Key Takeaways

· Nanjing Wujing Dynamics (Chinese quadruped robot startup), founded 8 months ago and once said to hold nearly 1 billion yuan in orders, has halted work; employees say unpaid wages total nearly 8 million yuan

· IDC estimates the FCC ban will leave US humanoid robot revenue 58% below baseline by 2030; the same day, Intel received a one-year waiver

· China's Ministry of Commerce: sales of embodied-intelligence robots on key platforms grew 22.6% in the first three quarters

· Polish apparel group LPP's new e-commerce warehouse in Tczew will deploy more than 1,500 shelf-climbing robots from Hai Robotics (Chinese warehouse robot maker)

· Rivian buys 3D scene-generation company Third Dimension AI; the team joins autonomous-driving R&D

· On tasks demonstrated only in human videos, FRED gets robots to a 50.0% success rate; the previous best world action model reached 19.4%

FRED: Self-distillation improves future-video representations, letting a world action model learn new tasks from human videosworld-model

A compute-saving approach has become popular for world action models (WAMs, policies that generate actions using the spatiotemporal priors of pretrained video diffusion models): run just one denoising step and pass the intermediate-layer "future frame" representation to an action decoder. Action quality then depends directly on the quality of that one-step representation. FRED self-distills toward representations of real future frames, and feeds historical observations into the video model to provide temporal context. On real robots it leads other WAMs on long-horizon tasks and zero-shot new-environment tasks. On tasks absent from the robot training set and demonstrated only in human videos, FRED reaches a success rate of 50.0%, versus 19.4% for the strongest competitor (HF 15↑).

Huiwon Jang et al. · arXiv 2610.30320 source

WholeBodyMem: Humanoids remember what they saw before acting, while walking and workinglocomotion

When a robot turns around or an object gets occluded, the cues vanish from view, yet existing whole-body VLAs look only at the current frame. WholeBodyMem tracks task objects with SAM3, selects keyframes based on object events, stores them in a budget-capped memory, and lets the action expert read it through cross-attention. The authors also release HuMeGo, a memory benchmark with 6 simulated tasks and 6 real-robot tasks on the Unitree (Chinese robot maker) G1. Real-robot success rises from 36.7% to 88.3%.

Chenyuan Chen et al. · arXiv 2610.31166 source

A single integration step is enough for flow-matching VLAsvla

Flow matching is the mainstream way VLAs generate actions. Standard inference solves an ODE over many steps, which lowers control frequency, and a number of works have added dedicated single-step generation objectives. Testing on six flow-matching VLAs, this paper finds that as long as the model is well trained and has enough capacity, one-step Euler integration is essentially on par with multi-step sampling, and multi-step sampling has no consistent advantage. So large VLAs can be deployed at low latency without retraining, distillation, or auxiliary objectives.

Shivakanth Sujit et al. · arXiv 2610.30735 source

DRILL: Training world models on policy returnworld-model

When a world model is used for reinforcement-learning post-training of a VLA, lower next-frame pixel MSE can actually yield worse policy gradients: the authors have a world model with 17% higher pixel error that produced a better downstream policy. DRILL is a bilevel framework: the inner loop runs one GRPO update on imagined trajectories, and the outer loop adjusts the world model to maximize the return of the updated policy. Across five manipulation simulators and two VLA backbones, the full meta-gradient version, DRILL-IMG, beats WMPO by 13.1 percentage points in average success rate.

Zhaolu Kang et al. · arXiv 2610.30814 source

The reasoning a driving VLA gives has little to do with the trajectory it actually drivesautonomy

Driving VLAs first describe the scene and reason in language, then output a trajectory; interpretability has long been their selling point. The authors ran counterfactual interventions on four driving VLAs. Tampering with the model's own written reasoning changes the trajectory far less than masking the camera input does; using photorealistic editing to remove the key objects named in the reasoning does not make the reasoning and trajectory change consistently either. The authors conclude the two are only weakly coupled, so the interpretability such systems can offer is limited.

Gabriel Manalu et al. · arXiv 2610.30953 source

Progress Probes: Monitoring execution with the VLA's own activationsvla

Most existing runtime monitors bolt on a VLM to watch the policy, at high compute cost and system complexity. This paper trains linear probes directly on the internal activations of the open-source π0.5, learning to estimate task progress from demonstration data and raising an alarm when execution deviates from expected progress. On simulated and real-robot manipulation tasks, it detects failures as well as or better than VLM monitors, and also beats methods that need failure data for training.

Atiksh Bhardwaj et al. · arXiv 2610.31181 source

Without seeing any robot data, pretraining on procedurally synthesized control problemsvla

The authors built three data pipelines that abstract robot control into mathematical problems, generated in bulk, to pretrain a Transformer for action prediction. The model has never seen real robot behavior, yet through in-context learning it makes non-trivial zero-shot predictions on unseen real-robot tasks. For comparison, a baseline trained on real-robot data with the same compute performs better, but only on the datasets it was trained on.

Mateusz Odrowaz-Sypniewski et al. · arXiv 2610.31275 source

OpenTac: Enter a description and get manufacturing files for a tactile skinperception

This resistive tactile skin adds a structured deformable overlay to its surface that converts tangential force into spatially uneven pressure: the differential signal between adjacent contacts resolves shear force, while the common-mode signal corresponds to normal force. A companion CAD tool automatically generates manufacturing files from 2D outlines, developable surfaces, or even text descriptions (such as "design a sensor for a size-9 insole"); the team has built more than 30 sensors this way and validated feedback control and policy learning on four real robots. For a 1,024-contact skin sampling only 16 points per readout, adaptive readout reaches kilohertz rates, 19.2x faster than row-by-row scanning of the full array. All hardware and software are open source.

Ariel Slepyan et al. · arXiv 2610.31661 source

GelSlim 5.0: One fingertip camera that images the object and measures contactperception

Once a hand touches an object, external cameras are often blocked. GelSlim 5.0 uses a UV-sensitive contact membrane and switches the lighting so that a single camera alternately captures a see-through visual image and a tactile image, eliminating extrinsic calibration between two cameras. The companion multimodal transformer, VisTacFusion, predicts the object's planar pose and contact geometry along two branches. On 12 known objects, vision plus touch cuts depth MSE by 61.2% compared with touch alone.

Xiangdong Wang et al. · arXiv 2610.31334 source

Rivian acquires Third Dimension AI (UK) ｜ M&A ｜ Amount undisclosedworld-model

Third Dimension AI was founded in 2024 by Tolga Kart and Piotr Sokolski. Its platform converts real road data into controllable, high-fidelity 3D environments, replicating complex driving scenarios in bulk and altering the conditions within them, so rare edge cases need not wait to occur in the real world. Rivian's senior vice president of autonomy and AI, James Philbin, said: "Simulation is central to how we train and validate autonomous driving systems." The team joins Rivian's autonomous-driving R&D, and Rivian says the technology will be used on both its passenger-car and robotaxi tracks. Sources: Rivian announcement source; TechCrunch source

Tractian (US) ｜ Series D ｜ $130 million ｜ Valuation $1.4 billionindustrial

Led by Sapphire Ventures, with Catalyst, part of M&G, as co-lead, and participation from N47, General Catalyst, Bosch Ventures, Y Combinator, and others. The Atlanta company collects industrial equipment data using vibration, ultrasonic, and other sensors, and an agent platform then diagnoses root causes of faults and recommends actions; customers include Eli Lilly, McKesson, and Whirlpool. Its "physical AI" refers to equipment health monitoring and does not involve robot hardware. Source: PR Newswire source

Formic (US) ｜ Series B + debt financing ｜ $50 million combinedindustrial

The $50 million is growth capital: the equity portion is a Series B led by Mercato Partners, with Light Ray Ventures, Lightscape Partners, Citi, and others participating; the debt comes from S2G Investments. The Chicago company's automation systems, including palletizing, run at more than 100 facilities in over 30 US states. Source: Business Wire source

Limiaoda (Suzhou) ｜ Series B1 + B2 ｜ Nearly 500 million yuan cumulativeindustrial ⚠️ Single-source claim

Limiaoda provides "lights-out factory" solutions for semiconductors and optical modules in Class 100 cleanroom environments; the company says its Series B3 round, worth several hundred million yuan, has been oversubscribed. By its account, orders in hand and intended for 2026 total nearly 1 billion yuan, of which cumulative orders from core customer Micron reach several hundred million yuan. Funds go to a semiconductor and optical-module equipment manufacturing base and an AI4S lab, as well as expanding the embodied-algorithm team. Source: Embodied Home (具身之家) source

Xingyi Intelligence ｜ Seed + Seed+ ｜ Tens of millions of yuanadjacent

Incubated out of Tsinghua University's Department of Computer Science, the company specializes in first-person-view hand-object interaction training data. The seed round was led by Shunxi Fund; the Seed+ round was led by Luster LightTech (Chinese machine-vision company), with Zhongguancun Capital and Yichuang Investment participating. Luster also works with it commercially through its subsidiary Yuanke Shijie, providing optical motion-capture and multi-view acquisition hardware. Source: Embodied Home (具身之家) source

Fujian Humanoid Robot Innovation Center ｜ Pre-A+ ｜ Amount undisclosedhumanoid

Announced on October 23; the investors are Fuzhou Gulou District Urban Construction Investment Development Group and Xiamen Hangzhi Liangshu Future Technology Incubator. Source: Guandian (观点网) source

Keda Intelligent ｜ Hong Kong IPO ｜ Filed againindustrial

After a first filing on January 19, this A-share company filed again with the HKEX Main Board on October 22, with Huatai International and Guoyuan International as joint sponsors. Its business has two main segments, power equipment and energy storage, and robotics; citing Frost & Sullivan data, the prospectus says its cumulative industrial robotic arm revenue from 2020 to 2025 ranks first in China. 2025 revenue was about 3.015 billion yuan with profit of 85.665 million yuan; first-half 2026 revenue was 1.524 billion yuan with profit of 111 million yuan. Source: Tonghuashun Finance (同花顺财经) source

LPP to deploy more than 1,500 Hai Robotics climbing robots at new Tczew warehouse in Polandindustrial

The partner this time is the logistics subsidiary of leading Polish apparel group LPP. The two sides plan to build a new e-commerce warehouse in Tczew, northern Poland, deploying more than 1,500 shelf-climbing warehouse robots to serve apparel e-commerce orders across Central and Eastern Europe. Hai Robotics says this is the largest project of its kind for its solution to date. In September, Hai Robotics disclosed a project for an unnamed European fashion retailer of the same scale, more than 1,500 units. Two weeks ago, Geek+ (Chinese warehouse robot maker) announced the deployment of more than 1,600 robots across two LPP fulfillment centers in Poland, meaning LPP now uses systems from two Chinese warehouse robot makers at once. No delivery date has been announced. Source: EE Times China (电子工程专辑) source

Tesla publishes Robotaxi safety report: 3.11 million unsupervised miles, with four cities listed for the next batchautonomy ⚠️ Vendor claim

From January 22 to October 18, Tesla's Robotaxi logged 3,112,541 miles with no safety monitor in the vehicle, and Tesla says there were zero injuries and zero airbag deployments; citing state police data, it estimates human drivers would have had about 9.1 injury-causing collisions over the same mileage (vendor claim). The two sides are not measured the same way: Tesla counts only collisions where the Robotaxi made first contact, the human benchmark does not distinguish who hit first, and the mileage also includes trips where a remote operator assisted or had taken over. The report lists Las Vegas, Jacksonville, Phoenix, and San Antonio as "coming soon," with no dates. Sources: Tesla Oracle source; Not a Tesla App; Drive Tesla Canada

Volvo: No timeline for removing safety drivers on Waabi's Texas freight laneautonomy

This Dallas-to-Houston lane began hauling freight for the Warp freight network in early October. Ceren Wende, a spokesperson for Volvo Autonomous Solutions, confirmed to Trucking Dive that there is no timeline for when safety drivers will be removed from trucks on the lane. In June, Volvo announced fully driverless operation on US highways in the first quarter of 2027; Wende said that plan will launch first with another technology partner, Aurora, and does not include the Warp lane. Wende said: "We will move forward only after full consideration and when we think it's appropriate."Source: Trucking Dive source

Advantech's "AI Factory Brain" is already running on its own 20 production linesindustrial ⚠️ Vendor claim

At GTC Berlin, Advantech (Taiwan-based industrial computing company) disclosed that the system is in use on its own 20 production lines, with per-capita output up 13.2% quarter on quarter in the third quarter; the company says 8.91% of that came from 5 AI agents in actual operation. The system connects to Nvidia's Factory Operations Blueprint (FOX) and real-time data from Cosmos and Nemotron; after a goal such as "raise output by 10%" is set, it analyzes bottlenecks on its own and coordinates the agents to propose improvement plans. Source: BigGo Finance source

Nanjing Wujing Dynamics halts work and owes wages; CEO says they cannot reach the controlling shareholderembodied

An employee told National Business Daily that Wujing Dynamics has owed wages since August, stopped paying into the housing provident fund in September and stopped social insurance contributions in October, affecting thirty to forty people and totaling nearly 8 million yuan; the company has halted work, and some employees are pursuing labor arbitration. On October 22, Yuan Haihui, publicly identified as CEO, told reporters that they currently cannot reach controlling shareholder Cui Xiaofei and are unable to comment; Cui's phone was continuously busy. The company was founded in February 2026 and makes industrial quadruped robots for power-line inspection, petrochemical, and energy and mining uses. In June, Nanjing local reports said it had signed framework orders of nearly 500 million yuan with several central state-owned enterprises, that local market backlog orders were also close to 500 million yuan, and that 110 million yuan in financing had been secured. In a June announcement, Junchuang Technology, a Beijing Stock Exchange-listed company, said it would set up a 32.01 million yuan fund with partners to invest in Wujing Dynamics: first into Chaoguangwei, whose legal representative is Cui Xiaofei, then swapping shares to hold Wujing Dynamics directly. After the August 24 business registration change, the fund does not appear on Wujing Dynamics' list of direct shareholders; Junchuang Technology had not responded as of publication. Of the so-called nearly 1 billion yuan in orders, half are framework contracts and half are "backlog." Source: National Business Daily (每日经济新闻) source

IDC: FCC ban to leave US humanoid robot revenue 58% below baseline by 2030humanoid

IDC's trade policy model shows that by 2030, US humanoid robot revenue will be 58% lower than the pre-restriction baseline, with cumulative lost revenue of more than $6 billion, and annual growth falling from 13% to about 9%. Household cleaning robots shrink by only 18%: they have a large stock of previously certified older models to sell, whereas humanoids turn over models quickly and depend more on certifying new ones. IDC estimates global humanoid shipments in the first half of this year at about 25,000 units, 26% more than the 18,500 for all of 2025, with Chinese manufacturers accounting for about 95%. On October 23, the FCC granted Intel a one-year waiver. In its application, Intel said that AGVs that have run in its fabs for more than twenty years would disrupt production if software, firmware, and hardware could not be updated, and that there is no proven US-made alternative, with replacement taking several years. Sources: Fortune source; The Economic Times source

Ministry of Commerce: Embodied-intelligence robot sales on key platforms up 22.6% in first three quartersadjacent

An official of the Consumption Promotion Department of China's Ministry of Commerce, citing commerce big data on October 23, said sales of embodied-intelligence robots on key platforms grew 22.6% in the first three quarters, exoskeleton assist devices grew 46.4%, and smart glasses grew 1.1 times. These are sales figures from the Ministry of Commerce's monitored platforms, which are not the same measure as production figures such as the Ministry of Industry and Information Technology's "more than 40,000 humanoids produced in the first half." Source: China News Service (中新网) source

Tesla cuts 2026 capex guidance from more than $25 billion to about $20 billionautonomy

Third-quarter capex was $4.7 billion, double the year-earlier level but below the analyst average of $7.5 billion compiled by LSEG. Free cash flow was $1.53 billion, versus analysts' expectation of a $4.92 billion outflow. Tesla did not say which spending would be cut. Ryan Lee, senior vice president of product and strategy at Direxion, said: "Capex coming in below expectations suggests it isn't going all in on the AI arms race the way Musk was expected to." On Tuesday this week, Musk called a Tesla–SpaceX merger a "good idea," and the two companies have set up special committees to discuss it. Source: Global News source

NEURA unveils 4NE1 Gen 4, opens 17 million euro training center with TU Munichhumanoid

The new-generation 4NE1 is 20% lighter than the previous generation, with a narrower body and finer hands, and its limbs are fitted with components NEURA calls Smart Limbs, which let an individual limb respond to environmental changes on the spot. It runs the company's vision-language-action model Neuron AI; NEURA says it is preparing for mass production but gave no timing or volume. TUM RoboGym is located at LabCampus next to Munich Airport, covers 2,300 square meters, runs for an initial three years, and has a total investment of about 17 million euros, of which NEURA contributes about 11 million. Training began in September; companies from more than 10 industries, including automotive, pharmaceuticals, medical devices, logistics, agriculture, and care, are training robots there to move cartons, fill bottles, and load and unload timber components. TUM President Thomas Hofmann said that working with a German robot maker is "also a clear signal for Germany as an industrial base in a time of geopolitical uncertainty." Source: AI Insider source

Wayve, Stellantis, Nvidia, and the French government announce collaboration to advance autonomous driving deploymentautonomy ⚠️ Plan-stage claim

The collaboration was announced at the Assises du Véhicule Autonome in Paris on October 23. The division of labor: Stellantis provides the L4-Ready platform and vehicle engineering and manufacturing; the Wayve AI Driver runs on Nvidia's Halos OS, built on the Hyperion compute and sensing architecture; engagement with unions and city authorities is led by the local partner responsible for operations. French Transport Minister Philippe Tabarot said France will adapt its regulatory framework to speed up autonomous driving, "while maintaining strict safety and data sovereignty requirements." The announcement does not state vehicle models, cities, scale, or timeline. Source: Stellantis Media source

inDrive teams up with Pony.ai and Nvidia, plans robotaxi tests in Kazakhstan in 2027autonomy ⚠️ Plan-stage claim

The three parties signed a memorandum of understanding. inDrive CTO Yuri Misnik told TechCrunch that testing will begin in the capital, Astana, in the first half of 2027, initially with safety operators. Pony.ai provides its seventh-generation Robotaxi, which inDrive will operate on its own app; Nvidia provides the Hyperion platform and the Halos safety system. "Our approach has always been asset-light," Misnik said. Nvidia automotive vice president Ali Kani said ride-hailing in emerging markets is already cheap, so driverless services must compete on cost. Sources: Tech in Asia source; TechCrunch

In a Bremen warehouse, a 24-year-old Chinese international student is crushed to death while maintaining automated equipmentindustrial

Bremen police report that the accident happened around 11 p.m. on October 13, when the victim was trapped while maintaining a lifting robot in a high-bay warehouse in the port area, suffering severe neck injuries; coworkers gave aid on the spot but could not save them. The victim studied at the University of Bremen; according to reports, the warehouse owner is ethnically Chinese and employed several Chinese students for warehouse robot maintenance and shelf inspection. Police are still investigating. Source: Sina News (新浪新闻) source