Today's Highlights

· Waymo's next-generation robotaxi is set to be the Toyota bZ4X, with sensors installed directly on the assembly line at Toyota's Takaoka plant; testing in Japan first in 2027

· Aeon's Kawaguchi logistics center is officially in operation with 90-plus robots deployed, cutting about 39 workers a day

· Agility's listing raise is finalized at about $465 million, roughly a quarter less than expected when the registration went effective

· Hyundai Mobis has signed to sell its lamp business for about KRW 600 billion and starts prototyping Atlas actuators next month

· UBTech (Chinese humanoid robot maker) tested GPT-6 Astra: all 3 tasks it controlled directly failed

· On 25 robot simulation tasks chosen based on US GDP, the strongest frontier model averages a reward of only 0.06

Gym-Anything for Robotics: Having Coding Agents Build Robot Simulations in Bulkbenchmark

Robot simulation environments have long been built by hand by experts; this paper hands the job to coding agents. Using US GDP data, the authors pick out economically valuable manual labor, have agents write the simulations, and then check each one with a round of adversarial review. The resulting RoboWorld contains 25 simulations, 25 test environments and 175 training environments; the strongest frontier model fails on every test task, with an average reward of only 0.06. The pipeline, RL environments and benchmark are all open-sourced.

Pranjal Aggarwal et al. (CMU) · arXiv 2610.40565 source

Cross-Controller Transfer: One Policy Reads Six Controller Types In Contextmanipulation

When generalist policies mix data from multiple robots, they assume each robot has only one controller, yet the same action command means something different physically once the gain or compliance changes. Chelsea Finn and Dorsa Sadigh's groups describe the current controller with a few "commanded action vs. actual motion" examples, which a transformer encodes into context so that a single policy adapts on the spot, covering six controller types. On real robots, average success rate is 33 percentage points higher than directly mixing the data, and reaches 50% on controller modes never seen before.

Jonathan Yang et al. (Stanford) · arXiv 2610.39280 source

Data "Inoculation": Deliberately Perturbing Teleoperation During Collectionmanipulation

Sergey Levine's group changes the data-collection step itself. While a human teleoperates, an online-trained residual policy adds perturbations to the commands, pushing in the direction of mistakes the eventually learned policy is likely to make; the human keeps correcting, so the data gains recovery demonstrations. On three real-robot tasks, policies trained on data collected this way match or exceed the baseline success rate under both behavior cloning and offline RL.

Qiyang Li et al. (UC Berkeley) · arXiv 2610.39054 source

Metatron: A Prompt-Schedulable Whole-Body Execution Layer for Humanoidslocomotion

Humanoid controllers are usually trained separately for each task. Metatron trains just one generative foundation policy, whose input is a "prompt" assembled from partial goals and can mix timed body-part targets, velocity commands and style. Once trained it is frozen; task logic is written as ordinary code, which a coding agent rewrites after analyzing motion data. On a real Unitree G1, one frozen checkpoint simultaneously supports walking, fall recovery, teleoperated object pickup and sitting on furniture.

Chen Tessler et al. · arXiv 2610.38027 source

ScheduleGen: Automatic Multi-Arm Demonstration Generation with Schedulingmanipulation

Synthesizing new demonstrations from a small number of examples is a common way to supplement multi-arm data, but existing algorithms ignore constraints between arms, so the generated trajectories often collide or fail. ScheduleGen folds planning and scheduling into data generation: across 28 simulated task variants, data generation succeeds 79.6% of the time on average, and the policies trained on it average 65.5%. With zero-shot sim2real and real-to-sim-to-real transfer, success rates on eight real-robot variants range from 50% to 80%. Yuke Zhu and Ajay Mandlekar are among the authors.

Neel Jawale et al. · arXiv 2610.39429 source

EgoContact: Using Egocentric Video to Test Where a Robot Should Touchbenchmark

Egocentric video is mostly used for pretraining; this paper turns it into an exam: given the frame before contact and the intent, the model must predict which pixel on the object the hand will land on. The dataset contains 22,500 human-verified interaction events, and existing methods in in-the-wild scenes often cannot even find the target object. Used as a memory bank for an existing affordance-transfer policy, it raises real-robot grasp success from 16% to 58%. Marc Pollefeys and Daniel Cremers are among the authors.

Adam Deryło et al. · arXiv 2610.39167 source

UMI-Live: An Open-Source Bimanual Handheld Data-Collection System With No External Calibrationmanipulation

UMI lets people demonstrate with handheld devices instead of teleoperating a real robot, but it has to be recalibrated whenever the environment changes. UMI-Live packs cameras, lidar-inertial sensing, onboard compute and power into the device, so no external tracking setup is needed on site; SLAM reconstruction runs in parallel in the background and directly produces trainable data. After fine-tuning π0.5 on its data, bimanual pick-and-place succeeded 10 out of 10 times. The hardware design, software and 40 hours of data are all open-sourced.

Govind Velamoor et al. · arXiv 2610.39385 source

SimRetract: Trained Only in Simulation, a Robot Arm Takes Over a Retractor From a Surgeonmanipulation

In open surgery, someone has to keep holding the retractor. SimRetract's policy is trained only on simulated demonstrations, with no human demonstrations and no fine-tuning on a real robot; the task is to take the retractor from the surgeon's hand and keep the tissue retracted. Across 50 phantom scenarios set up by surgeons, zero-shot grasp success is 90% and retraction is maintained through the handover 78% of the time; removing the instrument mask alone drops the real-robot grasp rate to 22%.

Xucheng Wang et al. · arXiv 2610.38234 source

Catch It All!: A Humanoid With Dexterous Hands Catches Thrown Objects It Has Never Seenmanipulation

Using only a single egocentric camera, a Unitree G1 zero-shot caught 9 types of objects unseen in training, including a plush toy that barely bounces, a bag of chips, and hard objects with strong bounce, with both overhand and underhand catches working. A safety curriculum lets the policy output actions at high frequency even when the wrist torque limit is as low as 5 Nm.

Arth Shukla et al. · arXiv 2610.39044 source

Agility Robotics ｜ SPAC merger (raise finalized) ｜ about $465 million ｜ pre-money equity valuation $2.5 billionhumanoid

When the registration went effective on October 13, the raise was expected to exceed $620 million; after the redemption period ended, the figure landed at about $465 million (before transaction costs). After SPAC trust redemptions about $164 million remained, versus about $424 million at registration; the Foxconn-led private placement at $10 per share is about $201 million; another $100 million is equity investment Agility received after the deal was announced. Churchill XI shareholders vote on November 2, closing is expected on November 4, and the company will list on Nasdaq as AGLT on November 5. On a podcast, CFO Michael Beer said the Digit 5's bill-of-materials cost at launch was about $150,000 per unit, corresponding to annual output of about 1,000 units, then falls to about $75,000, and to about $30,000 at 10,000 units a year. Chief Commercial Officer Daniel Diez said the company already has commitments at 9 customer sites and aims to reach 25 by the end of 2027. Source: AI Insider source

Hyundai Mobis ｜ Sale of lamp business ｜ about KRW 600 billion (about $449 million) ｜ Buyer: OP Mobilityhardware

The lamp business had revenue of about KRW 2.5 trillion last year, 4.1% of Hyundai Mobis's consolidated revenue. On September 30 it signed a share purchase agreement with the French auto parts maker OP Mobility, with closing targeted for the second half of 2027. On the third-quarter earnings call, the company said it will begin prototyping actuators for Boston Dynamics' Atlas next month, targeting mass-production supply in 2028, while expanding its robotics team and R&D staff. The reasons given for the sale are that the expanding product lineup is stretching R&D and capital spending thin, and that lamps turn over quickly and have limited links to electrification and robotics. Boston Dynamics' CFO has previously said that about 70% of Atlas's bill-of-materials cost is in actuators. Source: BigGo Finance source; Chosun Ilbo source

Opentrons (US) ｜ Series E ｜ $40 million ｜ valuation undisclosedadjacent

Avego led, with BroadOak Capital Partners participating; existing investors include SoftBank Vision Fund and Khosla Ventures. Opentrons makes lab-automation robots and says it has installed more than 11,000 units in over 80 countries, with 14 of the world's top 15 biopharma companies using them. The new money goes to software that lets AI design and run experiments, more automated instruments and data infrastructure. A week ago, Warp Sciences raised $215 million to "have robots do wet-lab work". Source: AI Insider source

Hangzhou Embodied AI Application Pilot-Scale Testing Base ｜ Capital increase ｜ RMB 200 million ｜ Unitree takes 10%adjacent

Unitree announced on October 29 that its wholly owned subsidiary Hangzhou Unitree Robotics plans to contribute RMB 20 million to the capital increase. The five parties together add RMB 200 million in new capital, raising registered capital from RMB 100 million to RMB 300 million; Hangzhou Hi-Tech Sci-Tech Innovation Group contributes RMB 73 million for a 41% stake and remains the largest shareholder. The base was founded in May 2025; in the first nine months of this year it had revenue of RMB 449,600 and a net loss of RMB 15.2047 million, and its debt-to-asset ratio at the end of September was 89.74%. Source: Robot Qianzhan source

Neuron Industries (US) ｜ Seed ｜ $10 million ｜ valuation undisclosedindustrial

FCVC and Obvious Ventures co-led, with Y Combinator and others participating. The company makes Cortex, an AI-native industrial controller: engineers describe equipment behavior in natural language, and the system generates readable control code, which is tested in simulation before deployment. It only made its public debut in August. Source: GlobeNewswire source

Coratia Technologies (India) ｜ Bridge round ｜ INR 180 million ｜ valuation undisclosedadjacent

MGF Kavachh led, and the money is mainly for fulfilling an INR 660 million contract signed with the Indian Navy. The company makes remotely operated and autonomous underwater robots; its customers also include Indian Oil and Tata Steel. Source: ETElectronicsWorld source

Aeon's Kawaguchi Logistics Center Goes Live With 90-Plus Robots Loading Trucks by Storeindustrial

Aeon's Kawaguchi Higashi XD logistics center officially began operation on October 30. The robot control software comes from Mujin, and more than 90 robots are deployed on site, including 82 AGVs that carry individual items to a staging area. The center supplies about 200 My Basket small supermarkets in the Tokyo metropolitan area. The step of loading trucks by store is handled by robot arms, which decide the stacking order themselves based on each item's weight and shape to keep stacks from collapsing. Aeon says the overall process automation rate reaches 50%, saving about 39 workers a day. Hiroki Yamamoto, president of logistics subsidiary Aeon Global SCM, said the robots are "3 to 4 times" as efficient as people. Aeon plans to raise the automation rate to 80% by 2030 and then roll it out to logistics facilities for the group's supermarkets and drugstores. Source: Digital Today, citing ITmedia source

CiDi's 117 Driverless Mining Trucks in Routine Operation at Jilangde Coal Mineautonomy ⚠️ Vendor claim

As of August 2026, 117 driverless mining trucks from CiDi (希迪智驾, Chinese autonomous mining-truck company) are organized into 15 routine work groups at the Jilangde open-pit coal mine in Xinjiang, with driverless transport accounting for 53.6% of hauling. They have driven a cumulative 3.9 million km, completed 23.63 million cubic meters of rock and soil stripping and hauled 1.967 million tonnes of raw coal, with driverless and manned vehicles sharing the same routes and loading and unloading together more than 100,000 times in total. Inspections found that if a mining truck suddenly loses power, the parking brake may fail and the truck could roll on slopes; both delivered and pending vehicles have been retrofitted. Source: Phoenix New Media source

Linjiedian's Dexterous Hand Performs 72 Steps on Its Own Production Line, 30 Seconds Slower Per Cycle Than a Skilled Workerembodied ⚠️ Vendor claim

At a real workstation in the Shaoxing Linjiedian Wolong joint production base, the OmniHand dexterous hand autonomously ran performance tests on whole dexterous hands, with loading, code scanning and picking up a screwdriver to tighten, 72 steps in all. Proof-of-concept data shown at the breakout session of AgiBot (Zhiyuan, Chinese humanoid maker) Innovation Day on October 29: task success rate 98%, 8.5 minutes per cycle versus 8 minutes for a skilled worker, and continuous operation for 8 hours. The company's own wording is that it "preliminarily meets the conditions to join the line"; it has not formally joined the line. Linjiedian also disclosed that cumulative dexterous-hand mass production this year has passed 15,000 units as of October, and that the whole-unit failure rate fell from nearly 10% at the start to below 0.5%. At the end of September its wording was "cumulative deliveries of more than 15,000 units"; this time it changed to "cumulative mass production this year". Source: Shanghai Securities News source

Shunheng Intelligence's SynDexy C1 Makes Its First Small-Batch Deliveryhumanoid ⚠️ Vendor claim

Beijing-based Shunheng Intelligence (瞬恒智能) has delivered a first small batch of its dual-arm wheeled humanoid SynDexy C1, used for precision assembly steps at the factories of consumer electronics and electronic component customers; the number of units was not disclosed. The company says key-task success rates in earlier tests exceeded 98%. The robot has 22 degrees of freedom and carries in-house visuo-tactile sensors and SynDexy M1, a manipulation model fusing vision, touch and force. It took 6 months from the start of R&D in April to delivery in October, and the chief scientist is Fang Bin, a professor at Beijing University of Posts and Telecommunications. Source: Leiphone source

Kodiak's Long-Haul Driverless Readiness Reaches 99%autonomy ⚠️ Vendor claim

ARM is a metric Kodiak defined itself, measuring how many of the claims and how much of the evidence in its driverless safety case have been substantively completed; the value announced on October 30 is 99%. It was 84% in February, 93% in August and 96% in September. Once the safety case is finished, the first route is the 219-mile I-45 stretch from Dallas to Houston, run for IKEA, and the target is still to start before the end of 2026. CEO Don Burnette said: "This is the final sprint." In the Permian Basin, Kodiak's technology is installed on 35 trucks running with no one in the cab, owned and operated by customer Atlas Energy Solutions. Source: GlobeNewswire source

Tesla Robotaxi Now Runs 20 Hours a Day, and Cybercab Is on the Interstateautonomy

6 a.m. to 2 a.m. the next day is Tesla Robotaxi's new window for taking rides, announced by the official account on October 29, without saying which cities it applies to. On October 17 the end of service had just been extended from 11 p.m. to midnight, and last week it was extended again to 1 a.m. The same day, passenger Jesse Richards posted a video showing a Cybercab carrying passengers at 55 mph on an Austin interstate in the rain; until then only the Model Y had gone on highways. Texas Department of Motor Vehicles registration data show Tesla's Texas Robotaxi fleet has 1,080 registered vehicles, with Cybercabs now outnumbering Model Ys. Source: Drive Tesla Canada source; source

Venti's Driverless Yard Trucks Are Heading Into US Rail Terminals, With 130-Plus Plannedautonomy ⚠️ Plan-based

A US Class I railroad plans to put 130-plus Venti driverless trucks into service at 8 terminals by 2027 for short-haul container shuttling within the yards, with no one in the cab and no dedicated lanes. When the contract was signed in July the figure was "100-plus" for 2027, raised on October 1. The US fleet has not started work yet, and neither the railroad's name nor the contract terms have been made public. Venti's trucks have run at Singapore's PSA port for nearly three years; founder and CEO Heidi Wyle said that to enter the US, it first has to teach a 60-foot-long tractor to reverse with only 4 inches of clearance on each side. Source: FreightWaves source

Waymo's Next-Generation Robotaxi Uses the Toyota bZ4X, With Sensors Installed on Toyota's Own Lineautonomy

When it launches robotaxi service in Tokyo in 2027, Waymo will use the Toyota bZ4X. The vehicle comes off the line at Toyota's Takaoka plant in Toyota City, Aichi Prefecture, and Waymo's sensors and compute unit are installed directly on the final assembly line. Waymo's previous approach was to buy finished vehicles, ship them to another plant, and add sensors, computers and wiring harnesses there; The Verge estimates the new approach can eliminate duplicated labor, custom brackets and secondary shipping. According to Waymo, testing and validation will start in Japan next year, after which the vehicles will go to the UK and other international markets. The vehicles it runs on the road today are the Jaguar I-Pace and the Zeekr RT (Geely's EV brand), and a third, based on the Hyundai Ioniq 5, is about to be put into service. The two companies are also studying putting Waymo technology into Toyota's privately owned cars. In mid-September, Waymo had already agreed with GO and Nihon Kotsu to launch in Tokyo in 2027, starting small and gradually expanding to about 100 vehicles. Source: The Verge source

AgiBot Proposes "Worken" as a Work-Hour Unit and Self-Reports 9,700 Shipments in H1humanoid ⚠️ Vendor claim

At its Innovation Day on October 29, AgiBot co-founder, president and CTO Peng Zhihui proposed Worken, a unit of effective robot intelligent working hours, calculated from three factors: number of active robots, autonomous working time, and the output-to-energy-efficiency ratio. In their words, "Tokens measure how much a machine has thought; Workens measure how much a machine has acted." AgiBot disclosed that it shipped 9,700 units in the first half of 2026, claiming over 43% global share. The first-half humanoid shipment figure reported at the end of September was "over 8,600"; this time's figure does not say whether it counts only humanoids. The new model Genie G2 Max has a rated payload of 20 kg per arm and 40 kg with both arms, and the company says its per-box cycle time in mixed-line palletizing at a food company is under 35 seconds. The manipulation model GO-3 Preview is said to lead on three leaderboards: RoboDojo, RoboTwin and RoboColiseum. Source: Sina Finance source

UBTech Publishes Six GPT-6 Astra Tests: All Three Direct-Control Tasks Failembodied

Of the six tests, the real-robot ones are grabbing a Rubik's cube, grabbing a lucky bag and autonomous obstacle avoidance, and the simulated ones are carrying a basket, bowl stacking under direct control, and bowl stacking combined with a world model. With GPT-6 Astra controlling directly, cube grabbing, basket carrying and bowl stacking were all not completed. When Astra instead does the planning and calls UBTech's in-house Thinker-VLA for execution, or Thinker-WM is added to predict action outcomes, lucky-bag grabbing, obstacle avoidance and bowl stacking were all completed; on bowl stacking alone, Astra token consumption and estimated API cost fell by 83.6% and 74.2%, respectively. UBTech's conclusion is that there is still a gap between "can think" and "does well". In another test on October 19, GPT-6-Astra scored 100% on LIBERO-Long++, but on a real robot inserting a plug it failed in all 11 attempts. Source: The Paper source

Goldman Sachs: General Embodied AI Needs About 10 Million Hours of Interaction Dataadjacent ⚠️ Estimate

The Goldman Sachs report "China AI Robotics" estimates that general embodied AI needs roughly 10 million hours of high-quality multimodal interaction data to reach mass-production-grade capability. The figure comes from an October 28 expert call with a robot algorithm engineer. The report likens the maturity of general robot manipulation to autonomous driving before 2015, or around 2017 to 2018, and expects physical AI to progress 2 to 4 times faster than autonomous driving. The three engineering bottlenecks it lists are 3D spatial perception, inference latency and inference cost, and it says current spatial precision is "not sufficient to support millimeter-level physical interaction". Source: ANI source

The UK Gets Its First Insurance for Household Humanoid Robotsadjacent

Following Tokio Marine Nichido's launch of physical-AI robot insurance in Japan, UK insurer Adrian Flux has added household humanoid robots to its home insurance line, underwritten by Tokio Marine's Tokio Marine HCC. The policy covers damage to and theft of the robot itself, damage the robot causes to the home and belongings, personal liability for injuring people, and emergency retrieval or safe shutdown after a sudden malfunction. The two companies expect household robots to enter UK homes from 2027. Managing director Gerry Bucke said: "Robots are starting to enter homes, and we want to make clear who is responsible when something goes wrong." Adrian Flux launched a driverless-car policy in 2016. Source: The Register source

Wen Shilei, ByteDance's Head of Commercial GenAI R&D, Joins an Embodied-AI Companyadjacent

Wen Shilei has completed the departure process and joined the embodied-AI company Octopus Dynamics as a partner. Wen spent ten years on vision technology at Baidu and joined ByteDance in 2021. ByteDance's commercial GenAI line has lost all three of its core leaders within 4 months: in July, department head Peng Bingyue left and later joined Google; in the same month, China head Yuan Zehuan left to start a company in world models. Source: 36Kr source