivery capacity at its own factory, iterating on its world model and building physical-interaction data. The company says embodied-AI-related orders amount to several hundred million yuan, without a breakdown of specific amounts.Sources: Embodied Home source; PEdaily source

MOVIN (South Korea) | Series A | KRW 10 billion | valuation undisclosedadjacent

Led by IMM Investment, with IBK Industrial Bank and Voyager Ventures participating and existing investors such as NAVER D2SF adding more. MOVIN uses lidar plus in-house AI for wearable-free, marker-free motion capture; one year after launch its product generated KRW 1 billion in revenue, 70% from overseas. The company says Boston Dynamics and the RAI Institute already use its technology for robot teleoperation and training.Sources: Asia Economy source

PlusAI (US) | SPAC merger | registration filed | amount undisclosedautonomy

Autonomous heavy-truck software company PlusAI has filed a registration statement for its merger with Texas Ventures Acquisition III, projecting annual revenue of over $40,000 per truck, with SuperDrive planned to launch on trucks in the US in early 2027. It neither builds vehicles nor operates fleets; its OEM partners include Traton, Iveco and Hyundai. Most of first-half revenue came from a Traton milestone agreement worth up to $25 million, with another $7.2 million from a safe note revaluation; the filing itself states that if its commercialization assumptions prove too optimistic, it may not generate operating cash flow.Sources: Trucking Dive source

Provision Labs acquires Basilisk Industries (India) | M&A | amount undisclosedadjacent

Anjali Sardana, founder of Pronto, started physical-AI data company Provision Labs in San Francisco in October and has now bought Mumbai hardware startup Basilisk; Basilisk founder Shourya Watwani will lead Provision's hardware. The company says that within weeks of the merger it built a 3D hand-tracking model and a first lightweight haptic glove. Neither company confirmed whether data collected by Pronto's home-services business will be used.Sources: Bharat Fast source

BMW's in-factory autonomous vehicle moves pass one millionautonomy

At four plants (Dingolfing, Leipzig, Regensburg and MINI Oxford), new vehicles coming off the line are driven away by Embotech's in-factory autonomous driving system, which has now completed more than 1 million vehicle moves in total. It currently averages about 3,500 a day, with system availability above 99.5%. The project started as a pilot in 2022; Embotech supplies perception, prediction and planning software, plus a separate deterministic safety architecture that has passed independent TÜV certification. Embotech CEO Andreas Kirtatos said: "It took us 15 months to complete the first 500,000 automated vehicle moves, and the next 500,000 took only 7 months." BMW board member for production Raymond Wittmann said: "We are currently rolling this technology out worldwide."Sources: ZDNet China source

AgiBot's convenience store 24-hour livestream wraps up, 6 G2 robots complete 1,730 taskshumanoid ⚠️ vendor-reported

The livestream that began on November 2 ended on the 3rd. According to AgiBot (Zhiyuan Robotics, Chinese humanoid maker), 6 Genie G2 robots completed 1,730 tasks at a store in Xuhui, Shanghai, including 683 customer receptions, 369 item fetch-and-deliveries and 147 coffees made; the company gave no figures for human takeovers or per-task success rates, and Interesting Engineering also said the stream cannot prove uninterrupted autonomous operation throughout. Ren Guanghui, vice president of the AgiBot Embodied Research Center, admitted to Allweather Tech that compared with skilled store staff, the robots still lag clearly in fine-manipulation efficiency, on-site adaptability and exception handling, and that stocking, replenishment and cleaning assistance will be deployed first.Sources: Allweather Tech source; Sina (via NetEase Intelligence) source; Interesting Engineering source

Ant Lingbo says its sorting robots are in routine operation at multiple Guoda Drugstoresembodied ⚠️ vendor-reported

The robots only began sorting medicines in September at real Guoda Drugstores (Chinese pharmacy chain) stores in Shanghai; at the Wuzhen Summit, Ant Lingbo said the solution is already in routine operation at multiple stores, taking orders, sorting and delivering in aisles 80 cm wide without changing the displays. Neither the number of stores nor per-store volumes were disclosed.Sources: Ifeng source

Japan plans single certification for L4 and L5 vehicles without steering wheels, applying to vehicles produced from Decemberautonomy

Autonomous vehicles with no steering wheel or pedals are currently reviewed one model at a time in Japan. Citing Nikkei Asia, EE Times China reports that the Ministry of Land, Infrastructure, Transport and Tourism is considering revising its safety certification standards so that L4 and L5 models are reviewed under uniform requirements; a subordinate working group meets on November 4 to discuss the details, and the new standard is intended to apply from vehicles produced in December. Passenger vehicles such as taxis and buses must still undergo dummy crash tests. Cargo vehicles may be exempted, provided the vehicle can avoid an accident when the system fails. In June this year the UN's WP.29 reached agreement on safety standards for driverless vehicles, expected to take effect in December, and Japan's revision is meant to align with it. The Japanese government aims to have at least 10,000 L4 buses and taxis on the road by fiscal 2030; according to an Autohome (Chinese auto portal) creator-platform article, most of Japan's existing L4 routes are low-speed short routes in suburbs, and the earliest approved one, the Eiheiji line in Fukui, is only 2 km long.Sources: EE Times China source; Autohome source

Baidu releases world model Apollo World, used to train Apollo Goautonomy

At Shenzhen AI Day on November 3, Baidu released Apollo World. It models vehicles and pedestrians in a virtual city as AI agents that make decisions and interact, generates camera and lidar data in real time, and lets driverless vehicles run an entire trip from order acceptance to drop-off between arbitrary start and end points. Citing Baidu, Reuters reports it generates up to 100,000 km of simulated driving per day, equivalent to about three years of a test vehicle with a safety driver, and it has already been used to train the Apollo Go (Baidu's robotaxi service) fleet. Wang Liang, chief architect of Baidu's Intelligent Driving Group, said a world model should not only be a "painter that generates cool videos" but also an "examiner" that verifies the safety of driverless driving. Frank McCleary of Arthur D. Little told Reuters that robotaxi companies "still have to bear the capital expenditure of training models ... and vehicles."Sources: Reuters (via The Economic Times) source; China Fund News source

South Korea's first bipedal humanoid data factory opens with only 12 robotshumanoid

On November 3, South Korea's Ministry of Science and ICT unveiled the AI humanoid data factory at the headquarters of the Korea Institute of Machinery and Materials (KIMM) in Daejeon. Jointly built by KIMM and ETRI, it is South Korea's first facility dedicated to producing data for bipedal humanoids: people teleoperate robots to do work, and the recorded motion and task data is used to train humanoid models. It currently has only 12 humanoids on site. Maeil Business reports that the ministry originally wanted to match the scale of China's facilities, but part of the budget was cut in the budget review; it now plans to add budget, expand the floor area from about 2,000 to 3,000 square meters and the robots to more than 100, and open it in phases to industry, academia and research from the end of 2027. Citing Interact Analysis data, Seoul Economic Daily says that as of end of April China had at least 90 humanoid data-collection and training centers operating or planned, of which at least 64 are operating and 13 have deployed more than 100 robots.Sources: Seoul Economic Daily source; Maeil Business source

FCC list: 13 operators of foreign-made robots conditionally clearedindustrial

According to an announcement by India-listed company ARAPL, on October 30 the FCC published a list of foreign-made advanced robots conditionally approved by the US Department of War, covering 13 operators in total. ARAPL subsidiary Humro's Maia2000 automated pallet truck and Atlas AC2000 automated forklift are on the list and are therefore not subject to the Covered List restrictions. The approval is conditional on companies continuing to carry out their submitted US production plans and passing updated product reviews.Sources: SMEStreet source

China's securities regulator plans to loosen support buybacks; Unitree's share price has fallen to half its early-listing leveladjacent

According to South Korea's JKN, the China Securities Regulatory Commission (CSRC) announced on November 2 that it plans to amend buyback rules: a share price falling below the issue price may serve as a reason for a buyback, and the restriction barring support buybacks within six months of listing is removed; companies listed for less than 36 months that trigger support conditions must carry out the price-stabilization measures they promised at listing. Before the rules are formally amended, the CSRC will encourage and supervise buybacks company by company among major index constituents and leading tech stocks. JKN says Unitree's share price has fallen to about half its level when trading began in August, and the STAR 50 index has fallen 40% since June.Sources: JKN source

ISO 21423 published, giving mobile robots from different vendors a common communication protocolindustrial

A3 announced the publication of ISO 21423:2026. Developed by ISO TC 299, the standard specifies how industrial mobile robots from different vendors, fleet management systems and other equipment exchange position, speed, direction, status, task and availability information, and also covers elevators, automatic doors, alarm systems and ERP. Carole Franklin, A3's director of robotics standards development, said: "Many large industrial users have been asking for an interoperability standard for years." It is the first non-safety robotics standard A3 has helped develop.Sources: Robotics & Automation News source

Japan's industrial robot orders rise 46.6% in the third quarter, a record high for the fourth straight quarterindustrial

Data from the Japan Robot Association (JARA) show orders in July to September reached JPY 325.4 billion, up 46.6% year on year and above JPY 312.7 billion in April to June. The statistics cover member companies such as Fanuc, KUKA, ABB and Yaskawa Electric, and exclude service robots.Sources: EE Times China source

Cainiao's climbing robot now picks bins from four sidesindustrial ⚠️ vendor-reported

The new version shown at CeMAT ASIA 2026 can pick bins from four faces, front and back on both sides of the aisle; Cainiao (Alibaba's logistics arm) says operating efficiency improves by a further 15% and storage density by 20%. Yang Bin, general manager of the logistics robot business unit, said overseas orders already account for 40% and are expected to rise to 80% within two to three years.Sources: Sina Finance (via China Business Journal) source