Today's Highlights

· Amazon invests over $100 million in Indiana to build a fourth robotics factory; its US production lines have made over a million units cumulatively

· Waymo's registered driverless vehicles in Texas grew 49% in three weeks, reaching 1,102

· Uber, WeRide (Chinese autonomous driving company), and AVOMO obtain Spain's first nationwide L4 passenger permit, service to launch in Madrid

· EU vote on Tesla FSD (Supervised) postponed from October, to be revisited December at the earliest

· IFR: Chinese-made robot manufacturers took 55% of China's industrial robot installations in 2025, surpassing half for the first time

· GPT-6 Astra navigates zero-shot using only monocular RGB, achieving 79.0% success rate on R2R-CE, surpassing the previous best supervised result

GPT-6 Astra Zero-Shot Vision-Language Navigation Evaluation · autonomy

Zero-shot success rate on R2R-CE reached 79.0%, 13.0 percentage points above the previous strongest zero-shot result and 6.9 points above the strongest supervised result. In a Codex harness, the authors gave the model only monocular RGB, with no navigation fine-tuning, no waypoint predictor, and no pre-built map — the model decides on its own when to observe, how to move, and when to stop. The paper also notes shortcomings: matching local landmarks correctly doesn't guarantee reaching the correct endpoint, and route execution and goal confirmation remain unreliable. The same model, used directly as a manipulation policy on RoboDojo this month, achieved only a 22.48% average success rate.

Guangzhao Dai et al. · arXiv 2609.29861 source

Coding Agents Solve General Task and Motion Planning · manipulation

HF 5↑. Across 16 environments with comparable off-the-shelf planners, TAMP programs written by coding agents achieved average success rates of 56% to 95%, versus 47% for hand-built planners (TAMP stands for task and motion planning, solving discrete task decisions together with geometric motion constraints). Claude Code (Opus 5) and Codex (GPT-5.6 Sol, GPT-6 Astra) iteratively wrote and tested programs across 28 simulated environments in KinDER and PDDLStream; once frozen, the programs were tested on 100 unseen instances, totaling 98,000 trials. As object counts increased, the agent-written programs still outperformed the planners, using on average an order of magnitude less compute per instance. Code and full prompts are public.

Matteo Merler et al. · arXiv 2609.30233 source

World Action Agent: Letting VLMs Rehearse Before Acting in a Visual Action Workspace · vla

HF 3↑. Average success rate on LIBERO-Pro was 75.6%, surpassing end-to-end VLAs, code-as-policy agents, and a visual harness baseline built on the same backbone. Each action from a general-purpose VLM first becomes an editable proposal, which is rehearsed and corrected before execution. Its interaction trajectories can also be used to train smaller models: after fine-tuning, Qwen3.5-9B's out-of-domain success rate rose from 1.7% to 43.3%.

Yehang Zhang et al. · arXiv 2609.29964 source

RAPID: Watch One Human Demonstration, Auto-Write and Debug the Robot Program · manipulation

Authors include Leslie Pack Kaelbling and Tomás Lozano-Pérez. RAPID derives three things from a single human visual demonstration: a testable task specification, action primitives, and an interactive execution environment; a coding agent then repeatedly generates, verifies, and revises the program. The program uses an object-centric relational representation, so it still works when object poses, shapes, or materials change. Real-robot deployment on a Franka arm tested all 8 non-prehensile contact-rich tasks.

Yuyao Liu et al. · arXiv 2609.30249 source

Self-Adaptive VLA: Adapting to Hardware Drift On-Site at Deployment · vla

Authors include Chuang Gan. Wear and calibration error can cause VLA performance to drop in the field. This post-training method compresses the policy's rollout in a drifted environment into a single context token, which modulates policy output via AdaLN; multiple tokens can also be ensembled for round-by-round self-correction. Across four precision-sensitive bimanual and dexterous manipulation tasks, under execution bias and joint encoder drift, it recovers over 80% of the base policy's performance.

Hongxin Zhang et al. · arXiv 2609.30092 source

Res-HIL: Human-in-the-Loop Residual Reinforcement Learning · manipulation

Using only 20 initial demonstrations and ten minutes of online training, Res-HIL outperformed both full-policy human-in-the-loop RL and residual fine-tuning without human guidance across five contact-rich tasks, and also outperformed an imitation policy trained on five times as many demonstrations. The imitation policy stays frozen, with only a residual correction layer learned on top. Each time a human takes over, they both directly supervise the residual and retroactively shape the reward for the preceding autonomous actions.

Mariia Iavorskaia et al. · arXiv 2609.30023 source

Rolling-WAM: Spreading Joint Video-Action Denoising Across Multiple Replanning Cycles · world-model

A world action model (WAM, i.e. a model that predicts future frames and actions jointly) normally runs a full round of joint denoising on every replan, which drags down control response latency. Rolling-WAM spreads this round of denoising across several consecutive cycles, speeding up steady-state replanning by 4.5x.

Yinghua Zhou et al. · arXiv 2609.30247 source

Real-Time Force Modulation for Whole-Hand Grasping · manipulation

Authors include Sangbae Kim. This controller infers contact across all links of the hand using object tracking plus proprioception — no tactile sensing needed at these sites — and repeatedly recomputes contact force allocation under friction and actuator constraints. In real-robot experiments on a 27-DoF arm-hand system, it maintained grasps under human-applied perturbation and re-grasped after drops.

Sang Min Kim et al. · arXiv 2609.30082 source

Ego-Exo4D-HM: Adding 4D Human Motion Reconstruction to Ego-Exo4D · benchmark

Ego-Exo4D contains both first-person and multi-view external video, but originally came with only sparse 3D human pose annotations. This dataset adds large-scale 4D human motion reconstruction to that collection, and releases the reconstruction pipeline alongside it, for use in robotics research that learns skills from human video.

Abhiram Maddukuri et al. · arXiv 2609.30187 source

Newstart Precision (Zibo) | ChiNext IPO Under Inquiry | Targeting RMB 823 Million · hardware

Newstart Precision updated its prospectus; its review status is "under inquiry," with Orient Securities as sponsor. Per Frost & Sullivan data, by 2025 sales it ranks first among Chinese-brand precision planetary reducer makers in China, shipping over 700,000 units a year and ranking second globally. Revenue from embodied AI robot applications was RMB 96.6282 million in 2025, up 1235.46% year-on-year, and RMB 59.3891 million in the first half of this year. Zhiyuan Robotics (Chinese humanoid startup) was its largest customer in 2025, with sales of RMB 37.0215 million, 7.46% of core revenue. The sponsor's estimate puts its share of the humanoid robot reducer market at no less than 23%, assuming each humanoid needs 12 to 32 reducers, with 22 taken as the midpoint baseline. The IPO proceeds are mainly earmarked for a manufacturing base with annual capacity of 1.5 million precision planetary reducers and transmission system modules. Source: Ifeng Finance via Unicorn Insights source

Simate | Multiple Consecutive Rounds | Several Hundred Million RMB · embodied ⚠️ company claim

This physical AI startup, only a few months old, has completed several hundred million RMB across multiple consecutive funding rounds. Founder and CEO Zhang Ying was previously the core technology lead at a major autonomous driving company; the team also includes HKUST assistant professor Zhan Fangneng. The company claims to have topped the RoboDojo leaderboard, surpassing GPT-6 Astra and DeepMind, and says its in-house AutoResearch system lets AI agents autonomously propose hypotheses and run experiments. The report did not give Simate's actual success rate on RoboDojo. Source: Sina Finance via AI Era source

Feather Robotics | First Round | $7.6 Million · humanoid

Led by Gradient Ventures. The two outlets label the round differently: TechCrunch calls it pre-seed, while Dealroom's headline calls it seed. Feather makes a modular developer humanoid priced at $29,990, which TechCrunch says is roughly half the price of the Unitree H2 Edu; the robots are already cooking in restaurants and cleaning labs, generating over $1 million in revenue. Co-founder Hoa Mai said: "Today you can't buy a Tesla robot to develop on."Source: Dealroom source; Yahoo Finance via TechCrunch source

DeepCtrls | Strategic Investment | Amount Undisclosed · adjacent

CATL and Aramco Ventures invested in DeepCtrls, a physical-world AI company; the round and amount were undisclosed. Source: ANTARA News source

Konnex | New Round | $18.5 Million · adjacent ⚠️ single-party claim

Konnex says it aims to build the "financial layer" for physical AI, with mainnet launch as its next step; it positions itself as crypto-financial infrastructure. Source: Chainwire source

Kodiak Sets Its First Long-Haul Driverless Route: Dallas to Houston · autonomy ⚠️ company claim

The route runs 219 miles from the Lancaster, Texas hub to Houston, with the main leg on I-45 and city streets at both ends. Kodiak says that since launching its driverless-operations rollout in August, this end-to-end delivery run has been consistently completed without human intervention, and that its safety observer hasn't touched the wheel even on the city-street segments. The company's target is to complete safety validation by year-end and begin unsupervised highway operations; as of late August, its Autonomous Readiness Measure (ARM) for long-haul scenarios was 93%. An observer still rides in the truck, so these runs count as human-supervised deliveries. The same week, Waabi COO Lior Ron said its system could be deployed "today," pending only Volvo's final truck validation by year-end, with a Texas launch set for early 2027. Source: Stock Titan source; Kodiak press release; BetaKit source

Waymo's Texas Fleet Grows 49% in Three Weeks · autonomy

As of September 24, Waymo had 1,102 registered autonomous vehicles in Texas. Over the entire summer, that number only rose from roughly 600 to just over 700, then grew 49% in the last three weeks alone. Per TechCrunch's analysis of state registration data, roughly 80% of Waymo's roughly 4,000 robotaxis nationwide operate in California and Texas, delivering about 500,000 paid trips per week; most of the newly added vehicles are Zeekr RT-based Ojai conversions. Registration counts don't equal vehicles actually in service: as reported on the 22nd, Tesla's Austin fleet has 67 registered Cybercabs, but only 8 robotaxis actually carried passengers in the past 7 days. Internationally, Waymo plans to launch a driverless paid service in Tokyo in 2027 with GO and Nihon Kotsu, starting with a small fleet and scaling to roughly 100 vehicles, with the timeline dependent on national and local approvals. Source: TechCrunch source; Yahoo Finance source

Uber, WeRide, and AVOMO Win Spain's First Nationwide L4 Passenger Permit · autonomy

The permit covers commercial robotaxi operations in Madrid, clearing the regulatory barrier to paid rides. Under the arrangement, WeRide provides the technology, local operator AVOMO manages the fleet, and Uber integrates the service into its own app. A launch date has not been disclosed. Source: Yahoo Finance source

Pony.ai Interim Report: Robotaxi Revenue Up 534%, Losses Widen · autonomy

First-half robotaxi revenue was $20.643 million, up 534.0% year-on-year; Q2 passenger fare revenue rose 849.3% year-on-year. As of June 30, the global fleet numbered 1,975 vehicles, and the company maintains its year-end guidance of over 3,500 vehicles across more than 20 operating cities. Total revenue was $70.47 million, with a net loss of $98.861 million, 9.1% wider than the same period last year; operating cash outflow was $118.2 million. These financial figures were already disclosed with the August 18 earnings announcement; the interim report published September 23 filled in operational details. Source: Autohome source

Over 300 Robots Deployed at Chimelong Spaceman Park · humanoid

Following Zhiyuan Robotics' 20,000th robot delivered to Chimelong in Hengqin, over 300 robots went into service on the opening day of Spaceman Park, handling tour guiding, tea ceremony demonstrations, rock-paper-scissors, and table tennis. Accounts of the human-to-robot ratio differ: a National Business Daily reporter observed one staff member stationed beside each robot; Zhiyuan co-president Xiong Yan said "the ratio of robots to people deployed today is roughly 3:1," with a target of 10:1. Zhiyuan CEO Deng Taihua said the robot fleet at Chimelong Hengqin will expand to 1,000 units for phase two by year-end. Chimelong Group vice president Jiang Minling said the robots "still have some shortcomings" in outdoor settings and rain. Source: Sohu via National Business Daily source

SenseTime Shanhui Robot Kiosks: 0.2 to 0.3 Staff Per Store · embodied ⚠️ company claim

SenseTime's "Shaomai Gou" robot kiosk concept has opened over 20 locations across Shanghai, Hefei, Shenzhen, Qingdao, and Yancheng. SenseTime Shanhui says the average staffing per store is 0.2 to 0.3 people, versus about 2 people per shift at a traditional convenience store; well-located Shanghai sites pay back in roughly six months to a year. On September 23 it launched SenseMart OS, a retail physical operating system, aiming to sell the solution to unmanned-retail operators and to reach around a hundred self-operated stores by year-end. Its after-sales engineering team currently numbers only about a dozen people; co-founder Yi Shuai said people capable of robot after-sales service, maintenance, and quality inspection are "still fairly scarce." Source: Sina Finance source

Amazon Builds Fourth Robotics Factory in Indiana · industrial

Amazon will invest over $100 million in Greenwood, Indiana, to build a plant producing warehouse robotics equipment — its fourth robotics manufacturing site. Before August it had only two, both in Massachusetts; in August it added one in Austin, Texas, which the company says involved investment in the billions of dollars. The Greenwood facility spans 585,000 square feet, handling parts machining, welding, powder coating, and final assembly on-site, with welding performed by robots; it's expected to employ about 300 people and come online by 2028. Amazon says its US production lines have produced over 1 million robots cumulatively, deployed across more than 300 facilities worldwide, helping process 75% of its customer orders. Last October, The New York Times reported that internal Amazon documents showed plans to avoid hiring more than 600,000 US employees by 2033 through automation; Amazon denied this at the time, saying the documents came from only "one team." Source: SiliconANGLE source; Magzter via The Wall Street Journal source

Tesla Optimus Scale-Up Bottlenecked by Hands and Suppliers · humanoid ⚠️ single-party claim

Following its supplier audit in Ningbo earlier this month, The Information reports that Optimus production has increased roughly tenfold in recent months but still struggles to reach stable mass production. The hand's complex mechanical structure is the main bottleneck, with supplier capacity and automation equipment failures also causing delays; the year-end target of 1,000 units per week is under pressure. Tesla's Q1 update stated that the first-generation Optimus line at Fremont is designed for annual capacity of 1 million units, while noting that installed capacity does not equal actual output, which depends on factors including parts supply. The Q2 update only said production was "expected to start later this year," with no date, quantity, or price given. Source: The Information source; Yahoo Finance source

IFR: Chinese-Made Manufacturers Take Majority Share of China Installations for First Time · industrial

Following IFR's announcement yesterday that global operational industrial robots reached 5 million units, segment data from the same World Robotics 2026 report shows: in 2025, Chinese-made manufacturers installed 195,000 units in China, up 15% year-on-year, accounting for 55% of the Chinese market, up from about 30% five years ago. Foreign suppliers' installations in China rose 27% year-on-year, a faster growth rate than Chinese-made manufacturers. The EU's full-year installations totaled 60,500 units, down 11% year-on-year — less than a sixth of China's 354,000. Source: Tech Times source

EU Postpones Tesla FSD Vote · autonomy

On the agenda for the EU Technical Committee - Motor Vehicles meeting on October 6, the Netherlands' Article 39 authorization application is listed only as "for continued discussion." The EU-wide vote to approve Tesla FSD (Supervised) has therefore been postponed to December at the earliest. Approval requires a qualified majority of at least 15 of the 27 member states, representing 65% of the population. Sweden and other countries are concerned about the system exceeding speed limits; countries that have already approved it maintain that responsibility for complying with traffic rules still rests with the driver. Source: Investing.com source; Reuters

UBTECH's Almaty Plant: $340 Million of $440 Million Is a Technology Transfer Fee · humanoid ⚠️ single-party claim

Kazakhstan's NERO Group and UBTECH plan to jointly build a robotics factory in the Almaty Technology Park special economic zone, with total investment of $440 million. Of that, $340 million is a technology transfer fee paid to UBTECH for humanoid robot manufacturing technology, with the remaining $100 million being NERO's own investment. NERO CEO Yerlan Nabiyev said at the Kazakhstan-China Investment Forum that localized production of service and humanoid robots is set to begin in Q2 2027, with planned annual capacity of 1,000 units, and that around 60 schools in Almaty will receive educational robots this year. The two sides signed the agreement in July, marking UBTECH's first overseas factory. Kazinform reported on September 25 that Kazakh President Tokayev met with UBTECH senior vice president Zhong Yong. Domestically, UBTECH's Beijing-Tianjin-Hebei embodied AI robot smart manufacturing service base broke ground in Jinghai, Tianjin on September 22. Source: Kursiv Media source; Kazinform source; CNR source

Two Quadruped Robot Reports: Both Rank Unitree First, but Differ on Everything Else · adjacent

Counterpoint and IDC released global quadruped robot reports on the same day, September 21. By shipment volume, Counterpoint ranks Zhiyuan's Kuotuo second, with a 16.3% share; by hardware revenue, IDC ranks DEEP Robotics second, with a 9.2% share. The totals don't match either: for shipments in the first half of this year, Counterpoint counted close to 35,000 units, while IDC counted over 49,000. The revenue figures diverge in opposite directions: Counterpoint calculates industry revenue of nearly $700 million for 2025, while IDC puts it at only $490 million, because Counterpoint's figure includes sensors, scheduling software, and maintenance services, while IDC's counts only the hardware itself. The ranking gap comes from average selling price: Unitree's 2025 average quadruped price was RMB 30,300, while DEEP Robotics' Jueying X unit price is RMB 287,500. Source: cyzone.cn via AIX Finance source

Galbot Psi-R2.5: Generating Human-Hand Data in Reverse for Human-Robot Alignment · embodied ⚠️ company claim

On a phone-case assembly task, Galbot (Chinese embodied-AI startup) says using post-training that combines human involvement with reinforcement learning, a few iterations brought the success rate to about 99%, taking 1 to 2 working days, with the approach already deployed at customer sites. On the data side, it runs its world model Psi-W0 in reverse: starting from real robot data, it generates corresponding human-hand manipulation videos, producing frame-aligned "strongly paired human-robot data," which it then uses to train a converter that turns human manipulation into robot imagery and actions. Galbot also notes that some especially complex tasks still aren't directly solved, and that in-context learning has only been demonstrated on specific tasks. Source: BAAI Community source

Mifeng Pai Launches: Ordinary People Wear a Device to Collect Data, Paid by Valid Duration · adjacent ⚠️ company claim

Mifeng Technology launched its physical AI data crowdsourcing platform Mifeng Pai on September 23. Users take on tasks, wear a MEgo device, and perform specified actions in retail, warehousing, manufacturing, and home settings, getting paid based on verified valid duration. In one month of internal testing, 20,000 users registered and submitted 13,000 task entries. Chairman and CEO Yao Maoqing said "clothes-folding data is already oversaturated" — clients are now asking for data on specialized skills like plumbing and auto repair. He said over 95% of the data collected passes verification for use, and meeting a demand of 10 million hours of data would require several hundred million RMB in fixed-asset investment. Source: Yangtse Evening News via Red Star News source

Wang Xingxing: The Bottleneck Is a Few Millimeters of Error · humanoid

Unitree founder Wang Xingxing said at the Hangzhou Digital Trade Expo on September 24 that getting a robot to follow a command and nominally complete a task was already achievable last year; the biggest problem now is that the match between the model's input/output and the physical world isn't precise enough, leaving errors of a few millimeters in the work. "Whoever solves this problem in the future will have completely solved the robot problem." He defined the industry's "ChatGPT moment" as robots completing 80% of tasks via voice command in 80% of unfamiliar scenarios. Source: Jiemian News source