Key Takeaways

· At the Zhibula copper mine, 5,400 meters above sea level, all 60 battery-electric mining trucks running EACON-style autonomy from Yikong Zhijia (Chinese autonomous mining-truck firm) had moved to driverless operation by the end of September

· Kodiak has delivered its first Western Star 49X driverless trucks, bringing Atlas's self-operated driverless fleet to 37 trucks

· The Netherlands' RDW approved a Tesla update: the driver-assistance speeding margin drops from 50% to 10%, and FSD is renamed

· AIRoA-0 used no robot data at all and reached an average success rate of 78% across ten real-robot tasks; the strongest comparison VLA reached only 27%

· CNPC Kunlun Manufacturing and Unitree (Chinese humanoid robot maker), among others, plan an industrial embodied-AI joint venture, with the CNPC side holding 60%

· Boston Dynamics redesigned the second-generation Stretch using 100,000 hours of deployment data, cutting internal cabling by 81%

· Ouster shipped more than 27,000 sensors in the third quarter, with revenue of $59 million, up 50% year over year

AIRoA-0: Learning long-horizon bimanual manipulation from handheld-gripper data alonemanipulation

The training data is 16,151 hours of handheld demonstrations. No robot data was collected at any point, and there was no teleoperation, reinforcement learning, or on-robot correction; the robot appears only at test time. The demonstrations were recorded with YUBI, the team's open-source handheld gripper. The test site is not among any of the training-data collection sites. Using only two wrist cameras, the robot reached an average success rate of 78% across ten tasks, including sorting small parts and changing the mop pad on a robot vacuum; the strongest of four VLAs post-trained on the same task demonstrations reached only 27%. The dataset, pretrained weights, and post-training recipe are all released, and the authors say post-training on a new task takes under four hours on a single 8-GPU H200 node.

Yueh-Hua Wu et al. (AIRoA) · arXiv 2611.03815 source

Rationale text unchanged, actions altered: a monitoring blind spot in reasoning VLAsvla

Reasoning VLAs output a written rationale as they act, and supervisors judge safety by reading these rationales. In Alpamayo (driving) and InstructVLA (manipulation), the authors found an action-sensitive direction nearly orthogonal to the text gradient and applied visual perturbations along it: across 154 driving scenes, 52.6% showed new collisions while the rationale text stayed identical word for word; an unguided action attack achieved only 4.55%. The conclusion is that monitoring must verify the action channel directly.

Zihao Feng et al. (including Z. Morley Mao) · arXiv 2611.04166 source

RAFT: Learn reasoning before actions, so few-shot post-training no longer loses robustnessvla

Fine-tuning a generalist VLA on a few demonstrations raises nominal success rates but degrades robustness to changes in appearance, objects, and instructions. RAFT first trains a low-rank adapter on task-relevant reasoning annotations, then freezes it and fine-tunes on action demonstrations. On a real-robot DROID setup, out-of-distribution success is 75.0%, versus 27.9% for ordinary fine-tuning.

Shresth Grover, Andrey Kolobov, Dhruv Shah · arXiv 2611.04368 source

GSBR: Leaving online RL room to surpass the imitation policymanipulation

From Sergey Levine's group. Online RL is regularized toward a pretrained diffusion or flow policy, and if the constraint is too tight it cannot learn better actions. GSBR applies a penalty based on how far an action lies from the base policy's support: no penalty inside the support, and a penalty that grows smoothly with distance outside it, estimable by sampling from the base policy alone. Across 19 tasks on four benchmarks, average success rises from the base policy's 33% to 95%; the code is open-sourced.

Charles Xu et al. (including Sergey Levine) · arXiv 2611.05295 source

PyRUA-DataFarm: Letting a coding agent collect data in place of teleoperatorsbenchmark

Given a real-robot calibration clip of the target scene, the agent rebuilds the task in simulation, writes the scene, the expert controller, and the success criterion itself, and a video model then renders each rollout as photorealistic multi-view video. On five real-robot tasks, a policy trained only on farm data succeeded in 93 of 100 trials, versus 84 for a policy trained on the same amount of teleoperation data; with an unfamiliar tablecloth and colored lighting, the two scored 87 and 46. Counting all costs, one hour of three-view data averages $107, slightly below the $117 teleoperation quote in China.

Kewei Zhang et al. · arXiv 2611.05320 source

RoboTryOn: "Trying on" human videos as robot demonstrationsmanipulation

HF 4↑. The idea borrows from virtual try-on: keep the demonstrated motion, objects, and background, replace the person in the frame with the robot specified by a reference image, then infer action labels with an inverse dynamics model that uses depth features. Across several humanoid embodiments, average success on new tasks with no real-robot demonstrations rises from 2% to 60%. Matching the performance of training on teleoperation data alone costs roughly one third as much to collect.

Seungyong Lee et al. (including Jinwoo Shin) · arXiv 2611.05163 source

DreamX: A video diffusion model as a general latent-action learnerworld-model

HF 46↑, the highest community buzz among today's papers. The authors turn a pretrained video diffusion model into a latent-action encoder, compressing each latent frame from 1,200 tokens to 4, a 300x reduction. On LARY-Bench it is state of the art among latent-action models; in one setting, the encoded actions transfer zero-shot to other embodiments and environments, with humans and robots sharing a single action interface.

Jisoo Kim et al. · arXiv 2611.05377 source

LoopQ: VLA-specific quantization methods lose to naive baselinesbenchmark

Holding checkpoints, quantized layers, and the inference engine fixed and changing only the quantization method and kernels, the authors ran more than 180,000 episodes of LIBERO and LIBERO-plus on four VLAs, with hardware spanning H100, L4, and Jetson AGX Thor. At 4 bits, the naive RTN, AWQ, and GPTQ match or beat the VLA-specific ActQuant and QuantVLA. Quantization reliably saves 2 to 3x memory, but speed has not kept up: on an H100 at batch size 1, no quantized π0.5 or OpenVLA-OFT configuration is faster than the 16-bit version, and on Thor every 4-bit configuration is slower.

Syed Umer Bukhari et al. · arXiv 2611.05093 source

ReflexArc: Adding a 10-millisecond tactile reflex to humanoid whole-body controllocomotion

The first force peak of an impact arrives within tens of milliseconds, too fast for a 50 Hz whole-body policy to react. ARC uses whole-body touch sensing to recognize hazardous contact and, within 10 ms, sends the computed withdrawal posture straight to the motors, then folds the reflex into policy training to maintain balance and complete the follow-up recovery. On the Unitree G1, compared with the same reflex at 20 ms latency, the first force peak falls by 34.1% and contact impulse drops from 0.2632 to 0.1225 N·s.

Mingyu Wang et al. (including He Wang, Li Yi) · arXiv 2611.04424 source

Houmo AI ｜ Series B ｜ several billion RMB ｜ valuation undisclosedhardware

Both Shanghai Observer and Zhidx reported this Series B of several billion RMB. Houmo AI (Chinese edge AI chip startup) makes AI chips for on-device large models; Shanghai Observer says the M50 has been adopted by more than a hundred customers, and by the specifications announced at WAIC in July, the M50 delivers 160 TOPS per chip at 10W. Source: Shanghai Observer source; Zhidx source; Wallstreetcn

Corridian (US) ｜ Seed ｜ $38 million ｜ valuation undisclosedautonomy

Led by Point72 Ventures, with participation from Balerion Space Ventures, Harpoon Ventures, DCVC, Marlinspike, and others. The company was founded in September 2025 by four former General Dynamics executives and builds large autonomous underwater vehicles that carry military cargo in standard shipping containers over voyages of thousands of nautical miles; its first flagship vessel is due in 2027. Retired Admiral Lisa Franchetti, former US Chief of Naval Operations, is senior military adviser. CEO Mike Guay said: "The US relies on surface and air to move supplies at scale, and that is a serious weakness." Source: AI Insider source

Syncere (US) ｜ Pre-seed ｜ $11 million ｜ valuation undisclosedembodied

Led by a16z Speedrun, with participation from Draper Associates, Physical World, Sequel, SeaX Ventures, and others. The product, Lume, is a floor-lamp-shaped home robot starting at $2,699; it can move around the home and, the company says, fold clothes, tidy rooms, and make beds. The company says pre-orders have reached several million dollars and that the product is in production and has begun shipping. CEO Aaron Tan said home robots need real data to become reliable, yet to get into homes they must first be reliable enough; "we built Lume to break that cycle." Source: AI Insider source

Peppermint Robotics (India) ｜ Pre-Series B ｜ INR 1 billion ｜ valuation undisclosedindustrial

The round includes both new shares and secondary transactions. The company started with floor-cleaning robots and is expanding into material handling and cleanroom automation for factories, warehouses, and pharmaceutical plants; it plans to launch a new model within 12 to 15 months and an outdoor cleaning robot in 2028. Source: Inc42 source

Saigan Technology (Shenzhen) ｜ Pre-A++ ｜ tens of millions of RMB ｜ valuation undisclosedhardware

Led by Donghu Hi-Tech Investment. Founded in 2023, the company makes flexible smart sensors and electronic skin, for use in humanoid robots, consumer electronics, automotive electronics, and healthcare. Source: Eastmoney source

Deta Intelligence ｜ Galaxea AI and Lenovo Capital take stakes ｜ amount undisclosed ｜ valuation undisclosedworld-model

Qichacha (Chinese company-registry database) shows that Beijing Deta Yuanchuang Intelligent Technology has added shareholders including Shanghai Lenovo Future Venture Capital Partnership and Galaxea AI (Chinese embodied-AI startup), raising its registered capital to RMB 1.3504 million. Founded in January this year, the company builds native humanoid-robot foundation models. Source: Stockstar source

Analog Devices acquires Alif Semiconductor ｜ Closed ｜ $1.35 billion, all cashhardware

In September, ADI announced it would acquire Alif for $1.35 billion in cash, plus up to $200 million in contingent consideration; the deal has now closed. Source: TradingView source

All 60 battery-electric mining trucks at the Zhibula copper mine move to driverless operationautonomy

At the Zhibula copper mine, above 5,400 meters in altitude, 60 Sany (Chinese heavy-machinery maker) battery-electric wide-body mining trucks fitted with Yikong Zhijia's system reached mine-wide driverless operation at the end of September. The owner is Tibet Julong Copper, a subsidiary of Zijin Mining (Chinese mining giant). Yikong had earlier disclosed that its deployments extended to this altitude, and International Mining now lays out how the mine ramped up: the vehicles arrived in September 2025 and were first driven by people, with on-site validation of operation without safety drivers completed within two weeks; by March 2026, 52 trucks were in routine driverless operation. Slippery freeze-thaw road surfaces and ice-and-snow glare interfering with perception are what make this mine harder than lowland mines. The neighboring Julong copper mine is still retrofitting 18 NHL NTE240 diesel rigid-frame trucks in the 220-tonne class, some of which are already running driverlessly. When Yikong published its fleet figures in September, it recorded the 60 high-altitude vehicles under the Julong copper mine; both mines are operated by Julong Copper. Source: International Mining source

Kodiak delivers first Western Star 49X driverless trucks; Atlas fleet reaches 37autonomy

Atlas Energy Solutions' owned and operated driverless trucks rise to 37. The 2 newly delivered trucks are Western Star 49X units built by Daimler Truck North America, fitted with Kodiak's seventh-generation hardware, delivered in September and announced on November 3. They are the first two of the 65 Western Star trucks in Atlas's initial order of 100. Kodiak says it will complete the order on schedule by mid-2027. Source: Kodiak press release (GlobeNewswire); GuruFocus source

Xiaomi humanoids work 8 hours straight at a car-plant sorting stationhumanoid ⚠️ Vendor-reported figures

At the center-console side-panel sorting station in Xiaomi's car factory, two full-size Xiaomi humanoids worked continuously for 8 hours, taking panels out of delivery bins and inserting them into sorting racks; Lei Jun (Xiaomi founder) said this meets the station's full-day production needs. This is a single-station production-line validation, not yet deployment at scale; Xiaomi did not disclose cycle time, the number of human interventions, or how batteries are managed, and the video was sped up by as much as 100x. Kuai Keji says Xiaomi's robots previously could work for only 3 hours at a stretch. Source: CLS source; Kuai Keji source; Interesting Engineering source; Gasgoo source

HAI Robotics says Shanpan has signed more than 20,000 units worldwideindustrial ⚠️ Vendor-reported figures

At the end of August, HAI Robotics (Chinese warehouse-robot maker) put the figure at more than 10,000 units; at CeMAT it updated this to more than 20,000 units signed, with the largest single project exceeding 5,000 units. Units signed are not the same as units in operation. The newly launched double-deep solution can store 45,000 totes in 1,000 square meters, and the HaiQ platform can dispatch up to 6,000 robots on one site. Source: Ifeng Tech source

Epiroc autonomous drill rigs enter Fortescue, KGHM, and other minesindustrial

The Autonomous SmartROC D65 MKII is already operating at customer sites in Australia and the US, including Fortescue's Solomon and Eliwana, Luck Stone's Rockville, and KGHM's Robinson mine; Epiroc has not disclosed fleet numbers. Source: International Mining source

More than twenty Cybercabs on the road for testing in Detroitautonomy

The Detroit News reports that more than twenty Cybercabs set out from Tesla's collision repair center in Southfield and are testing on roads including I-696 and M-10. The cars carry temporary steering wheels and are driven by human test drivers. Tesla has not officially announced testing in Detroit and has given no passenger-service date. Waymo removed front-seat staff locally in October, but carries only employees. Source: EVXL source

Dutch RDW approves Tesla update: speeding margin cut to 10%, FSD renamedautonomy

The speeding margin cap on Tesla's driver-assistance system falls from 50% to 10%. The Netherlands Vehicle Authority (RDW) confirmed on November 4 that it approved the software update, and "FSD Supervised" is formally renamed "Tesla Assisted Driving" at the same time. According to the RDW announcement, the configuration approved on April 10 allowed the system to exceed the recognized speed limit by up to 50% on grounds of "safety and traffic flow," meaning 150 km/h on a 100 km/h road. The new cap resets to zero each time the vehicle is started, and other speed-related features are off by default, though RDW did not say which ones. RDW says that after verification it folded these changes directly into the existing type approval, and stresses that the rename does not change the system's functions: it remains driver assistance, and responsibility stays with the driver. The number of countries recognizing this approval, including the Netherlands, is 10, up from 8 in early October. Opposition has centered on speeding: in June Sweden demanded the over-limit setting be removed or it would not allow the system; France refused to recognize the Dutch approval over speeding on urban roads. The conditions Germany's transport ministry set out for its support on October 6 were dropping the "Full Self-Driving" name and limiting the speed tolerance to 10%; CNMO says only a few weeks remain before an EU-wide vote. Source: CNMO source

CNPC Kunlun Manufacturing and Unitree, among others, plan an industrial embodied-AI joint venturehumanoid

The Beijing Municipal Administration for Market Regulation began publicly posting a new joint-venture filing on November 4, with the notice period running to November 13. The parties are China National Petroleum Corporation (CNPC) Kunlun Manufacturing and its concert party Kunlun Capital, Unitree, and Beijing Fanjie Intelligence; the joint venture will offer industrial embodied-AI robot solutions. The shareholdings are 60%, 20%, and 20%, with joint control by the three sides. Kunlun Manufacturing was registered in Xiong'an in 2023 and mainly makes energy equipment; its ultimate controller is CNPC. Fanjie Intelligence was founded this year, incubated by the Beijing Weishi Embodied Intelligence Research Institute and a team from the Chinese Academy of Sciences' Institute of Automation, and works on integrating world models with embodied foundation models; CEO Han Jinchi is a member of Academician Qiao Hong's team. When Unitree listed on the STAR Market in August, CNPC was on the strategic placement list. Source: Zhidx source; CLS source

Boston Dynamics launches second-generation Stretch, focused on less downtimeindustrial

The second-generation Stretch has 81% less internal cabling than the first generation. The design for this truck-unloading robot draws on more than 100,000 hours of deployment data since the first generation entered commercial use in 2023, over which it handled tens of millions of boxes. The changes target the causes of unplanned downtime rather than pursuing faster cycle times. Some joints now rotate continuously beyond 360 degrees, the electrical connections are replaced with automotive-grade connectors, and the vacuum gripper's valves and sensors are also upgraded. For reference, the first generation had a maximum pick of about 23 kg, a design cycle rate of 600 to 800 boxes per hour, and battery runtime of up to 16 hours. Herald Economy reports the company plans to expand supply to logistics firms such as DHL. Source: Herald Economy source

Ouster Q3 revenue of $59 million, with robotics customers driving shipmentshardware

Third-quarter revenue was $59 million, up 50% year over year and 9% quarter over quarter. Of this, product revenue was $56 million, the 15th consecutive quarter of growth. The company shipped more than 27,000 lidar and camera sensors in the quarter, with lidar about 27% of the total; demand came mainly from industrial and robotics customers in warehouse automation, defense, construction, surveying, and data collection. CEO Angus Pacala said: "Our Stereolabs camera shipments hit a record, thanks to the capacity we invested in to meet the surge in demand in robotics." GAAP net loss was $14 million, versus $22 million a year earlier; fourth-quarter revenue guidance is $61 million to $64 million. Source: AI Insider source

GGII: at least 15% of Chinese humanoid makers will go bankrupt or be acquired within three yearshumanoid ⚠️ Estimate-based figures

Seoul Economic Daily, citing a Yicai report: GGII (Chinese robotics research firm)'s general manager gave this forecast at an annual conference in Shenzhen on November 3, when China had more than 150 humanoid OEMs. GGII estimates that China's humanoid shipments this year will exceed 50,000 units, nearly triple last year's 18,000, with a market size above RMB 10 billion; the problem is that units are coming off the line faster than factories are adopting them, and most products handle only basic tasks and have "not yet entered a stable commercial stage." The same day, Xiao Yuanqi, deputy head of China's National Financial Regulatory Administration, warned of overheating in AI investment, saying excessive concentration of capital has pushed up financing costs in other industries. Source: Seoul Economic Daily source

JD.com's Cao Peng: embodied-AI spending expected at RMB 7 billion to 8 billion next yearadjacent ⚠️ Plan-based figures

Cao Peng, chairman of the JD Group Technology Committee, told Zhejiang Merchants magazine that the money will go mainly to three areas: data collection and foundational training, real-world deployment, and model R&D. At the Wuzhen Summit he disclosed that JD's embodied-AI data collection already covers more than 100 sub-scenarios and has gathered more than 100,000 real operational tasks; its two-year goal is to collect 10 million hours of real-world human video plus 1 million hours of robot-body operation data. JD says its own JoyAI-Sim simulation evaluation agrees with real-robot evaluation results about 90% of the time, at 3.2 times the evaluation efficiency of real robots. Source: Zhejiang Merchants Online source

LiberAI releases world action model Liber-0world-model ⚠️ Vendor-reported figures

Pretraining used more than 2.2 million hours of human experience data, sourced from its in-house LiberSense data glove, cross-embodiment robot data, and internet human video. The company says it ran 600 zero-shot real-robot tests on 60 tasks randomly generated by GPT-6, with an overall success rate of 36%, of which complex multi-step tasks reached only 10.5%; the same tasks handed to GPT-6 Astra for direct control succeeded 5% of the time. The company was founded in December 2025; founder Liu Songming holds a PhD from Tsinghua's computer science department, where he studied under Zhu Jun. Source: Ifeng Tech source; Robotics & Automation News source

Tesla's next-generation Optimus appears ahead of China International Import Expo; Musk denies December pre-ordershumanoid ⚠️ Plan-based figures

The China International Import Expo opens on November 5, and Tesla's next-generation humanoid has already appeared at its booth. NetEase Tech calls it Tesla's first robot aimed at mass production, with a long-term target cost of $20,000 to $30,000 per unit. On November 1, when someone on X asked whether orders could be placed in December, Musk replied with a single word, "No"; Gasgoo says the delivery window he confirmed is late 2027. Source: Sina (NetEase Tech) source; Gasgoo source

Skild AI has more than 60 paying customers, ARR above $100 millionworld-model ⚠️ Vendor-reported figures

According to AI Magazine, Skild AI's ARR passed $100 million 10 months after its first commercial deployment, with more than 60 signed paying customers, Foxconn among them. The company was valued at more than $14 billion in January. Source: AI Magazine source

Exotec survey: 94% of warehouse-automation buyers say switching vendors is too costlyindustrial ⚠️ Vendor-reported figures

The survey covered 250 warehouse-automation decision-makers in the US and Canada; a quarter regret their original choice, and 94% say the cost of switching vendors is prohibitively high. Among respondents who bought low-priced systems, 24% said they had picked the wrong vendor, versus 5% in the comparison group. Exotec, which commissioned the survey, is itself a warehouse-robot seller. Source: MarketScale source

Wang Xingxing: next year is likely the tipping point for the robotics industryhumanoid

At the China International Industry Fair in October, Unitree founder Wang Xingxing said the tipping point could arrive within a year or two at the earliest; at the Wuzhen Summit on November 4 he narrowed the timing to "most likely next year." He described fine manipulation such as disassembling phones and threading needles as "the last few millimeters problem," and also said that the utilization rate of robot data, Unitree's included, is still fairly low. Source: Kuai Keji source