Today's Highlights

· Tesla's Robotaxi arrives in Fort Worth with no safety monitor in the car, covering a service area of about 143 square miles

· With the instruction left blank, all 6 VLA checkpoints still moved in every episode, and adding "don't move if there's no task" did nothing

· Changing one constant in PhysX lifts RoboRwin's five container tasks from 89% to 98% success

· GenRobot (Chinese embodied-AI data startup) closes a US$100 million Series A+, with SAIC (Chinese automaker), Baidu and Galbot (Chinese humanoid startup) among the investors

· Unitree (Chinese robot maker) A2 wins the Mount Tai climb in 43 minutes 41 seconds; a G1 that entered overheated and was carried down the mountain on a stretcher

VLAs don't know when not to move: they go to work even on an empty instructionvla

With the instruction left blank, six VLA checkpoints still moved in every episode of the LIBERO-Goal simulation, and five of them often completed a task nobody had asked for. The authors fixed the initial state and varied only the instruction (empty, neutral, gibberish, infeasible, wrong task); on two GR00T checkpoints, the instruction determines which task gets done but not whether the robot moves at all. Prepending "don't move if no task is given" to the instruction stopped none of the six checkpoints; a gate comparing the first action chunk against the empty-instruction action chunk stopped the robot on five of them, while fine-tuning only prevented task completion and still did not reliably stop movement. Notebookcheck has already reported on this.

Minsu Jang et al. · arXiv 2610.41609 source

Some failures in simulation benchmarks should be blamed on the physics enginebenchmark

On RoboTwin, the authors traced episodes where objects are "flung away" to a single constant in PhysX's collision detection: the distance algorithm discards triangles below a fixed area threshold, and at centimeter scale it drops legitimate triangles too, so the engine reports roughly 2 cm of penetration and the solver then ejects the object at several meters per second. Changing only that constant cut episodes with the artifact from 12 to 0 out of 100, and lifted five container tasks from 89% to 98% success. The audit plugin SimuGuard supports LIBERO, RoboTwin, RoboCasa and RoboDojo; the authors recommend rerunning benchmarks on PhysX 5.3.1 and earlier, and for LingBot-VA, for example, the reported 83.1% actually corresponds to a range of 75.8% to 98.5%.

Alex Zhou et al. · arXiv 2610.41857 source

MolmoSpaces-CommonSense: the common-sense shortfall of zero-shot manipulation policiesbenchmark

A new benchmark from Allen AI tests common sense in manipulation, with questions involving occlusion, containment and support relations, which part of an object to grasp, and hidden objects. Several kinds of zero-shot generalist manipulation policies lose more than half their success rate as distractors increase, and in cluttered scenes they moved the wrong object 30% of the time; even with very specific prompts, they picked the correct part of the object only about half the time. Using a VLM to judge success or failure is likewise unreliable, and the authors argue existing benchmarks may overstate how usable zero-shot policies actually are.

Arjun Guru et al. (Allen Institute for AI) · arXiv 2610.41377 source

CaPtain: training a 4B open-source VLM into an agent that writes robot codevla

HF 18↑. Code-as-policy agents have until now either called closed frontier models or handed manipulation off to a VLA, and small open-source VLMs have a zero success rate out of the box. CaPtain first runs SFT on trajectories synthesized with GPT-6 Astra, then multi-turn RL, using a "physics critic" that reads privileged simulator state to assign credit to each API request; Qwen3.5-4B goes from 0% zero-shot to 73.8% on held-out LIBERO tasks. Code, data and weights are all open-sourced.

Yijia Fan et al. · arXiv 2610.40799 source

WorldBridge: action-label-free video doubles a world action model's cross-task success rateworld-model

HF 12↑. On RoboTwin 2.0, training on in-distribution video-action data alone gives 25.07% success on out-of-distribution tasks; adding video of those tasks without action labels raises it to 55.90%; adding one action-labeled demonstration per task reaches 70.70%. Based on this, WorldBridge separates a video expert from an action expert and connects them with structured attention, reaching 96.23% in the full-data setting on RoboTwin 2.0.

Jun Cen et al. · arXiv 2610.40748 source

REA: the fast policy in operation never waits for the "brain"vla

HF 11↑. When a large model acts as the "brain" and a VLA as the "cerebellum," the common approach asks the brain after every failure, which stalls the fast policy. REA instead routes using each policy's pre-recorded measured success rate, consults the brain for arbitration only at execution boundaries, and sets a shared budget for recovery. It reaches 95.9% success over 60 LIBERO-Pro tasks and 3,000 episodes, with 25% fewer control steps than the strongest monolithic baseline, and also runs on a real Unitree G1.

Pengfei Zhou et al. · arXiv 2610.42053 source

RoboArchitect: a coding model automatically rewrites the control code around a policy without changing weightsmanipulation

HF 5↑. A coding model watches robot replays, diagnoses failures, and repeatedly revises the harness wrapped around a frozen policy; average success rises from 9.7% to 43.3% in unseen scenes and from 12.7% to 40.2% on unseen tasks. Evolved behaviors include lifting the arm clear when it gets stuck and reaching in from the side when the area above is blocked; the harness also carries its own memory, recording effective grasp depths.

Shengguang Wu et al. · arXiv 2610.42141 source

∇Mimic: differentiable simulation gets humanoid motion tracking results in 20 minuteslocomotion

First-order policy learning has always suffered exploding gradients on contact-rich whole-body control, and the standard baseline SHAC collapses within minutes. ∇Mimic applies per-sample median-relative clipping at every step of backpropagation, together with NablaEngine, a differentiable GPU engine written for the Unitree G1 (adjoint computation 11.5 times faster than MuJoCo Warp); with a single GPU and a single policy tracking all 40 LAFAN1 motions, it reaches in 20 minutes the 0.82 score that the previous world-model method needed 1.2 hours to reach. The engine and code are open-sourced.

Weyl Lu et al. · arXiv 2610.41654 source

GenRobot ｜ Series A+ ｜ US$100 million ｜ Valuation undisclosedadjacent

Strategic investors include SAIC Motor, Malaysia's Eastworlds, Baidu Group, Shoucheng Holdings and Galbot, and financial investors include Photosynth Capital, SUNHELM, CITIC Jinshi, CSOP, CCB International and Huatai Zijin; the investment community says the round closed in the third quarter. Business registration changes in early October had already shown Baidu and Galbot taking stakes, and this time the amount was disclosed. The company provides data and evaluation for embodied AI and world models, collecting human first-person skill data through a crowdsourced production line; the company says its effective data deliveries rose from 100,000 hours per quarter to 1 million hours per quarter, and expects to reach 10 million hours per quarter in 2027. It will release the collection devices Ego 2.0 and Dex 2.0 during CoRL. Source: Pedaily source; FourWeekMBA source

RoboScience ｜ Series A2 ｜ Hundreds of millions of yuan ｜ Valuation undisclosedembodied

Amber Capital, Hongruida Investment and Shanghai Semiconductor Industry Investment co-led, with 360 Strategic Investment, Shuang Capital, Guolian Minsheng Capital, Jinding Capital and others following. The two founders were Stanford classmates: CEO Tian Ye previously led Apple's Apple Intelligence device-cloud inference system and Core ML at Apple, and chief scientist Shao Lin is an assistant professor at the National University of Singapore. Its in-house Visics model uses 3D point-cloud trajectories of objects as an intermediate representation, with the upper layer reasoning about object motion and the lower layer converting trajectories into control commands for each robot body; it comes with an in-house simulation engine, RoboMirage. Tian Ye said the cost of acquiring a single piece of data has been pushed down to CNY 0.05, 1/20 to 1/200 of traditional approaches. Source: Pedaily source

Yiwei Technology (02065.HK; Chinese robotics chip designer) ｜ Hong Kong IPO subscription ｜ About HK$673 million net proceeds at the midpoint ｜ Offer price HK$9.27–11.58hardware

After passing its listing hearing in late September, Yiwei Technology opened its subscription on October 31 and closes on November 9, and is expected to list on November 12 under Chapter 18C as a specialist technology company, with no greenshoe. Two cornerstone investors subscribed for a combined HK$99 million. About 70% of the proceeds will go to a robot-dedicated AI SoC. The company's 2025 revenue was CNY 407 million, with a net loss of CNY 155 million, wider than the previous year; its gross margin for the first four months of this year fell to 23.9% from 30.7% in the same period last year, which the company attributes to higher wafer and packaging-and-testing prices. Source: Finet source

ManuPrax Robotics ｜ Two consecutive rounds ｜ Nearly US$10 million cumulative ｜ Valuation undisclosedembodied

Alphaist Partners and Yaotu Capital each led one of the two rounds, with the angel investors and the actual controller of a components industry company following. Founded only in September, the company builds the "cerebellum" for dexterous manipulation, taking a reinforcement learning route, with the goal of a single manipulation model adaptable to different dexterous-hand hardware. CEO Li Puhao was born in 2001 and is studying in Tsinghua's Department of Automation; CTO Chen Zeyuan holds a master's from Peking University and has done research at Amazon FAR, AgiBot (Chinese humanoid startup) and JD.com. Qichacha shows the Beijing Institute for General Artificial Intelligence indirectly holding 5%. Source: DoNews source; Zhidx source

Tianxiu Intelligence ｜ Angel+ round ｜ Amount undisclosed ｜ Post-money over CNY 200 millionhardware

Led by Hangzhou Sci-Tech Innovation Group. The company spun out of Tianjin University and works on machine olfaction: olfactory chips, olfactory models and odor databases, with products including explosion-proof sniffing inspection robots, drones and electronic noses, and more than 30 customers including PetroChina, CNOOC and the Ministry of Emergency Management. Its manufacturing base in Binjiang, Hangzhou began operating on October 11. Source: Embodied Home source

Lingshu Zhifu ｜ Seed round + Angel round ｜ Over CNY 10 million ｜ Valuation undisclosedhardware

The seed round included Oriza Holdings and Huawei Hubble (Huawei's investment arm), and Oriza added to the angel round. A team from Fudan University's polymer science department is pursuing an ionic-skin route to flexible tactile sensors for data-collection gloves and dexterous hands; a materials pilot line is under construction, and samples are being sent to customers. Source: Embodied Home source

Tesla's Robotaxi arrives in Fort Worth, and Texas-registered vehicles reach 1,204autonomy

At 20:00 Beijing time on November 2, Tesla opened Robotaxi service in Fort Worth with no safety monitor in the car. It is the seventh city with unsupervised passenger service and the first new city in about three months; Tesla's previously announced next batch was Las Vegas, Phoenix, Jacksonville and San Antonio, not Fort Worth. The service area is about 143 square miles, behind only Austin's 272 and Houston's 254 among unsupervised cities, and starts with the Model Y. Drive Tesla Canada counts another 124 Model Ys registered by Tesla in Texas, taking Robotaxi registrations from 1,080 to 1,204, of which 631 are Model Y and 573 are Cybercab; the site compares this with a Waymo figure of 1,259 vehicles, and cautions that registrations do not equal vehicles on the road taking rides. Cybercabs are also appearing more often outside the Austin service area: a camera in the southwest recorded 33 sightings last week on the community tracker Robotaxi Tracker, versus only 1 in the last week of September, and Tesla has not announced an expansion. Source: Not a Tesla App source; Drive Tesla Canada source; TeslaNorth source

FFR sold and shipped 351 robots in Octoberhumanoid ⚠️ Vendor-reported

The 351 units are more than a third above September's 265 and above the company's October 26 forecast of "over 300"; cumulative deliveries since starting in late February are 1,168 units. The figures come from Faraday Future's November 1 weekly investor update on its website, the first since Jia Yueting (Faraday Future's founder) became FFR's CEO; Jia said October had a positive gross margin, that November and December would each average over 400 units, and that the full year would reach about 2,000 units. The basis is combined "sales and shipments" of robot equipment, with no breakdown by humanoid, biomimetic or other categories and no unit price given. Of the four industry segments, education is in the lead, while security patrol is still at proof of concept. Source: Securities.io source

LEAPTING has supplied over 50,000 cleaning robots cumulatively to Indian solar plantsindustrial ⚠️ Vendor-reported

This corresponds to 10 GW of installed capacity, with customers being solar developers and EPC contractors in India. Many parts of India are short of water, and these robots dry-clean panels with brushes and connect to SCADA for remote scheduling. The company has set up an operations center in Agra with more than 100 technical support staff. Source: MVA Pulse source

EACON's driverless mining trucks run at copper mines above 5,400 meters altitudeautonomy

EACON (Chinese autonomous mining-truck company) has extended its deployments at the Julong and Zhibula copper mines in China to above 5,400 meters, higher than Quellaveco in Peru (about 3,500 meters) and Collahuasi in Chile (about 4,400 meters). Research cited by the company shows that among workers tested at Andean mines at 3,800 to 4,500 meters, 73% had nighttime blood oxygen below 85%; driverless operation moves drivers out of high-altitude cabs and into centralized monitoring posts. Source: International Mining source

AgiBot's 6 Genie G2 robots work a 24-hour shift in a convenience storehumanoid

AgiBot previewed this livestream when it released the embodied Agent framework Genie Master in late October; at 10:00 on November 2, 6 Genie G2 units started work as "probationary employees" in an operating convenience store, handling greeting, making coffee, hot-food service, picking and restocking, and cleaning tables, built on the GO-3 foundation model. The whole thing was livestreamed for 24 continuous hours. This is a public capability demonstration, and AgiBot has not provided success rates, the number of human interventions or plans for routine deployment. Source: CNMO source

Agility says over 30 prospective customers are in talks or trials for Digithumanoid ⚠️ Vendor-reported

Three more manufacturing customers are paying for application evaluations, and Agility did not name them, describing them only as a global automaker with annual sales above 3 million vehicles, a leading auto parts and technology supplier, and a global consumer goods manufacturer. The pipeline also includes two potential collaborations as a subcontractor for a branch of the US military, which are early-stage and not yet signed. Digit 5's disclosed multi-year committed orders exceed US$300 million, and the company itself notes this is potential value that will be realized over time. Source: Robotics & Automation News source

South Korea's IL puts humanoids on an affiliate's injection molding line for validationhumanoid

The test line is an operating injection molding line at the group's automotive electronics plant IL Mobility, and IL will show the process from process diagnosis and data collection to on-site validation at RoboWorld 2026 from November 4 to 7. This is still a proof of concept, and the number of units and cycle time have not been disclosed. Source: BigGo Finance source

Mount Tai climbing competition: quadrupeds reach the summit all the way, humanoids stumblehumanoid

The Unitree A2 from the Beijing Lingyi Technology team ran the 3.9-kilometer, 3,100-plus-step course in 43 minutes 41 seconds, including a 2-minute penalty, winning the legged speed event at the first CMG World Robot Mount Tai Climbing Competition on November 2. The Shanghai Gaoyi Technology team's A2 and B2 took second in the speed and load-carrying events respectively, with the B2 carrying an effective load of 66 kilograms. The humanoid group's official stage only went as far as Doumu Palace, about 1.6 kilometers and 600-plus steps. The Yongda Zhicheng team entered a Unitree G1 developer edition that overheated midway, and the operator withdrew for safety, with the robot carried down the mountain on a stretcher; Unitree itself did not send a humanoid to compete. The Zhishen Lingyue team's E04 was first across the finish line at Hutian Pavilion, but on the last flight of steps it was "one step forward, three steps back," and another robot finished while being led by a rope held by its operator. Source: Guandian source; Jimu News source; MyDrivers source

London's mayor calls for a cap on the number of Robotaxisautonomy

Sadiq Khan is lobbying the central government to give Transport for London (TfL) more powers. The Guardian reports that under current rules for fully driverless vehicles, TfL can only approve or refuse the use of vehicles in London and cannot manage or license Robotaxi services. Khan said: "London is not San Francisco. We are a city of nearly 10 million people, 400 villages and narrow roads." The issues he raised were air quality, congestion, and the livelihoods of ride-hail and black cab drivers. Freenow by Lyft hopes to launch fully driverless Baidu Robotaxis in London as early as next year; Uber has offered limited passenger service with safety drivers using Wayve vehicles since September, and driver unions have protested outside Wayve's and Waymo's London headquarters in recent weeks. Source: The Guardian source

Samsung SDS breaks ground on Gumi AI data center, starting Samsung's 60 trillion won Yeongnam planadjacent

Initial power capacity is 60MW, expandable to 120MW, with completion at the end of 2028 and liquid cooling for the high-density GPU zone. Samsung SDS announced an investment of 427.3 billion won for it in January. The total of Samsung's Yeongnam investment plan announced in July is 60 trillion won (about US$44.6 billion), with Gumi alone covering humanoid robots and AI data centers; this data center will provide training compute for the group's manufacturing robot retrofits. Source: The Korea Times source; Maeil Business source

Japan's economic and fiscal policy minister takes a ride in a Wayve car, then commits to doing more on autonomous drivingautonomy

On November 2, Minoru Jonai, accompanied by Wayve CEO Alex Kendall, took a roughly 15-minute test ride around the National Diet in Nagatacho, Tokyo. Jonai said the government "must support private companies with a speed that is no less than theirs." Autonomous driving is among the 62 priority items in Japan's growth strategy, with a combined public-private investment target of 8.2 trillion yen through fiscal 2040 and a goal of capturing 25% of global autonomous vehicle sales in the 2030s. Source: Guandian source

MaxInsights founder: annualized revenue from about US$1 million to about US$250 million in a yearadjacent ⚠️ Single-source claim

MaxInsights turns human first-person video into robot training data, and just raised a US$30 million seed round in October. Founder Yuan Yuguang said in an interview with Silicon Valley 101 that cumulative data is close to 2 million hours, peak monthly capacity is 450,000 hours, and a 10-person team coordinates about 4,000 collectors worldwide; the first order was only 200 hours, from a US large-model company. Yuan said Dyna Robotics found its own scaling law using about 1 million hours of human first-person data, most of it bought from MaxInsights. Source: BigGo Finance source

Ma'anshan real-scene training challenge puts its trial rounds in factories and supermarketsadjacent

The challenge is hosted by the Ma'anshan municipal government, the Chinese Institute of Electronics and the World Robot Cooperation Organization, with 13 events across 5 scenarios. Trial rounds from November 16 to December 3 take place in local manufacturers, logistics companies, hotels, supermarkets and communities, with the finals from December 4 to 6 at the Ma'anshan Sports and Convention Center; events include millimeter-precision insertion, linen folding and supermarket restocking, and registration closes on November 27. Liang Liang, deputy secretary-general of the Chinese Institute of Electronics, said the industry's focus has shifted from "jumping on stage and running on the field" to "working in factories and being used in supermarkets." Source: Sina Tech source; NetEase source

Follow-up on Meta's humanoid prototype Max: Fourier non-committal, Meta hiring robot hardware roles in Shanghai and Shenzhenhumanoid ⚠️ Rumor-based

After foreign media reported that Meta is working with Fourier (Chinese humanoid robot maker) on the humanoid prototype Max, Fourier told The Paper (Chinese news outlet) on November 2 only that "if there is any new news, we will announce it through official channels." The Paper found that Meta posted a robot supplier development engineer role in Shanghai in September and a robotics studio product design engineer role in Shenzhen in October, the latter with responsibilities spanning hardware from early design, validation and system integration to mass production. Source: The Paper source

Aitu Technology's sewing robot "Aini" priced at CNY 308,000embodied

CNY 308,000 is the China launch price set for "Aini" by Aitu Technology, a unit of Jack Technology (Chinese sewing-machine maker). It targets real workstations in garment factories, doing single-ply fabric picking and wrinkle correction, and after teaching by workers it can switch between operations such as attaching pockets, buttonholes and sewing on buttons; the complete production line "Production Line King," launched at the same time, integrates driverless cutting, AI delivery, multiple Aini sewing units and a quality inspection module. The company has not disclosed customers or shipment volumes. Source: China Business Journal source

Krafton releases robot AI Ludi 0.2world-model

Ludo Robotics, the robotics research subsidiary of game company Krafton, released it at the Bay Area Robotics Symposium on October 30, US local time, adding the ability to read gestures and conversational context beyond the 0.1 version from August: if told midway through fetching a drink to switch to a different one, it stops and picks the other. The next version, Ludi 1.0, aims to unify cognition, dialogue and body control in a single neural network. Ludi 0.2 is still a research-stage system with no commercialization timeline. Source: Aju Press source