Today's Highlights

· Boston Dynamics gives Atlas a new 13-degree-of-freedom four-finger hand, dropping the pinky, and it can use a drill, an electric wrench and a welding torch

· California signs SB 1246: robotaxi operators will be fined if a vehicle blocks police or fire crews for more than 30 minutes

· Qiyuan (Chinese personal-robot maker) delivers the first Q1 and T1 units in Shanghai, Shenzhen and Hangzhou on October 1

· The RPG framework raises success on 22 manipulation tasks from 28.6% to 95.0% through simulated practice, without changing model weights

· Bloomberg Intelligence estimates that Pony.ai (Chinese robotaxi company) and WeRide (Chinese robotaxi company) have lower fleet utilization than Waymo and will still be loss-making through 2028

RPG: Robots practice on their own in a reconstructed simulation, with no weight updates · manipulation

Writing skills, designing rewards and tuning perception have traditionally required humans. RPG hands these tasks to a multimodal LLM: it first identifies the manipulation abilities to practice from an offline dataset, builds matching practice tasks in simulation, then adds or modifies symbolic skills based on failure diagnosis and rewrites the system prompt. Tested on 22 manipulation tasks with changed initial states, success rose from 28.6% after the first round of practice to 95.0% after 15 rounds, higher than ASPIRE (75.5%) and CaP-Agent0 driven by GPT-6 Astra Pro (60.0%). The frozen system succeeded in all 30 real-robot trials.

Yen-Jen Wang et al. · arXiv 2610.02204 source

PyRUA-Lean: GPT-6 Astra as a robot agent, using 65% fewer tokens · vla

HF 13↑. The agent writes classical robot primitives and VLA policies into Python cells for execution, with built-in conditional checks and local retries, and sends images back only when needed. Across 700 simulated task instances in LIBERO-PRO, RoboTwin 2.0 and RoboCasa365, with the same LLM-call budget, success rose from 63.1% to 71.7%; on instances both agents completed, the agent made 49% fewer LLM calls and used 65% fewer input tokens.

Ruiyang Si et al. · arXiv 2610.01939 source

InterEvolve: Evolving reward programs at test time so humanoids can do untrained tasks · locomotion

HF 11↑. An LLM agent rewrites staged reward programs based on execution feedback, a numerical optimizer tunes the constants inside them, and the result is handed to a forward-backward behavior foundation model with an object residual for execution, with no retraining throughout. The evolved skills run autonomously on a real Unitree G1 using onboard egocentric perception.

Zhuo Lin et al. · arXiv 2610.02196 source

Recova: Turning manipulation failures into reusable recovery skills · manipulation

Task execution and failure recovery are first practiced together in a reconstructed digital twin. At deployment, the agent monitors progress and invokes recovery policies; when no suitable recovery exists, a human demonstrates once and the demonstration re-enters the learning loop. After data collection in parallel across four real-robot stations, DAgger fine-tuning raised average success from 23.8% to 77.5%, and adding recovery skills brought it to 87.5%. On one of the tasks, the share of human interventions fell from 87.5% to 0 over four rounds of collection.

Isabella Liu et al. · arXiv 2610.01178 source

λ0: Pretraining humanoid loco-manipulation on 500 hours of human whole-body data · vla

The team used lightweight wearable devices to collect HumanVerse-500, totaling 500 hours, with egocentric video recorded in sync with body and hand motion. Training has three stages: first learning interaction from egocentric data, then learning body-hand coordination from HumanVerse-500, and finally adapting to specific robots. λ0 self-reports SOTA on SIMPLE and 4 real-robot tasks, and promises to release code, models and data.

Chongyang Xu et al. · arXiv 2610.00438 source

ECoMEM: Using explicit "concept memory" to fix VLA forgetfulness · vla

Once an object leaves view, VLAs often forget it. In ECoMEM, a Writer maintains a table of "concept" records based on evidence, and a Reader converts the records into tokens fed to the VLA. On RoboMME's 16 tasks, it leads on 15. On two new real-robot tasks it reaches a success rate of 86.1%, versus only 8.6% for a VLA without memory. Stanford's Mac Schwager and Jiajun Wu are among the authors.

Yize Liu et al. · arXiv 2610.00801 source

UniWAM: Merging a reasoner, a world generator and an action predictor into one model · world-model

UniWAM writes low-level actions as natural language, with VQA data, human egocentric data and robot demonstrations each supervising different components. The authors self-report SOTA on multiple benchmarks and report a log-linear scaling law for mixed human and robot data training.

Jiayi Chen et al. · arXiv 2610.02054 source

MIKASA-Robo-VLA: 90 tasks built to test VLA memory · benchmark

Of the 90 tasks, 80 hide the cues that actions depend on, and the remaining 10 are controls. In 28 of the tasks, the information gap exceeds the widest window among the surveyed fixed-context VLAs, which is 16 frames. The dataset contains 22,500 trajectories, offered in both RLDS and LeRobotDataset v3 formats. The π0.5 baseline without history frames averages a success rate of only 0.211.

Egor Cherepanov et al. · arXiv 2610.00604 source

When a VLA's chain of thought is corrected, its actions may not be corrected with it · autonomy

TRUST is an offline-trained value model that predicts from a chain-of-thought prefix whether the final reasoning will be correct, and steers reasoning accordingly. On Nvidia's driving VLA Alpamayo 1.5, reasoning accuracy rose from 75.9% to 90.0%, and the collision rate on a hard subset fell by 30.4%. Applied to the manipulation VLA DeepThinkVLA, reasoning accuracy rose similarly, but closed-loop results on LIBERO-Plus were essentially unchanged.

Sathwik Karnik et al. · arXiv 2610.00601 source

Kinematic MeanFlow: A robot foundation model that generates actions in one step · vla

Applying MeanFlow directly makes performance collapse. K-MF splits the time derivative across two sub-intervals and outperforms multi-step flow matching in most settings. As a result, the action-head latency of GR00T-N1.6 falls by 67.5%~74.4% and end-to-end latency by 30.3%~54.9%, with code to be released in the IntelChina-AI repository.

Jiawei Fan et al. · arXiv 2610.00864 source

Yinjia Technology | Hong Kong IPO | Listing application filed | Valuation about RMB 2.78 billion · autonomy

Shanghai Yinjia Electronic Technology (Chinese driver-assistance supplier) filed with the HKEX main board; its shareholders include Zhenghai Capital, Shenzhen Capital Group (Chinese state-backed VC), and Oriental Fortune Capital. Revenue in 2025 was RMB 481 million, versus only RMB 111 million in 2023; driver-assistance customers grew from 16 to 45, and unit sales rose from 130,000 to 416,000. By Frost & Sullivan's 2025 revenue measure, it ranks ninth among suppliers of driver-assistance and visual-safety solutions in China. Its general-purpose robot perception and localization solution is still at the customer-validation stage, before any supplier nomination. Source: Securities Times source

Extend Robotics (UK) | Early-stage round | £2.6 million · embodied

Led by Skyworks Venture Capital Fund, with participation from Neo Venture (Europe) and Zip Capital. The company builds AMAS, a teleoperation platform, and does not sell robots, instead charging customers by completed output. It has 35 subscription customers, revenue has doubled year over year for three years running, and it already uses humanoid robots for quality inspection on an Apple packaging line. Source: Ventureburn source; Dealroom source

FORT Robotics | SPAC merger | S-4 confidentially filed · adjacent

FORT announced its SPAC merger listing plan on October 1 (previously reported) and has since confidentially filed its S-4 registration statement. Source: AI Insider source

First Qiyuan Q1 and T1 units delivered to customers · humanoid

In September we reported that Qiyuan's two personal robots went on sale from RMB 19,999. On October 1, the first batch was delivered at an authorized store on Huaihai Middle Road in Shanghai, with deliveries in Shenzhen and Hangzhou at the same time and other stores to follow. Authorized experience stores now cover 7 cities. The base Q1 does not open its SDK; only the RMB 26,999 Explorer edition has a secondary-development interface. Source: TOM Tech source; Pandaily source; Huxiu source

Zhuhai Chimelong robot park opens with 300-plus humanoid robots on duty · humanoid

In late September, AgiBot's (Chinese humanoid robot maker) 20,000th robot rolled off the line and was delivered at Chimelong in Hengqin (previously reported), and these robots have now entered the park: 300-plus humanoid robots handle check-in, guidance, performances and playing table tennis with visitors. According to CGTN, there is currently 1 operator for every 3 robots, and AgiBot's goal is 1 person overseeing 10. The 1:3 human-to-robot ratio is the real cost of current operations; 1:10 is the target. ⚠️ Vendor-reportedSource: travelsdubai citing CGTN source; Qazinform source

Orion Star (Chinese service-robot maker) has supplied 4,000 units in South Korea in total · adjacent

Orion Star entered South Korea in 2022 through Woowa Brothers, operator of the "Baedal Minjok" delivery app, and has supplied 4,000 units in total over about 4 years. It was initially mostly delivery robots and has now expanded into guide, cleaning and logistics robots, placed in science museums, schools and hospitals. A pizza restaurant in Busan has deployed 5 delivery robots on each of its two floors. Source: Asia Economy source

Tesla's Texas Cybercab registrations reach 158 vehicles · autonomy

126 vehicles were reported on September 26, and the latest batch adds 32 more, versus only 69 on September 23. The figures come from Texas DMV records compiled by a blogger; registration only means a vehicle has completed plating formalities, not that it is carrying passengers on the road. ⚠️ Compiled figuresSource: BASENOR source

Barrett plans to deploy the UNIT AI platform across warehouses from 2027 · industrial

US third-party logistics provider Barrett is expanding its partnership with UNIT AI, and from 2027 will use the latter's platform to coordinate inventory, fulfillment and returns across multiple warehouses, paying under a "warehousing-as-a-service" model. ⚠️ Planned figuresSource: PR Newswire source

Atlas gets a four-finger hand: 13 degrees of freedom, pinky removed · humanoid

Boston Dynamics has released a new-generation hand for Atlas: three fingers with 3 degrees of freedom each and an opposable thumb with 4, for a total of 13 degrees of freedom. The thumb can slide along both the length and width of the other fingers, and the fingertips and palm are covered with pressure-sensing tactile sensors. The previous 7-degree-of-freedom hand was mainly for grasping; the new hand can rotate objects in-hand, catch slipping items, and operate an electric drill, an electric torque wrench, an angle grinder, a nail gun and a welding torch. One fewer finger saves 3 actuators, and cost, size and failure points all fall accordingly. CTO Zack Jackowski once had team members tape their pinky and ring fingers together for a day and note what they could not do, concluding that four fingers were enough. The actuators are fully enclosed and use a single model, with no cables running through the joints, which the company says makes mass production and repair easier. Source: Robotics & Automation News source; Interesting Engineering source; The Robot Report source

California's SB 1246 takes effect: robotaxis blocking emergency crews for over 30 minutes will be fined · autonomy

Governor Newsom signed SB 1246. When a robotaxi malfunctions, the operator must provide local on-site support to emergency responders; blocking police or fire crews for more than 30 minutes will be penalized. Remote drivers must be located in the US and hold a US driver's license. The bill's author, state Senator Dave Cortese, said: "When autonomous vehicles crash, break down, block the road in an emergency or interfere with law enforcement and emergency response, there must be clear accountability." According to TechCrunch, Waymo's US commercial fleet is about 4,000 vehicles, about 1,200 of them in the San Francisco Bay Area, and most earlier incidents involved Waymo. Source: TechCrunch source; Wangtong News source

Bloomberg Intelligence: Chinese robotaxi fleet utilization lags Waymo · autonomy

Bloomberg Intelligence estimates that Pony.ai and WeRide have lower fleet utilization than Waymo, and expects both companies to remain loss-making through 2028, with fleets possibly reaching the scale needed for profitability only in the late 2020s. The report also notes that Waymo's utilization did not rise as its fleet expanded, but rose markedly only after average trip distance lengthened. Insufficient fleet density makes the service less convenient for riders, and higher operating costs will continue to weigh on margins. ⚠️ Estimated figuresSource: KLSE Screener citing Bloomberg Intelligence source; Bloomberg source

Skild AI releases S1: learns a new task by watching a video · world-model

S1 understands tasks using video as a prompt, with no weight changes and no task-specific post-training, and can handle unseen multi-step tasks of up to about 10 minutes. The company says S1's per-step success rate on new multi-step tasks is about 66%, versus 9% for comparable systems, and that a short video is worth about 380 human demonstrations. The company also says annualized revenue reached $100 million 10 months after its first commercial deployment, with more than 60 deployment partnerships. Co-founder and CEO Deepak Pathak said, "Learning from experience instead of being preprogrammed is the leap that has already happened in robotics." ⚠️ Vendor-reportedSource: IT Brief New Zealand source

Shenzhen Nanshan intelligent robot park named to China's first batch of national demonstration base candidates · adjacent

China's Ministry of Industry and Information Technology (MIIT) published its first list of candidates for national emerging-industry development demonstration bases, and the intelligent robot demonstration park in the Nanshan Park of Shenzhen High-tech Zone was selected; the creation period is 2 years, with assessment and acceptance afterward. Nanshan's "Robot Valley" brings together more than 200 industry-chain companies and 32 specialized "little giant" firms (government-designated niche champions), and within its roughly 28 square kilometers, 90% of components can be delivered within half an hour. Source: Sina Finance citing Southern Plus source; Sohu source

Bolt sets a target of 25,000 Lucid robotaxis · autonomy

Bolt plans to deploy at least 25,000 Lucid driverless vehicles in Europe, reaching 100,000 on its platform before 2035. These are all still targets, with no firm orders. Lucid's mid-size platform has been delayed to the second half of 2027, and under EU type-approval rules, a single driverless model can register only 1,500 vehicles a year. ⚠️ Planned figuresSource: Fleet Europe source; The Auto Wire

Galaxy General's (Chinese embodied-AI startup) bipedal humanoid ET1 priced from RMB 79,000 · humanoid

The ET1 stands 1.25 meters tall and comes in three tiers: the standard edition at RMB 79,000, the Pro edition at RMB 139,000, and the flagship edition at RMB 179,000. The Pro edition adds a 40 TOPS Orin Nano module and supports "learn motions by watching once". The company says that after opening pre-orders in early September, it received more than 200 within 24 hours. ⚠️ Vendor-reportedSource: Pandaily source

US Army's xTechHumanoid competition selects 5 winners · humanoid

The competition received 103 white papers, and 10 teams reached the finals, 9 of which gave demonstrations at Marine Corps Base Quantico in September. Army Chief Technology Officer Alex Miller said that using xTech to survey the industry is meant to move promising technology forward quickly, "rather than getting bogged down in multi-year research". Source: AI Insider source