Key Takeaways

· XPeng (Chinese EV maker) Robotaxi has been approved for L4 demonstration operations across all of Guangzhou, still hailed by invitation code at first

· Agility's Digit will move into two new GXO warehouses in North Carolina and Tennessee; the first site has handled more than 100,000 totes to date

· Volkswagen China obtained an L3 road-test permit in Hefei, targeting delivery of its first L3 solution in the second half of 2027

· 9 of 12 open-source VLA codebases fine-tune directly in bf16; restoring an fp32 master copy lifts π0.5's success rate from as low as 5.1% to as high as 80.9%

· Robotis signed a supply contract with Micro Duck for about 750,000 actuators, 15 per robot duck

Full-parameter VLA fine-tuning without an fp32 master copy silently erases most updatesvla

Of 12 open-source VLA codebases, 9 have a default or documented recipe that runs AdamW directly on bf16 weights. Updates smaller than 1/512 of a weight get rounded to zero, and the authors measured at least 79% of updates vanishing this way on LIBERO. Affected setups include the default SmolVLA configuration in LeRobot and the π0/π0.5 tutorials; in a sample of full-parameter SmolVLA fine-tuned models on Hugging Face, 99% store the action expert in bf16. After adding an fp32 master copy, π0.5 on LIBERO-Object rises from 89.0% to 97.7% at the default learning rate, and from 5.1% to 80.9% with the learning rate reduced tenfold. Fixing SmolVLA in LeRobot takes a one-line change.

Shubham Singh · arXiv 2610.31859 source

RynnUnify: a 16B unified embodied model spanning understanding, visual planning, and controlworld-model

HF 31↑, the highest community buzz today. The model puts spatial reasoning, physical prediction, and grounding into a single 16B model, uses interleaved video-and-language reasoning for long-horizon visual planning, and adds a separate world-action branch that outputs robot actions directly. On the new Rynn-VP benchmark it scores 9.75 points higher than Cosmos 3 Nano. The control branch reaches a 51.5% success rate on 20 RoboCasa365 composite tasks and an average of 46.7% across three real-robot tasks.

Minghao Zhu et al. · arXiv 2610.32965 source

RepresentNav: a navigation policy builds its own representation space and predicts the next moment before actingautonomy

Navigation policies usually take off-the-shelf features directly from visual foundation models. RepresentNav instead uses a three-layer transformer to build its own representation space from all layers of several visual foundation models; before each action, it predicts the next moment of the scene in that space in a JEPA fashion. Its success rate on VLNVerse fine-grained tasks is 83.80%, 20.05 percentage points above the previous best. This prediction step also cuts collisions by about a quarter and allows future depth and 3D occupancy to be decoded, even though 3D occupancy was never used as a supervision signal.

Jiawei Chen et al. · arXiv 2610.31762 source

DexPort: turning the robot's dexterous hand itself into a handheld data collectormanipulation

Collecting data for high-degree-of-freedom dexterous hands requires a full robot setup for teleoperation, while human-hand videos need retargeting. DexPort uses the same robot hand for both collection and deployment: the operator controls only wrist motion and selects or fine-tunes a few finger skills, and what gets recorded are joint commands the robot can execute directly. The authors collected more than 2,500 demonstrations. Training on commands rather than states raises success from 11% to 64%; adding data collected outside the lab raises it from 36% to 78%. Guanya Shi and Carmelo Sferrazza are co-authors.

Amber Xie et al. · arXiv 2610.32969 source

Safety filters for humanoid robots can themselves be exploitedlocomotion

The mathematical "no-collision" guarantee a safety filter provides holds only when its safety parameters match the task and environment. The authors show that an ordinary user, with no extra privileges, can make the filter fail simply by inserting specific goal points. They also built SIREN, an automated vulnerability-finding tool: it found attacks in all vulnerable scenarios in the SPARK dataset, 3 times as many as baseline methods, and caused a real humanoid to collide while performing a task.

Tiffany Bao et al. · arXiv 2610.31923 source

Backtrack: a frozen VLA senses impending failure and backs up to retryvla

Once a VLA's own actions carry it into states rarely covered by training data, it often keeps going wrong. Backtrack attaches a failure-prediction head to the VLA's existing internal representations, adding only 0.48% to 0.88% of parameters. When failure is predicted, the arm retraces its just-travelled trajectory and hands control back to the original policy to retry, with no recovery demonstrations and no policy retraining. On LIBERO-Long, failure rates of SmolVLA, π0.5, and π0-FAST each fall by 19% to 48%.

Thomas Landeg et al. · arXiv 2610.32941 source

EgoForesight: noisy egocentric action labels can still be used to scale up trainingvla

Egocentric human video offers broad coverage, but the action labels obtained through reconstruction and retargeting are noisy. EgoForesight assigns noise levels and weights to action supervision according to the reliability of each data source, then has a teacher network that has seen clean future frames distill task outcomes into the student. At 30,000 hours of pretraining data, open-loop action error on real data falls 20.5% and RoboCasa few-shot success is 8.1 percentage points higher, with a steeper slope than the baseline as data volume grows.

Chubin Zhang et al. · arXiv 2610.31845 source

FORCA: a coding agent that reads wrist force sensingmanipulation

The agent reads both vision and wrist force/torque data and writes Python control programs with contact detection and force feedback. After each execution, it diagnoses failures against the force curves and camera images and revises the code, with no demonstrations, task training, or simulation. Across five real-robot tasks, it grasps more accurately than the same agent using vision alone and loses contact less often when wiping and peeling. Wenzhen Yuan is a co-author.

Xiao Tan et al. · arXiv 2610.31841 source

CASBOT (Chinese humanoid robot startup) ｜ Series A++ ｜ several hundred million yuan ｜ valuation undisclosedhumanoid ⚠️ Vendor-reported

The round was co-invested by Zhilai Investment, Yonghua Investment, Maotian Investment, Yuanhe Chenkun, Guoke Kezhi, and Luning Investment, about 16 months after its Series A of nearly 100 million yuan led by Lens Technology (Chinese glass and electronics maker). Founded in 2023, the company's products include the full-size humanoid CASBOT 02 and the wheeled-arm W2, with the body, embodied brain, and dexterous hand all developed in-house. Orders come mainly from shoe-making robots: in August 2025, the first unit entered a shoe factory in Putian for a proof of concept on insole stamping; the company says total shoe-making robot orders now exceed 400 million yuan, with production scheduled at the scale of several thousand units. It also has a downhole charging robot for mines, which has entered a real mine for on-site training. Sources: InfoQ source; Sohu source

MoSense (Chinese tactile-skin startup) ｜ Angel+ round ｜ hundreds of millions of yuan ｜ valuation undisclosedhardware

Co-led by Hony Capital (Chinese private equity firm) and Oasis Capital, with follow-on investment from LimX Dynamics (Chinese humanoid maker), Loongson Venture Capital, CAS Star (Chinese deep-tech VC), Huashan Capital, and Pudong Venture Capital. The company was founded only in May 2026; its angel round involved Sequoia China, GL Ventures (Hillhouse's venture arm), and AgiBot (Chinese humanoid robot maker), so it has completed two rounds in under half a year. The team was incubated out of a championship-winning HKUST robotics competition team and makes MoSkin, a whole-body electronic skin for humanoid robots, along with multimodal tactile algorithms and a tactile world-action model. After the two rounds, its shareholders already include two complete-robot makers, AgiBot and LimX. Source: Guandian source

INFIFORCE (Chinese embodied-AI startup) ｜ investment from Zhejiang Venture Capital ｜ nearly 100 million yuan ｜ valuation undisclosedworld-model ⚠️ Vendor-reported

The investment was led by Shen Qianyang of Zhejiang Venture Capital, who previously took part in investing in Zhang Xue Motorcycles (Chinese motorcycle brand). INFIFORCE builds the embodied brain AtomBrain, focused on causal world models; chief scientist Chen Jiayu is an assistant professor at the University of Hong Kong. The company says its business covers 8 industries and more than 100 real-world scenarios, with cumulative orders above 500 million yuan, but it did not break down customers or delivery progress. Source: 36Kr source

Noitom Robotics (Chinese motion-capture company) ｜ Pre-A+++ round ｜ amount undisclosed ｜ valuation undisclosedadjacent

Huirong Fund was the sole investor. Founded in 2012, the company started in motion capture and now turns human-performed collection into an embodied training-data service, with three data systems: large-space interaction, dexterous manipulation, and egocentric in-the-wild collection, using cross-embodiment mapping so data can be reused across robots. Source: PEdaily source

SKAI Worldwide raises stake in SKAI Intelligence (South Korea) ｜ stake rises to 66.92% ｜ 25.35 billion wonadjacent

SKAI Worldwide is buying 6,500 shares for 25.35 billion won, raising its stake from about 42% to 66.92%, with the deal scheduled to close on January 28, 2027. Ninety percent of the consideration is paid in 30-year perpetual convertible bonds, with only 2.535 billion won in cash. SKAI Intelligence builds digital twins on Nvidia Omniverse and generates synthetic data for industrial robot training. Source: The Asia Business Daily source

Misumi Ventures (US) ｜ new early-stage fund ｜ $50 millionadjacent

Japanese industrial components supplier Misumi set up this fund in the Bay Area, planning to invest in 20 to 30 seed and early-stage companies, with initial checks of $500,000 to $1.5 million. For the first few years it will follow rather than lead, and the Misumi Group is the sole LP. Misumi Americas CEO Dave Evans said: "We watched a generation of talented hardware teams get stuck at the transition to mass production. That is a supply chain problem." Portfolio companies get access to Misumi's 30 million standard parts and 22 factories. Source: Robotics & Automation News source

XPeng Robotaxi receives citywide L4 demonstration operation qualification in Guangzhouautonomy

XPeng has been approved for Guangzhou's intelligent connected vehicle application demonstration operation, allowing its Robotaxi to conduct L4 demonstration operations across all of Guangzhou. The company frames this step as moving from technical validation to operational and business-model validation, with order response, dispatch, pickup and drop-off, and exception handling all to be tested in real mobility. It will still open gradually through invitation codes at first. The vehicle is the "XPeng Youyou" (小鹏悠游), based on the XPeng GX, with 4 in-house Turing chips delivering 3,000 TOPS, running the second-generation VLA. He Xiaopeng (XPeng's CEO) previously said XPeng will not directly operate a large fleet itself, with revenue coming from hardware sales, technical services, and commission sharing; the company is in talks with several leading European mobility platforms. The announcement gave no fleet size and did not say whether safety operators are on board. Sources: XPeng official source; EE Times China and others source

GXO expands Agility Digit to two new warehouseshumanoid

Digit will enter two GXO e-commerce fulfillment warehouses, in Statesville, North Carolina and Lebanon, Tennessee, handling repetitive material movement, going live in early 2027. The two companies announced the expanded partnership on October 26. The new sites will start with Digit 4 and switch over once Digit 5 is commercially available. The two signed a multi-year RaaS agreement in 2024, starting at the Flowery Branch, Georgia warehouse, where Digit places totes onto conveyor lines and has moved more than 100,000 to date. Agility CEO Peggy Johnson said: "Customers are no longer asking whether humanoid robots can do the work, but deciding where to scale." Automotive World's assessment is that roughly 100,000 totes in about two years is not much by warehouse standards, and the weight of this expansion lies in showing it can be replicated across multiple sites, which is exactly the evidence Agility needs before going public. Agility says Digit has logged more than 65,000 hours of real-world work, its RoboFab factory in Oregon has annual capacity of up to 10,000 units, and its merger with Churchill Capital Corp XI is expected to close in the fourth quarter. Sources: Securities.io source; Automotive World source

AgiBot Genie G2 Max enters JD's Zhilang warehouse, destacking 80 boxes per hourembodied ⚠️ Vendor-reported

JD Logistics (logistics arm of Chinese e-commerce giant JD.com) is using AgiBot's newly released Genie G2 Max for inbound handling in its Zhilang warehouse; the figures from the two companies are 80 boxes destacked per hour and 75 boxes palletized, with a task success rate above 99%. Another deployment is on the Qiyuan T1 production line at Huguang Co., where it delivers shoulder-joint housings and hip-joint assemblies to final-assembly stations. Neither JD nor AgiBot disclosed the number of units deployed, and the success rate and cycle time are self-reported by the vendors. Sources: JD Logistics blackboard bulletin, via Sina Finance source; Mydrivers source; SmartHey source; Shanghai Observer and others source

Tesla expands its Houston Robotaxi service areaautonomy

The new service area extends southeast from its former northwest corner, passing through the central city and covering downtown Houston and Bellaire; Tesla announced on the official Robotaxi X account on October 26 that it is in effect. Houston and Dallas opened unsupervised passenger service on April 18, just over six months ago. There is no Cybercab in the Houston fleet yet. Source: Drive Tesla source

Three Lingxi X2 robots go on regular duty at Chengdu's Kuanzhai Alleyhumanoid ⚠️ Vendor-reported

In a partnership between AgiBot and Chengdu Cultural Tourism Group, three Lingxi X2 robots handle greeting and inquiries, Sichuan-dialect cultural guided tours, and convenience retail. The retail post is driven by an in-house whole-body VLA and handles picking up and handing over items. AgiBot says it already supports more than 20 SKU types with a success rate above 98%, at about 30 seconds per transaction. Source: Gasgoo source

Shenmou's arm-equipped drone "Longzhua" wins a State Grid power projectembodied ⚠️ Vendor-reported

Following this month's win of a State Grid (China's national utility) cable-tunnel project by its arm-equipped quadruped robot "Fumo," Shenmou's aerial-work drone "Longzhua" has now won State Grid's "High-Stability Drones for Aerial Work" project. The airframe carries a robotic arm and dexterous hand to clear foreign objects such as bird nests and plastic sheeting from overhead lines, handling detection, removal, and re-inspection with a single aircraft. The contract value and quantity were not disclosed. Citing the company's research, it says annual procurement demand of this kind in a single province is about 150 million to 200 million yuan. Source: Ikanchai source

Charger Logistics plans to buy up to 500 Aurora driverless trucksautonomy ⚠️ Planned figures

The two sides signed a memorandum, under which Charger intends to buy up to 500 trucks equipped with the Aurora Driver, expanding on a contract signed in July. Hirschbach previously signed a memorandum of similar scale for 500 trucks, with Charger the second; neither is binding. Source: Business Wire, via News-Press NOW source

QuantGroup signs a framework agreement to supply up to 2,000 store-robot solutions to Chewujieworld-model ⚠️ Planned figures

Hong Kong-listed QuantGroup (量化派, 02685.HK), a Chinese fintech company, will supply Chewujie with no more than 2,000 embodied-intelligence solutions for car showrooms and 4S dealerships, handling greeting, model explanations, and lead collection. QuantGroup does not build robots; partners supply the hardware, while it provides the models and the store dispatch system. The 2,000 figure is a cap, and actual delivery depends on pilot results. The payback estimate in the 36Kr article comes from a trial restaurant in Beijing: monthly labor cost per store fell from 35,000 yuan to 20,500 yuan, and assuming a solution cost of 200,000 yuan, payback comes to 13 to 14 months. Source: 36Kr source

Volkswagen China obtains an L3 road-test permit in Hefeiautonomy

On October 26, Volkswagen Group China announced it had obtained an L3 autonomous-driving road-test permit in Hefei, and real-road testing and data collection have begun. Under the group's earlier plan, L3 technology will be ready in 2027, with the first L3 solution delivered as early as the second half of 2027. The software is developed by Carizon, the joint venture between Volkswagen and Horizon Robotics (Chinese automotive chipmaker), using a white-box licensing model to build a unified solution on Horizon's large model; the in-house C7H chip is designed to support up to L4. This is a test permit; for production vehicles to go on sale, a vehicle-type approval from the Ministry of Industry and Information Technology is still required. On the same day, what XPeng received was an L4 demonstration operation qualification. Sources: Autohome source; Guandian source

Unitree (Chinese robot maker) closes at a market cap of 271.8 billion yuan and issues another unusual-trading noticehumanoid

On the afternoon of October 26, Unitree's shares surged, with market capitalization at one point approaching 300 billion yuan, closing at 672 yuan for a market cap of 271.8 billion yuan. The unusual-trading notice issued that evening is the second recently; over the last 7 trading days, the stock has risen more than 65% cumulatively. The notice lists a trailing price-to-earnings ratio of 465.22, versus an average of 38.54 over the past month for the general equipment manufacturing industry it belongs to. The notice states that as physical AI moves from lab demos to large-scale deployment, it still faces challenges in perception accuracy, motion-control stability, data accumulation, and hardware-software coordination, and that the maturity of the technical path and business model "still needs time to be validated." The timeline founder Wang Xingxing gave at Unitree's first partner conference on October 22 is much nearer: he expects major breakthroughs within one to three years, hopes Unitree robots will enter ordinary households at the end of 2026 or in 2027, and plans to incubate more than 200 partner companies within five years. Sources: Sina Finance source; KrASIA source

Toyota and Honda test humanoid robots in Japanese factorieshumanoid

CNMO, citing foreign media reports, says the test tasks range from wire-harness installation to final inspection. Toyota is using its in-house ELEY, a wheeled-base robot weighing about 50 kg with two fingers on each hand. Honda showed its Avatar Robot at the Japan Mobility Show 2026, which is remotely operated by a human with AI assisting fine finger movements; in the on-site demo, it picked up a screw and used a wrench to tighten it into a transmission housing. After checking multiple reports from October 26, shattered.io said Toyota has not disclosed the number of units tested or a timeline, and that it could not find primary material showing the robots already working on a production line. Sources: CNMO, via Sohu source; Financial Times source; shattered.io

Guangdong's industrial robot output grows 34.2% in the first three quartersindustrial

Guangdong's industrial robot output in the first three quarters grew 34.2% year on year, 5.1 percentage points above the 29.1% national growth rate across China published by the National Bureau of Statistics. Over the same period, the province's 3D printing equipment output grew 71.0% and integrated circuits grew 29.7%. Source: Science and Technology Daily, via Sina Finance source

OneTrack releases Iris-1, an industrial video world modelworld-model

Iris-1 is built specifically to understand surveillance video from warehouses and factories, describing events on site, supporting natural-language search of footage, and flagging segments that need review. The company says that in an internal test on safety events held out from training, its accuracy in identifying real collisions is 3 times that of the runner-up, Gemini 3.8 Flash. CEO Marc Gyöngyösi said: "General-purpose AI wasn't built for that world." Customers are already using it to check protective-equipment violations, falling heavy objects, and accident root causes. ⚠️ Vendor-reportedSource: AiThority source

Pronto founder starts a physical AI data company in the USadjacent

Anjali Sardana, founder of Indian on-demand home services platform Pronto, has founded Provision Labs in San Francisco, collecting data from practitioners across industries worldwide to train and test physical AI. Inc42 notes that Pronto's earlier pilot of recording video in users' homes drew scrutiny; the company said at the time that participation in the pilot was entirely voluntary and covered fewer than 0.01% of users. Inc42 also named three issues this kind of data business must face: informed consent, privacy, and worker compensation. Source: Inc42 source