Key Takeaways

· NHTSA has opened a docket on a petition over Tesla Robotaxi speeding in school zones, estimating about 1,000 vehicles affected

· Shanghai has set a target: push 8,000 humanoid robots into factories by 2028

· The US FCC will no longer grant certification to new models of foreign-made humanoid and mobile robots; Rainbow Robotics became the first Korean company to apply for an exemption

· A London Assembly transport committee report makes 13 recommendations, calling for TfL to gain powers to cap Robotaxi numbers and revoke licences

· Frost & Sullivan: full-size embodied-AI humanoids made up only about 17% of unit sales in the first half but nearly 60% of revenue

· Reflex lets a Unitree (Chinese robotics maker) G1 catch a box thrown at it using only its own vision, with a 65% success rate at 2.5 meters

Reflex: a humanoid robot catches a flying box using egocentric vision · locomotion

HF 125↑, the most popular item today. Catching a box requires a humanoid to coordinate its whole body, keep up with the object's motion, and perceive using only its own vision, all within about one second; most existing systems solve only part of this. Reflex learns in three stages: first, reinforcement learning trains whole-body catching; next, it learns to infer the box's dynamics from delayed, incomplete observations; finally, the controller is frozen and an RGB-D encoder is trained to recover this dynamics representation directly from images. In simulation, vision-only catching succeeds 85.7% of the time, versus 87.9% with privileged observations; on a real Unitree G1, catch success at 2.5 m and 5 m is 65% and 40%, roughly 76% and 59% of human performance at the same distances.

Taoyang Jia et al. · arXiv 2610.12772 source

ATHLETE: teaching a humanoid to play tennis with generic gait data plus ordinary tennis videos · locomotion

There is no dedicated tennis motion-capture data. ATHLETE stitches LAFAN1 gait data together with stroke motions extracted from ordinary video into reference trajectories, then uses teacher-student distillation and PPO fine-tuning to turn them into a reactive policy that needs no reference motion input at deployment. On a real robot, forehand and backhand hit rates are 90.00% and 86.67%, and it can rally with a human for multiple consecutive shots.

Zhihao Tian et al. · arXiv 2610.13572 source

AlphaNIST: with only a binary success reward, one policy learns all 20 tasks on the NIST assembly board · manipulation

Borrowing from AlphaGo Zero, the authors ask how much of an assembly policy needs to be specified in advance. In simulation, a single policy trained from scratch, with no demonstrations, progress rewards, scripted primitives or per-task tuning, reaches a training success rate of 98.2% across the 20 tasks on NIST Assembly Task Board #1. It can also assemble ten pairs of parts in a single episode in an unprescribed order, and spontaneously learns to choose grasp poses based on later assembly needs and to flip parts using surrounding geometry. Results so far are in simulation only.

Octi Zhang et al. · arXiv 2610.13354 source

Splat2Act: scan a room with a phone and adapt π0.5 to the items in your home · vla

π0.5 achieves 0% pick success in the deployment scenes the authors scanned. Splat2Act needs only a phone scan of the scene and photos of the user's items; it builds a digital twin with 3D Gaussian Splatting, automatically finds surfaces and containers where items can be placed, and generates demonstrations. Across 101 campus scenes it produced more than 68,000 pick demonstrations. After two-stage fine-tuning, success in the target scenes is 62.5%; on a real Stretch 3, it reaches 63.9% across 12 subtasks, close to the 66.7% in simulation.

Xiaomeng Ye et al. · arXiv 2610.13945 source

Forward Dynamics Models: supervising actions with future states, with no need to generate the future at inference · vla

VLAs lack supervision from future states during training. WAMs predict the future first and then output actions, which raises inference cost and lets visual prediction errors propagate into actions. FDM conditions future prediction on action features but not the reverse, so ground-truth futures can supervise the policy while prediction errors cannot flow back into actions, and inference skips future generation entirely. It works with both VLM and video-generation backbones, with an average LIBERO success rate of 99.4% and real-robot tasks improving by up to 13% in absolute terms over VLA and WAM methods.

Zipei Ma et al. · arXiv 2610.13117 source

LeWAM: end-to-end JEPA, one model that both outputs policy and does planning · world-model

Earlier JEPA pipelines first train representations and dynamics and only add a controller at the end. LeWAM jointly trains next-latent-state prediction and action flow matching, constraining latent states to an isotropic Gaussian; the same model can execute directly and can also use its learned dynamics to evaluate sampled actions. Compared with earlier JEPA visual planning, success on contact-rich tasks rises from 28.6% to 89.7%, long-horizon tasks from 10.9% to 46.2%, and planning is 32.3 times faster.

Minghao Fu et al. · arXiv 2610.12818 source

SmolWAM: compressing Cosmos Policy to run on edge devices · world-model

The strongest world action models rely on video diffusion transformers, which edge devices cannot run. Starting from the 2B-parameter Cosmos Policy, SmolWAM keeps 12 of 28 transformer blocks, rebuilds the inference interface, and reads out actions and future visual latents together in a single rectified-flow step. Trained for only 3.6 GPU-hours, it succeeds in 95.5% of 400 LIBERO episodes, takes 57 ms per 16-step action chunk and uses 2.4 GiB of memory, 11 times faster than the teacher model.

Timi Owolabi et al. · arXiv 2610.13740 source

LAPT-Scale: the more human video, the higher the real-robot success rate · vla

Egocentric human video has no robot action labels. LAPT-Scale learns continuous latent actions from 23,300 hours of video, uses them as implicit supervision for VLA pretraining, and then aligns to the target robot embodiment. As the annotated human video used for pretraining grows from 1,000 to 5,000 hours, real-robot success rises from 63.3% to 85.0%.

Zhiyuan Xiang et al. · arXiv 2610.13732 source

Spa-Bench: VLAs can act, but often can't understand "the second one on the left" · benchmark

A real-robot benchmark that tests six types of spatial-reference tasks on an SO-101 arm and releases data from 4,092 episodes: 1,200 teleoperated demonstrations and 2,892 evaluation episodes from 5 VLA policies. The three fully evaluated policies succeed 67.5% to 73.3% of the time on in-distribution spatial configurations, dropping to 40.7% to 52.0% on out-of-distribution ones; replacing spatial descriptions with direct instructions raises success by 28.3 to 40.0 percentage points.

Justin Tien-Smith et al. · arXiv 2610.12696 source

Decart ｜ Amazon in talks to acquire ｜ about $7 billion · world-model ⚠️ Reported, unconfirmed

Reuters, citing the Wall Street Journal, says Amazon is in talks to acquire Decart at a valuation of about $7 billion, and a deal could come soon if talks go well and no other bidders emerge. Decart develops Lucy, a real-time video editing model, and Oasis, a simulation model for robotics and autonomous driving. In May it closed a $300 million round led by Radical Ventures with participation from Nvidia, at a valuation of nearly $4 billion. Source: AI Insider source

Winverse AI ｜ Seed round ｜ tens of millions of dollars ｜ post-money about $400 million · world-model ⚠️ Reported, unconfirmed

According to Bloomberg, citing people familiar with the matter, DeepSeek (Chinese AI lab) has agreed to lead the first funding round of this Hangzhou robotics AI startup, with funding commitments also from Meituan (Chinese delivery giant)-owned Longzhu Capital and others; the deal has not been finalized. Several foreign outlets headline it as a "$400 million seed round," but Bloomberg's original says the seed round is tens of millions of dollars, with a post-money valuation of about $400 million. Founder Hu Han previously led Tencent's vision-language and embodied foundation model team and holds a PhD from Tsinghua. This would be one of DeepSeek's largest venture investments since its founding; it has previously subscribed to shares in Unitree Technology's listing, and its own new funding round is reportedly at least 80 billion yuan. Source: Dealroom, Yahoo Finance relaying Bloomberg source; Investing.com China source

X Square Robot ｜ New round ｜ hundreds of millions of yuan · world-model

Led by SAIC Financial Holdings (investment arm of Chinese automaker SAIC), this is X Square Robot (Chinese embodied-AI startup)'s second round after a round of around one billion yuan in January this year. The company builds end-to-end embodied foundation models and open-sourced WALL-OSS-0.5 in June. Its industrial logistics sorting solution self-reports an efficiency of 1,816 items per hour and accuracy above 98%. The lead investor is SAIC Motor's industrial investment platform, and sorting, handling and assembly in automobile manufacturing are potential deployment scenarios. Source: Robot Dajiangtang source

Xianyin Future ｜ Angel + Pre-A + Series A ｜ hundreds of millions of yuan · embodied

This company builds cable-driven flexible space robots. Founded in September 2025, it closed three rounds in a row within four months starting in May 2026. The Series A was led by Alibaba and China Mobile's Chain Leader Fund; industry investors in the angel round include Zhiyuan Robotics (Chinese humanoid maker) and Trina Solar (Chinese solar manufacturer). The business targets on-orbit refueling, assisted deorbiting and on-orbit maintenance, with an on-orbit verification mission planned for 2027. Source: Pedaily source; Dealroom source

Oula Wanxiang ｜ Angel+ round ｜ hundreds of millions of yuan · embodied

The company was founded in March this year, and this is its fourth round in half a year, with new and existing shareholders including Chengyuanxiang Fund, Topband (Chinese electronics maker) and Hony Capital (Chinese private equity firm) participating. Founder Zhou Shunbo was Huawei's employee number one in embodied intelligence and head of Huawei Cloud's Physical Intelligence Innovation Lab. The company focuses on household robots with a "trainable" approach; its first product, a "mobile base plus dual arms" system, has entered iteration in maker households. Source: Zhidongxi source

Loomia ｜ Acquired by Unichem ｜ up to $6.3 million · hardware

South Korea's Unichem has signed to acquire all equity of US flexible-electronics company Loomia, whose LEL flexible circuit layer can serve as electronic skin for robot hands, grippers and housings; closing is expected on January 4, 2027. The price includes an earn-out: if, by the end of 2028, the robotics and automotive segments win volume-production orders of 1,000 and 3,000 units per year respectively, each segment gets an additional $500,000 in convertible bonds. Source: BigGo Finance source

Weita Dynamics ｜ Capital increase by Jiangte Motor ｜ 100 million yuan · embodied

Listed company Jiangte Motor (Chinese motor maker) plans to invest 100 million yuan in a capital increase in Weita Dynamics, entering the embodied-AI robotics field. Source: Eastmoney source

Shuangpai Robotics ｜ Series A+ ｜ amount undisclosed · adjacent

Led by Daosheng Capital with Lingdong Ventures following, the products are smart walkers and smart wheelchairs. Daosheng Capital partner Jia Jia said the company has delivered to nearly 30 countries, has nearly 10,000 units in orders on hand, and has production capacity booked through the end of next year. Source: Daily Economic News source

Kodiak driverless trucks: 35 running in the Permian, IKEA lane to go live this year · autonomy ⚠️ Company-reported

As of the end of Q2, Kodiak was commercially operating 35 driverless trucks in the Permian Basin for Atlas Energy Solutions, with cumulative paid driverless hours exceeding 40,000; the business began in December 2024, with no one on board. Next is hauling freight for IKEA between Dallas and Houston, with plans to remove the on-board observer later this year. Founder and CEO Don Burnette said removing the observer "does not depend on any single test or milestone," but on years of operating experience and a safety case. By the company's own readiness metrics, the evidence needed for the long-haul highway safety case was 96% complete at the end of September, versus 93% at the end of August. Closed mining-area routes and interstate highways are two different kinds of setting, and the IKEA lane is the real test of long-distance driverless operation. Source: Fox News source

JAKA says its wheeled humanoid is in volume production at an auto-parts plant, and launches a heavy-payload version at CIIF · humanoid ⚠️ Company-reported

JAKA Robotics (Chinese collaborative-robot maker) says its JAKA Kargo wheeled humanoid robot is in volume production at a leading global auto-parts company, but did not disclose the customer name or unit count. At the China International Industry Fair (CIIF), it debuted the heavy-payload Kargo50 and the Zu50 high-payload collaborative robot. At its booth, two wheeled humanoids worked together, one moving an engine into position and the other tightening screws. Source: Jiemian News source; Pandaily source

Northrop F-35 line: cobots cut bolt-tightening from hours to minutes · industrial ⚠️ Company-reported

Northrop Grumman has fitted collaborative robots to the F-35's Communications, Navigation and Identification (CNI) module to install front-housing bolts. Production line manager Lavell Catlin said the system "cut work that used to take hours down to minutes." Simple Flying notes that the speedup covers a single step in one subsystem, the company has not disclosed the effect on overall module lead time, and neither the US Air Force nor Lockheed has raised F-35 output as a result. Source: Simple Flying source

Canton Fair opens an autonomous bus demonstration route for the first time · autonomy

The 140th Canton Fair opens on October 15, and during the fair Guangzhou will run a circular autonomous bus route stopping at the Shangri-La Hotel, Pazhou metro station, the Canton Fair complex and other stops, operating daily from 11:00 to 16:00, with a bus every 15 minutes. Source: China News Service source

NHTSA opens a docket on a petition over Tesla Robotaxi speeding in school zones, giving its first fleet-size estimate · autonomy

On October 9 US time, NHTSA's Office of Defects Investigation opened a docket for a petition, numbered DP26010. The petition alleges that 2026 Tesla Robotaxis speed in school zones during school hours with the automated driving system engaged. NHTSA estimates about 1,000 vehicles are affected, including Model Y and Cybercab. The evidence is thin: the petition attaches three social media videos, NHTSA itself counted 4 incidents, and there were no collisions or injuries. Under the rules, NHTSA must decide within 120 days of receiving a petition whether to grant it, a deadline of about January 21, 2027; if granted, it becomes a formal defect investigation that could end in a recall. By Electrek's count, this is the third federal action against Tesla's driverless fleet in about five weeks. Electrek also argues that the 1,000 figure may come from Texas DMV registrations and that not that many vehicles may actually be running; last week, 319 Cybercabs were registered in Texas. Source: Electrek source

Shanghai: 8,000 humanoid robots in factories by 2028 · humanoid

At the CIIF opening ceremony, Shanghai launched its "Action to Accelerate Next-Generation Intelligent Manufacturing," comprising four major projects and 19 key tasks. The targets for 2028 are an industrial robot density of 600 per 10,000 workers, pushing 8,000 humanoid robots into factories, cultivating 500 advanced-level smart factories, more than 2,000 AI application scenarios in manufacturing, and smart-factory coverage of all key manufacturers. Shanghai's current smart manufacturing equipment output exceeds 150 billion yuan, and the action calls for the AI-enabled new smart manufacturing equipment industry to exceed 200 billion yuan by 2028. This is another provincial-level 2028 humanoid target after Guangdong set annual humanoid production of 80,000 units for 2028; Shanghai's figure counts units deployed in factories. Source: 21jingji.com source

FCC stops certifying new models of foreign-made mobile robots; Rainbow Robotics becomes first Korean company to seek an exemption · humanoid

On July 28, the US FCC added advanced mobile robots produced abroad to its "Covered List"; new models of humanoids, quadrupeds and AMRs cannot obtain equipment authorization, effectively barring their sale in the US. Previously certified older models can still be sold, and fixed industrial robot arms are outside the scope. There are two exemption routes. One is production in the US with US components exceeding 65% of cost, rising to 75% from 2029. The other is a determination by the US Department of War that the product poses no national security risk, yielding conditional approval, with applications due by January 1, 2028. The rule is applied by place of production rather than company nationality; Rainbow Robotics, a Samsung Electronics affiliate, had its mobile humanoid RB-Y1 exports blocked and filed a conditional-approval application in late September, the first Korean company to do so. According to the office of Korean lawmaker Seo Ji-young, about 20 Korean robot companies had planned to export to the US before 2027, and about half have already signed export contracts directly. Source: Seoul Economic Daily source

Frost & Sullivan: the top five humanoid makers by first-half revenue are all Chinese companies, and two sources' shipment counts differ by 11,000 units · humanoid

A Frost & Sullivan report says global humanoid robot sales in the first half of 2026 were 36,000 units with revenue of $540 million, both already above full-year 2025. By revenue, UBTech (Chinese humanoid maker) ranks first with 620 million yuan and 16,123 units sold; Zhiyuan Robotics (AgiBot) has 610 million yuan and 7,850 units; Unitree has 580 million yuan and 4,900 units; Galaxy General (Chinese embodied-AI startup) and Leju Robotics (Chinese humanoid maker) rank fourth and fifth. Full-size embodied-AI humanoids sold only about 6,000 units in the first half, roughly 17% of unit sales by the report's calculation, with revenue of $320 million, nearly 60% of the total. UBTech sold 921 units in this category, with revenue of 590 million yuan, the bulk of its first-half total revenue. Two other sets of figures do not match the report. Economic Reference News, citing IDC, says global first-half shipments were close to 25,000 units, up 432.1% year on year, and the same outlet says Zhiyuan shipped more than 8,600 units in the first half, versus the 7,850 that Frost & Sullivan gives for Zhiyuan. Source: Sina Finance via China Business Network source; Economic Reference News source

London Assembly report: TfL should have powers to cap Robotaxi numbers and revoke licences · autonomy

The London Assembly's transport committee has published a 66-page report with 13 recommendations. The committee says that under current plans it has not received sufficient evidence to support putting commercial driverless passenger services on the road. At present, passenger-carrying permits are issued by central government and TfL has only a consent role; the report calls for powers to issue and revoke permits and to cap numbers by area to be devolved to TfL when the Automated Vehicles Act is fully implemented in 2027. The report also asks operators to report collisions and near misses within 24 hours, requires remote assistance staff to be based in the UK and background-checked, and lists empty-running mileage as a core reporting metric. Source: TaxiPoint source

Werner hauls freight for driverless trucks; its CFO says insurance and liability issues must be solved before scaling · autonomy

US carrier Werner has piloted with Aurora and Kodiak since 2022 and now brokers freight to autonomous-driving companies, but is uncertain about the timing and scale of deployment. CFO Christopher Wikoff said the insurance issue must be resolved first. He also said: "We've been very clear with our driver teams that even if the most optimistic autonomous scenario comes true, we will need more drivers next year and ten years from now than we have today." Source: Trucking Dive source

UBTech and C&D Inc. form a joint venture, targeting at least 10,000 units sold by 2030 · humanoid ⚠️ Plan, not outcome

UBTech and C&D Inc. (Chinese state-linked supply chain conglomerate) have signed to build a joint venture, planning to complete sales of no fewer than 10,000 UBTech humanoid robots in cooperation by 2030. The robots will first go into C&D's own supply chain operations and real estate business for logistics warehousing, security patrols and similar scenarios, and the two sides will also explore supply chain finance solutions. C&D's 2025 revenue was 671.27 billion yuan. Source: Gasgoo source; Global Times source

Siemens opens 5 factories in China to embodied-AI testing · industrial

At CIIF, Siemens announced that it is opening 6 types of real-world scenarios at 5 factories in China for embodied-AI partners to test and validate, including unordered bulk-parts picking and feeding, material container handling, depalletizing and palletizing, high-precision screw tightening, and finished-product packaging. At the same time, it jointly released the national standard GB/T 47859—2026 "Reconfigurable Modeling Rules for Smart Factory Production Systems" and launched more than 20 new products, 13 of which were developed in China. Source: PR Newswire source

Wang Xingxing: GPT-6 is already being tested on robots, and the tipping point for embodied AI could come within one to two years · humanoid ⚠️ Speech, not verified

Unitree Technology founder Wang Xingxing said in a conversation segment at the CIIF opening ceremony that the team has deployed GPT-6 on robots for task testing, with "pretty good results," and that next year's GPT-7 "may directly crack the embodied-intelligence problem." He also admitted that what Unitree and its peers are doing in factories is still pilot work, and that large-scale application must first solve generalization and efficiency. He said: "The biggest mistake I made over the past decade or so was not believing in AI enough." Source: Guandian.cn source

A3 completes its mobile robot safety standards; Part 3 covers on-site users · industrial

ANSI/A3 R15.08-3-2026 has been published, completing all three parts of the R15.08 industrial mobile robot safety series. Part 3 sets out responsibilities at the site of use: risk assessment, change management, employee training and periodic inspection. Robot manufacturers are responsible for specifying the operating environment assumed in the design, and when the site environment changes, the operator is responsible for controlling it. Source: MarketScale source