Today's Highlights

· Draft rules from China's Ministry of Industry and Information Technology (MIIT): only vehicles equipped with L3 or higher autonomous driving may fit ADS marker lamps, and the "little blue lights" on L2 models would be banned

· Schaeffler's humanoid robot business has accumulated €400 million in orders, and 6 humanoids are assembling joint modules at its new Taicang plant

· 96 robots sharing execution experience lifted the success rate of a multi-stage phone-packaging task from 79.2% to 98.3%

· Frontier large models controlling real robot arms directly: benchmark scores doubled roughly every 44.5 days between July and September

· Dream Labs, founded by Nvidia's DreamZero team, raised a $90 million seed round; its robot-brain API costs $3 per hour

· Neolix (Chinese autonomous delivery vehicle maker) closed a Series E of nearly $400 million, with more than 30,000 unmanned vehicles in operation

Fleet-shared experience: 96 robots learning together lift phone-packaging success to 98.3%manipulation

HF 103↑. The larger a fleet gets, the harder it is to keep people watching to correct errors at all times; this paper lets deployed robots borrow experience from one another. The authors propose Collective Experience Learning (CEL): when one robot fails, its situation is used to retrieve relevant segments from the successful trajectories shared across the whole fleet, and these serve as the supervision signal. No additional expert demonstrations are collected, and no manual action corrections are needed. Success rates on five real-robot manipulation tasks range from 93.3% to 99.2%, and the robots used for testing did not take part in data collection. With per-robot collection time held fixed, success rates generally rise as the fleet grows; on the multi-stage phone-packaging task, 96 robots learning together lifted the success rate from 79.2% to 98.3%.

Chao Yu et al. · arXiv 2610.25736 source

Large models directly controlling real robot arms: benchmark scores double roughly every 44.5 daysbenchmark

Seven frontier multimodal large models, each run at several reasoning-effort levels, directly controlled real robot arms across six RoboDojo tasks. The authors say this is the first time so many frontier models have been compared side by side on the same set of real-robot tasks. Current models can already complete a range of manipulation tasks, but are not yet accurate or fast enough within practical action budgets. Progress is quick, though: between July and September, benchmark scores doubled every 44.5 days, the best execution rate rose to 4.4 times its earlier level, and the lowest execution cost fell to 1/6.2 of its earlier level. Authors include Sven Koenig.

Christopher Leet et al. · arXiv 2610.25287 source

Feeding tactile images to GPT-6 Astra beats specialized policies on simulated contact tasksmanipulation

Contact-rich manipulation has long been regarded as a stronghold of specialized policies. This paper feeds RGB images captured by visuo-tactile sensors directly into GPT-6 Astra through its existing vision interface for few-shot control. Across 19 tasks in the two simulation suites UniVTAC and OpenViTac, average success rates are 93.8% and 81.8%, which are 22.9 and 13.1 percentage points higher than the strongest specialized baseline in each comparison. Vision alone already leads the baselines (76.3% and 80.6%). The authors themselves note that the benchmarks show signs of saturation, and the results have not yet been validated on real robots.

Chen Yang et al. · arXiv 2610.26846 source

ScaleNav: pretraining language-guided navigation on internet first-person videoautonomy

HF 82↑. Real-robot navigation data is expensive, and simulation data suffers from a sim2real gap. ScaleNav first learns from internet first-person video how the camera moves step by step toward the target named in an instruction, then uses action-annotated simulation data to align this into executable action chunks. It reaches a success rate of 86.5% on single-view Short-Horizon OVON and 94.9% on the self-built HabitatGS-LN, and the model has also been deployed on a real quadruped robot.

Yihui Yang et al. (Zhejiang University) · arXiv 2610.25690 source

Ropedia W0: a mobile-manipulation world model trained on 10,000 hours of human data plus 3,000 hours of robot dataworld-model

The authors call this the first first-person mobile-manipulation world foundation model that covers both humans and multiple robot embodiments. Its training data is 10,000 hours of human first-person video plus 3,000 hours of robot data. The model encodes locomotion with camera poses and manipulation with projected skeletons, and achieves SOTA on five cross-embodiment benchmarks, including AgiBot (Chinese humanoid robot maker). The more human pretraining data is added, the better the evaluation scores on both the human and robot sides. After distillation it can generate minute-long rollouts in real time and can connect to VR interaction.

Tianqi Liu et al. · arXiv 2610.26855 source

LIBERO-Pro2.0: 17 policies lose 26.5 points on average under runtime interferencebenchmark

A high score on LIBERO does not mean a model is robust. The new version adds 22 kinds of static perturbations and 20 kinds of runtime interventions, for 1648 shifted tasks, each checked through six rounds of human review, along with 39390 training trajectories. All 17 evaluated VLAs, world action models and robustness-specific policies lose performance: under runtime interventions the average drop is 26.5 percentage points, and the largest is 32.5.

Xueyang Zhou et al. · arXiv 2610.24587 source

ContactBench 1.0: policies with added touch show no clear improvementbenchmark

The authors group everyday contact manipulation into eight skill categories and build 36 reference tasks. More than 40 objects can be reproduced with consumer-grade 3D printers, and simulated models with adjustable resistance are provided, so different labs can reproduce the same set of tasks. On six of the tasks, vision-only policies struggle to detect and respond to contact events, and visuo-tactile policies with tactile input added show no clear improvement either. Authors include Jeannette Bohg and Wenzhen Yuan.

Jingyi Xiang et al. · arXiv 2610.24512 source

FIBER: unifying humanoid loco-manipulation planning in the latent space of a behavior foundation modellocomotion

FIBER treats the executable latent space of a frozen behavior foundation model (BFM) as a shared planning representation: separately trained skill policies are first projected into it, and a vision-based flow planner is then trained offline. In simulation, with 128 variants tested for each of three task types (box carrying, door opening and grasping), the average success rate is 42.0%; replaying latent commands directly reaches only 11.6%, and kinematic planning only 3.5%. On a real Unitree (Chinese robot maker) G1, sim2real transfer and skill chaining were achieved. Authors include Guanya Shi and C. Karen Liu.

Yutang Lin et al. · arXiv 2610.24675 source

A quadruped robot solves a giant Rubik's Cube using only its feetlocomotion

An onboard RGB-D camera pieces together the cube's pose and the colors of all its faces while the robot walks and flips the cube. When turning a layer, one foot holds the cube while the other pushes; all of these whole-body skills are trained with reinforcement learning. The real robot completed three full autonomous solves, each taking about 30 moves and 8 to 9 minutes, and it recovers after external disturbances.

Lukas Amiet et al. (ETH Zurich) · arXiv 2610.24610 source

Neolix (Chinese autonomous delivery vehicle maker) | Series E | nearly $400 million | valuation undisclosedautonomy

BoYu Capital (Chinese investment firm) and a national-level fund co-led the round; Chedongxi (Chinese auto media outlet), citing investment-community sources, says the latter is the National AI Industry Investment Fund. Participants include Alibaba's AEF NextGen Hong Kong fund, Lisi Capital, GGV Capital, and Temasek-backed Vertex Growth, while existing shareholders including SoftBank Asia, Yizhuang International Investment and Gaorong Capital also followed on. By Neolix's own count, this is the largest single financing in China's autonomous driving sector in 2026, and the company's cumulative funding has passed RMB 10 billion. The previous round was a Series D of more than $600 million in October 2025, when cumulative deliveries had only just passed 10,000 vehicles; its unmanned fleet now exceeds 30,000 vehicles, covering nearly 20 countries and more than 300 cities. The new funds will go to its global operations network and an end-to-end autonomous driving large model. Source: Chedongxi source

Dream Labs | Seed | $90 million | valuation undisclosedworld-model

The team comes from Nvidia's DreamZero project; the CEO is Joel Jang and the CTO is Yunhao Ge. Two outlets differ on who led: SiliconANGLE writes that B Capital and Kindred Ventures co-led, while Jiqizhixin (Chinese AI news outlet) writes that B Capital led and Kindred followed. The Dream API, launched the same day and open to the public, is backed by the first model, Dream-1, at $3 per hour, with the lighter Dream-1-flash at $2 per hour. Dream-1 is a world action model: it learns to predict upcoming frames from videos of people doing everyday tasks, then generates robot actions from those predictions; dual-arm robots with grippers can use it zero-shot. The company says more than 20 hardware makers have joined. The company also cites a third-party evaluation, Robocurve's RoboDojo-RC Tier 1: Dream-1 scores 59% per attempt on average, versus 48% for Claude Opus 5.5 and Gemini 4 Argon and 47% for GPT-6 Astra, with a single Dream-1 attempt costing about $0.15. ⚠️ Vendor claimSource: SiliconANGLE source; Jiqizhixin source

Eka Robotics | Series A | $150 million | post-money $1.45 billionembodied

Led by a16z and F-Prime. The founders are MIT professor Pulkit Agrawal and former Google DeepMind researcher Tuomas Haarnoja, and the product is a grasping system: a force-sensitive gripper paired with a new kind of deep learning model. TechCrunch says it can pick about 1,500 items per hour weighing under 1 kg, versus about 600 for a person; Eka says its grasp success rate exceeds 95%. a16z partner Raghu Raguram says: "On some tasks they have already surpassed human level." UIUC professor Kris Hauser, after seeing a demo, said it can adapt to a large variety of objects without retraining. Source: TechCrunch source

Weiyi Zhizao (Chinese industrial AI inspection firm) | Hong Kong IPO subscription | up to about HK$800 million | offer price HK$16.63-21.46industrial

After passing the Hong Kong Stock Exchange listing hearing in October, Weiyi Zhizao (03275.HK) is taking subscriptions from October 20 to 23 and plans to list on October 28, with Orient Securities International as sole sponsor. The offering is 37.2758 million H shares, with 10% allocated to the Hong Kong public offering. There is only one cornerstone investor: Yibei Rhine Investment subscribed for about HK$15.7 million, less than 2% of the maximum raise. For the five months to May 31, the company's revenue was RMB 423 million, up 22.6% year on year, with profit for the period of RMB 22.5 million; revenue from outside mainland China over the same period was RMB 963,000. Source: The Paper source

Yita Power (YITA, Shenzhen) | Seed + Seed+ | over RMB 100 million | valuation undisclosedhardware

Investors include Matrix Partners China, SenseTime Guoxiang Capital, CCB International and L2F Light Source Entrepreneurs Fund, as well as two humanoid makers, Galbot (Chinese humanoid robot maker) and Fourier (Chinese humanoid robot maker). The company was founded in May this year and builds force-controlled dual arms and integrated joint modules. Its five released modules weigh 0.3 to 1.6 kg with rated torque of 2.2 to 62 N·m, and the company says costs are about 50% lower than the harmonic-drive solutions it compared against. Founder Zhong Xuegang was formerly R&D director for the powertrain segment at Fourier's humanoid robots and led the development of more than 20 actuators and joint modules. Source: Robot Qianzhan source

Yitai Embodied | AgiBot and Huawei's Hubble take stakes | amount undisclosedadjacent

According to Qichacha (Chinese business registry platform), this embodied-AI data company, founded in 2025, has added two shareholders: AgiBot Innovation (Shanghai), an affiliate of AgiBot, and Shenzhen Hubble Technology Investment Partnership, which is Huawei-affiliated. Registered capital has risen to RMB 1.273 million. Source: Shanghai Securities News source

Hertz | Steve Cohen investment | $80 millionadjacent

According to an Axios exclusive, Steve Cohen is investing $80 million in car-rental company Hertz to support its robotaxi plans. Source: Axios source

Schaeffler's Taicang plant opens: 6 humanoids assemble joint modules, humanoid business orders reach €400 million cumulativelyindustrial

Six humanoid robots have joined the joint-module assembly line at Schaeffler's Taicang digital and intelligent plant in China, handling material transport, part sorting, press-fitting, inspection and packaging; the joint modules they assemble are themselves components for humanoid robots. The plant opened on October 20 and covers about 7,500 square meters. It also has six types of industrial training scenarios, including visual inspection, material picking and part assembly, using digital twins to link scenario training, production validation and process optimization. Schaeffler disclosed the same day that it won new orders from two Chinese customers in the third quarter, bringing the group's cumulative humanoid robot business orders to €400 million. In September it had just registered an embodied-intelligence robot company in Taicang, focused on core components and subsystems for humanoid robots. The report does not say which company's humanoid robots were put into operation. Source: Gasgoo source

Waymo reportedly aims to expand its fleet to 20,000 vehicles by the end of 2027, and announced Madrid for 2027 the same dayautonomy ⚠️ Rumor

Semafor, citing people familiar with the matter, says Waymo plans to grow its fleet from about 4,000 vehicles to 20,000 by the end of 2027, reaching 8,000 first this December; Waymo declined to comment. Madrid is publicly announced by Waymo: Jaguar I-PACE vehicles will begin appearing on the streets in the coming days, driven by specialists at first, with the goal of launching fully driverless ride-hailing in 2027. Source: LavX citing Semafor source; Wangtongshe source

A Kodiak driverless truck runs 64 consecutive days without a power cycle or rebootautonomy ⚠️ Vendor claim

From June 30 to September 1, a driverless truck fitted with the sixth-generation Kodiak Driver in the Permian Basin stayed powered on and on standby for 64 consecutive days, with neither the autonomous driving system nor the vehicle ever power-cycled or rebooted; five other trucks each ran more than 45 consecutive days. These trucks belong to a 35-truck driverless fleet owned and operated by Atlas Energy Solutions. Kodiak has committed to expanding the Atlas fleet to 100 trucks by mid-2027. Source: GlobeNewswire source

AGCO signs with Agility to trial Digit at its Kansas combine plant in 2027humanoid ⚠️ Planned

Agricultural machinery maker AGCO and Agility Robotics announced a commercial agreement on October 20. The US pilot of the Digit humanoid robot is expected to begin in 2027 at AGCO's plant in Hesston, Kansas, which produces combine harvesters and hay equipment for the Fendt and Massey Ferguson brands. Digit will start with part sorting and stacking and with placing bins into storage, orchestrated by Arc, Agility's cloud fleet management platform. AGCO's manufacturing technology director Thomas Enghof said the partnership treats humanoid robots and AI as "teammates" in building flexible, intelligent factories. Source: Unite.AI source

AgiBot Expedition A3 Ultra runs 18 days and 827 orders at a retail warehouse on the Wuhan riverfronthumanoid ⚠️ Vendor claim

As of October 19, the full-size bipedal humanoid Expedition A3 Ultra had operated for 18 days at a retail warehouse in the "Jiangyu" bridge-pier film camp on the Hankou riverfront. Working 8 hours a day against the peak flow of night-tour visitors, it completed 827 orders in total, with an overall grasp success rate above 97%. The robot picks items in the warehouse and also chats with visitors; the partners are Wuhan Culture and Tourism Group and Zhongyun Weixun. That works out to about 46 orders per day. Galbot's Shanghai pharmacy instant-retail warehouse, disclosed in September, fulfills 400 to 500 orders per day; the two settings differ, and the scale differs by roughly tenfold. Source: Sohu source; Ziniu News source

MIIT proposes banning "little blue lights" on L2 vehicles, reserving ADS marker lamps for L3 and aboveautonomy

On October 20, the Equipment Industry Division I of China's Ministry of Industry and Information Technology (MIIT) released the draft "Installation Requirements for Exterior Lighting and Light-Signaling Devices on Automobiles and Trailers (Draft for Comment)": vehicles not equipped with an L3 or higher autonomous driving system may not be fitted with ADS marker lamps. The lamp color is set as blue-green. Compliant L3-and-above vehicles may optionally fit front, rear and side marker lamps, and trailers paired with L3 tractors may optionally fit rear and side lamps. Drivers cannot switch the lamps on or off manually; they light automatically only when an L3-or-above function is actually active and go off when the system exits, and lighting them while L2-or-lower functions are running is strictly prohibited. Many Chinese new-energy automakers had previously added blue exterior indicator lights to L2 models, with differing brightness, lighting patterns and mounting positions. Some cause glare for surrounding vehicles at night, and some other road users mistook the cars for being already in autonomous mode and then tailgated closely or cut in aggressively. According to First Financial, several automakers began receiving notice of the ban as early as July this year. Source: Kuai Ke Ji source

Unitree's share price deviates by more than 30% cumulatively over three days; the filing warns of valuation-reversion riskhumanoid

Unitree Robotics (688836.SH) announced that its closing-price gains deviated from the benchmark by a cumulative 30% or more over three consecutive trading days, which counts as abnormal trading volatility. The announcement states that the humanoid robot industry is still at an early stage of commercialization, with uncertainty in technological maturity, downstream demand cultivation and supply-chain support. If large-scale commercial rollout falls short of expectations, the company's revenue expansion and earnings release could be affected, and there is an "investment risk of valuation reversion." Source: 21st Century Business Herald source

Two German ministries sign a letter of intent with Nvidia to build industrial AI on open modelsadjacent

Germany's Federal Ministry for Economic Affairs and Energy and its Federal Ministry of Research, Technology and Space signed a joint letter of intent with Nvidia at the Invest in Germany summit, with Chancellor Merz present. The parties will build industry-specific open models on the basis of open models such as Nemotron and Cosmos. Nvidia will also extend its existing collaborations from German research institutions and industrial companies to small and medium-sized enterprises. The German government is also refocusing the "Industrie 4.0 Platform" on industrial AI, as the central network for this work. Source: AI Insider source

Counterpoint expects global quadruped robot shipments to exceed 90,000 units this year, with industrial-grade Zhiyuan Kuotuo ranked firstadjacent ⚠️ Estimate

According to Counterpoint data cited by Asia Economy, global quadruped robot shipments are expected to exceed 90,000 units this year, nearly double last year's 46,000, with Chinese makers holding more than 90% of the share. Of about 35,000 units in the first half, about 14,000 were used at industrial sites. In the overall ranking Unitree leads with 37%; counting only industrial-grade units, the ranking becomes Zhiyuan Kuotuo 27.8%, Deep Robotics (Chinese quadruped robot maker) 26.0%, Genisom AI 18.9% and Unitree 17.6%. The same report gives global humanoid robot shipments of 22,000 units in the first half, with AgiBot's share at 43%. Frost & Sullivan gives sales of 36,000 units and IDC gives shipments of nearly 25,000 units, which makes this the third differing measure. Source: Asia Economy source

UBTech and Ras Al Khaimah sign a memorandum, taking humanoid robots to the Gulf for climate testinghumanoid ⚠️ Planned

On October 19 in Shenzhen, the International Relations Office of the Government of Ras Al Khaimah and UBTech (Chinese humanoid robot maker) signed a strategic cooperation memorandum of understanding, witnessed by the Ruler of Ras Al Khaimah. UBTech will deploy humanoid robots locally and send engineering teams for on-site demonstrations, with the first phase covering industrial assembly, public services, elderly care and culture and tourism. The two sides will also explore setting up a "Ras Al Khaimah Humanoid Robot Climate Research Program" to test how robots actually perform in the Gulf climate. Autohome's headline said "$1 billion," but the joint press release gives no amount. Source: Industry Online (PR Newswire release) source; Autohome source

Chery sets a target of deploying 100,000 robotaxis overseas by 2030autonomy ⚠️ Planned

At the 2026 Chery Global Innovation Conference, Chery (Chinese automaker) set a target of deploying 100,000 robotaxis overseas by 2030, with the approach broken down by region. In the Chinese market it will enter through a vehicle platform plus operations; in the Middle East it will start with the vehicle platform and then gradually expand into operations; in Europe it will go straight to robotaxi operations; in Southeast Asia the vehicle platform, operations and algorithms will land together; in the Americas it will prioritize the vehicle platform. The same event also unveiled the Ziwei chip and an L4 robotaxi model. Source: DoNews source; Autohome source

The first engineering prototype of the native robotaxi built by Geely and Cao Cao Mobility rolls off the lineautonomy

This vehicle has been developed for robotaxi use from the outset; the first engineering prototype has rolled off the line, and a Cao Cao Mobility (Geely-backed ride-hailing company) custom version is planned for mass production in 2027. Source: D1EV source; Sina Finance source

Hyundai and Kia to run L4 logistics routes with two Xcient heavy trucks in early next yearautonomy

According to The Korea Times (relayed by Finimize), Hyundai and Kia, together with Hyundai Mobis, Hyundai Glovis and autonomous driving startups Mars Auto and RideFlux, signed a memorandum. Two Xcient trucks, one of them a hydrogen fuel-cell version, will be put into service in early next year and tested on two fixed routes: one of about 100 km, a loop mixing highway and urban roads, and another of about 400 km, a long-haul highway route. Source: Finimize source