Today's Highlights

· May Mobility goes public via SPAC, enterprise value $1.4 billion, 2025 revenue about $10 million

· NHTSA gives Tesla until September 30 to explain how Cybercab passed federal safety certification

· Tesla Robotaxi launches in western Miami; Mashable says the cars have no in-vehicle safety monitor

· South Korea releases five-year physical AI R&D roadmap, two regional projects total 1.4131 trillion won

· A dual-arm robot completes 8 cases of autonomous kidney tumor resection on phantoms, all 77 cuts with negative margins

· The Zhijie RX (Huawei-backed EV brand) obtains an L3-level autonomous driving road test license, already tested on urban expressways

Occupancy-Network-Guided Autonomous Robotic Partial Nephrectomy · perception

Autonomous soft-tissue tumor surgery has previously been limited to organ surfaces, since once tissue is cut or deformed, the system loses track of the anatomy. This work uses a conditional occupancy network trained purely in simulation to infer the complete 3D shape of the kidney, tumor, and margin tissue from a single-view partial point cloud, tracking continuously during cutting. On patient-derived hydrogel phantoms, a dual-arm robot completed 8 cases of autonomous tumor resection, 77 electrocautery cuts in total, all with negative margins, with an average absolute margin error of 1.61±0.48 mm. Results are currently limited to phantom experiments in an open-surgery setting.

Ethan Kilmer et al. · arXiv 2609.16186 source

XPACE: Jointly Learning World and Action from Heterogeneous Experience · world-model

XPACE was validated on XPeng's IRON humanoid robot, with a single model serving as both policy and simulator: a shared video backbone jointly predicts actions and future video, and also predicts the visual consequences of a given action. Action-unlabeled video is used to learn visual dynamics, while human and robot demonstrations are used to learn actions. The simulator then synthesizes trajectories around expert demonstrations that deviate and recover, feeding this data back to fine-tune the policy. The paper reports that heterogeneous training lets the robot learn human skills not present in the demonstrations, and that the self-generated recovery data further improves real-robot task completion.

Jiacheng Wei et al. · arXiv 2609.17372 source

ModAR: A Modality-Autoregressive World-Action Model · world-model

HF ↑4. World-action models (WAMs — models that jointly predict future observations and actions) typically represent the future using RGB images; ModAR instead sequentially denoises point trajectories, DINO features, and depth, generating actions last. Experiments found no stable benefit from additionally predicting RGB. ModAR trained from scratch achieves an average success rate of 75%, slightly above video-model-initialized Flex-π (72%), while using roughly 20x fewer training FLOPs.

Adam Hung et al. · arXiv 2609.17524 source

Re-running ACT's CVAE Ablation: The Original Paper's Large Drop Doesn't Reproduce · manipulation

The original ACT paper reported that removing the CVAE encoder caused the average success rate on two simulation tasks to fall from 35% to 2%. The author re-ran this using the original code and did not reproduce this drop. Training duration and checkpoint selection can flip which of the two policies performs better, and at inference time ACT already zeroes out this latent variable. Code and evaluation tools are public, and since many manipulation policies use ACT as a baseline, this reproduction result is directly useful to them.

Bo Kang · arXiv 2609.16745 source

TEMPO: Adding Temporal Context to VLAs for Dynamic Manipulation · vla

VLAs only see a single frame, so they cannot judge where an object is moving, nor distinguish between states that look similar but are actually at different stages of a task; the authors argue that a bigger model or lower latency alone cannot fix this. TEMPO leaves the backbone unchanged and adds two inputs: a motion summary extracted by a frozen video foundation model, and a segment of proprioceptive history. Success rate on a bottle hand-off task rose from 44% to 74%, alongside the release of TEMPO-Bench, with over 50,000 annotated frames.

Zhenyang Feng et al. · arXiv 2609.16864 source

Weave: Learning Whole-Body Dexterous Humanoid Loco-Manipulation from Human-Object Interaction Data · locomotion

A single policy simultaneously controls 29 body joints and 12 finger joints; human motion-capture data is first retargeted with contact awareness before training. Across nine objects, the success rate on trained interactions was 92.5%, while unseen sequences not included in training still reached 65.0%. The authors also release about 9,000 trajectories from real physical execution, roughly 23 hours in total, with contact annotations.

Liu Cao et al. · arXiv 2609.16683 source

Robot Data Factories: Turning Data Collection into Continuous Scientific Production · benchmark

Sami Haddadin, Ivan Laptev, Ian Reid, and others jointly propose that the core resource for physical AI is robot experience that preserves the "perception, action, consequence" chain — raw data alone is not enough. The framework links training grounds, task curricula, external ground truth, and continuously updated benchmarks into a "deploy, measure, learn" loop, and derives quantitative relationships between fleet scale, sensor frequency, and training compute; it has already been deployed at three physical training grounds — home, environment, and energy.

Sami Haddadin et al. · arXiv 2609.16705 source

AdaDE: Converting Dense VLAs to MoE and Deactivating Some Experts · vla

Onboard compute cannot keep up with increasingly large VLAs. AdaDE converts part of the FFN into MoE layers and dynamically deactivates experts based on routing statistics: after deactivating 40% of LLM parameters, average success rate on LIBERO remains at 95.1%, and average success rate across all 50 RobotWin2.0 tasks is 42.0%.

Muchun Niu et al. · arXiv 2609.16503 source

May Mobility ｜ SPAC merger and public listing ｜ up to $337 million raised ｜ $1.4 billion enterprise value · autonomy

May Mobility will merge with ACP Holdings Acquisition Corp., affiliated with Atlas Credit Partners, and list on Nasdaq under ticker "MAY." Total funds raised will reach up to $337 million: a $120 million PIPE has been fully subscribed, plus up to $217 million from the trust account, with the actual amount depending on shareholder redemptions; closing is expected by year-end. The company's 2025 revenue was about $10 million, with a 27% gross margin, full-year cash burn of about $93 million, and roughly $445 million raised cumulatively since its founding in 2017. The company is shifting toward an "autonomous-driving-as-a-service" model, with fleet partners bearing vehicle, depot, and operations costs while May collects licensing fees or per-ride payments; per Axios, its long-term goal is a gross margin of up to 70%. Operationally, May says it has completed over 550,000 commercial autonomous rides across the U.S. and Japan, with its partnership with Uber in Arlington set to launch in Q4 this year or Q1 next year. Revenue is two orders of magnitude below the $1.4 billion valuation, so the IPO narrative rests on the licensing model. Source: The Deep Dive source

Hello Robotaxi (Shanghai Zaofu Intelligent Technology) ｜ new funding round ｜ about $100 million ｜ post-money valuation near $3 billion · autonomy

Led by the Shanghai Guotou Leading Fund, with participation from Minjin Capital and Genesis Capital. The funds will go mainly toward R&D, focused on a large-scale compute cluster platform, iteration of end-to-end large models, and data supply-chain development. Source: Shanghai Observer source; Guandian.cn source

Robocurve ｜ Seed round ｜ $10 million · adjacent

Led by Initialized Capital, with participation from Notable Capital, Decasonic, Y Combinator, and Halcyon Futures, among others. Robocurve is registered as a public benefit corporation and positions itself as a third-party auditor, specializing in evaluating what frontier AI models can actually do when connected to real robots, saying it does not let the AI companies being tested influence the conclusions. The company says its research has found that on simple robotics tasks, large language models can outperform some VLA models. Source: AI Insider source

Turui Intelligent (Hangzhou) ｜ First round ｜ tens of millions of yuan · embodied

Joint investment from industrial capital including Dongli Weilai Technology and Tongli Technology, along with industry investors. Turui does embodied-robot ODM work while also supplying real-robot scenario data, and independently develops integrated joints, motion controllers, and data-acquisition training systems; the new funding will go toward a pilot production base and scaled-up production lines. Source: Gasgoo source

Shita Robotics (Shenzhen) ｜ Pre-A round ｜ amount undisclosed · hardware

Investors include Hubble Technology Investment (Huawei-affiliated investment arm) and Shenzhen Gaoxin Investment Group, per registration information from Tianyancha. DoNews describes Shita as a supplier of tactile sensing and data-acquisition products. Source: Stock Star source; DoNews source

Tesla Robotaxi Launches in Miami · autonomy

Tesla announced on X that its Robotaxi service has launched in Miami, with the published map covering only a small area of western Miami. Mashable reports that these vehicles have no safety monitor on board. Tesla previously ran limited deployments in Houston and Dallas in April, and Austin only expanded from a limited zone to full-city coverage last month. In the same week, NHTSA sent Tesla an inquiry regarding Cybercab certification (see Industry News). Source: Mashable source

Caterpillar Autonomous Haul Trucks Expand Operations at Luck Stone · autonomy

Aggregate producer Luck Stone is expanding its autonomous hauling partnership with Caterpillar. The decision is based on results from the Bull Run quarry: since going live in November 2024, the autonomous trucks have hauled over 3.5 million tons. Travis Chewning, Luck Stone's vice president of engineering and operational support, said autonomous hauling "makes the workplace safer and makes output to customers more consistent." Source: Global Mining Review source

Galbot (Chinese humanoid robot company) Says Commercial Robot Deployments Exceed 1,000 Units · humanoid ⚠️ vendor claim

At the SCO Digital Economy Forum on September 15, newly appointed chairman Yin Fangming made his first public appearance. According to Galbot, its commercial robot deployments have exceeded 1,000 units, with customers including CATL, Bosch, Siemens, Toyota, SAIC Motor, and BAIC, performing tasks such as loading/unloading, sorting, transport, and screw driving; service coverage spans about 50 cities, with nearly 200 "Galaxy Pods" deployed. The company did not break down how many of these 1,000 units are on third-party customer production lines — precisely the point Mech-Mind CEO Shao Tianlan previously questioned (reported earlier). Source: Shanghai Securities News source

Zhilai Says Its Polishing and Inspection Robots Have Entered Tier 1 Production Lines at Scale · industrial ⚠️ vendor claim

Zhilai Embodied Intelligence, founded in 2021, says its high-speed photographic inspection, force-controlled polishing, and ironing robots are already in large-scale use on production lines at multiple leading global auto-parts manufacturers; the report did not name specific customers or unit counts. Chairman and CEO Yu Qiong has nearly 30 years in the auto industry, having previously held senior roles at Ford STA and in procurement; the R&D team includes members from Alibaba's AI team and Nanjing University's reinforcement-learning group. Source: 36Kr source

NHTSA Requires Tesla to Explain How Cybercab Passed Certification · autonomy

Following last week's NHTSA audit of Tesla's self-certified compliance for the Cybercab (reported earlier), U.S. regulators have formally required Tesla to respond by September 30. The questions are specific: whether the vehicle temporarily had human driving controls such as a steering wheel and pedals installed during testing and certification; whether the vehicle can be operated by a human; whether passengers can move the vehicle via a touchscreen; and whether speed, operating area, and time of day are restricted. The answers will determine whether the vehicle's configuration at certification matches its configuration during commercial operation. This inquiry is part of an audit process and does not indicate NHTSA has determined Tesla violated any regulations. NHTSA is simultaneously revising rules covering pedals, mirrors, wipers, and lighting; until the revisions are finalized, manufacturers must either meet current standards or apply for exemptions. Source: Reuters source

South Korea's Five-Year Physical AI Roadmap: 1.41 Trillion Won Bet on Two Regional Factories · industrial

South Korea's Ministry of Science and ICT announced its 2026–2030 R&D roadmap on the 16th, with two physical AI projects in Gyeongsangnam-do and Jeollabuk-do totaling 1.4131 trillion won (about $1.01 billion). The Gyeongsangnam-do project focuses on precision manufacturing: using large action models as the core, training AI on motion data from workers and robots to directly control on-site equipment, then feeding results back into training. The Jeollabuk-do project builds a factory platform, with AI designing and operating a "lights-out factory" that integrates robots from multiple manufacturers. Seoul National University professor Cha Seok-won said 13 companies in Gyeongsangnam-do have already agreed to share tacit-knowledge data from veteran factory workers. Validated technology is planned to be rolled out to other factories, ultimately packaged into a "K-Manufacturing Factory" export product. Source: Seoul Economic Daily source

Zhijie RX Obtains L3-Level Autonomous Driving Road Test License · autonomy

Richard Yu (Huawei executive) confirmed via a repost that the Zhijie RX is undergoing real-world testing on complex road sections including urban expressways. The vehicle is equipped with 4 LiDAR units and 38 fusion-perception sensors, with Huawei's ADS full-domain redundant architecture; presale price starts at 299,800 yuan. This is a road-test license, not yet L3 commercial operation authorization. Source: Autohome source

Leapmotor's World-Model-Based ADAS Moves Down to Sub-100,000-Yuan Vehicles · autonomy ⚠️ plan-stage claim

At its Technology Day, Leapmotor unveiled a self-developed world-model-based driver-assistance system that does not rely on BEV perception or rule-based code, with LiDAR serving only as safety redundancy. Founder Zhu Jiangming said that as of the end of August, Leapmotor had delivered 350,000 LiDAR-equipped vehicles, all of which will be upgraded to the world-model version for free. Among these, the earliest roughly 10,000-plus units have a different hardware platform, and algorithm porting alone will cost at least 50 million yuan. Advanced driver assistance will also be extended down to sub-100,000-yuan models such as the A10. Source: Autohome source

Unitree's Interim Report: Research/Education 73.6% of Revenue, Industrial Applications About 9% · humanoid

Following rumors of tightened IPO thresholds that sent Unitree's stock below 500 yuan (reported earlier), Unitree closed at 469.8 yuan on September 15, its sixth consecutive trading-day decline, with total market cap at 190 billion yuan, down about 57% from its first day of trading. According to a 36Kr review of the interim report, Unitree's revenue in the first half was 1.152 billion yuan, up 48.54% year-over-year; non-GAAP net profit was 244 million yuan, down 19.34% year-over-year. By revenue segment, research and education accounted for 73.6%, commercial/consumer 17.39%, while industrial manufacturing, inspection, and logistics combined made up only about 9%. 36Kr also reports that AgiBot, Galaxea (formerly Star Chart), and Xingdong Jiyuan (Robot Era) have reportedly been reported by industry peers to regulators, though none of the three has responded. Source: 36Kr source

XPeng Publishes XCoT Technical Report: Reasoning Tokens Placed Directly in the Action Space · autonomy

Traditional VLAs have a time lag between scene understanding, semantic reasoning, and action generation. XPeng's foundation model team's approach is to reduce tokens describing the future, instead generating a small number of structured decision tokens stored directly in the action space, which the Control FFN responsible for trajectories reads directly. Decisions first converge on directional choices like decelerate, hold, or turn left, then Flow Matching expands these into multiple concrete trajectories. The training data — a mix of "reasoning-action" supervision — totals 3.62 million entries, validated in the autonomous driving domain. XPeng says this decision vocabulary can later be extended to robotics. Source: Gasgoo source

Songyan Power's HERON-CRA: Reinforcement Learning Raises Sock-Folding Success Rate to 97.8% · embodied ⚠️ vendor claim

Songyan Power's Scalabot brand released the second model in its HERON series, focused on memory and recovery from failures. Company-reported figures show that on the sock-folding task, success rate with imitation learning alone was 38.5%, rising to 97.8% with the RL Engine enabled, with the full process taking over a minute. Demonstrations covered both the ARX robotic arm and the wheeled humanoid Noetix M1. After watching the video, tech outlet Zhidongxi noted that the robotic arm's stability appeared insufficient, and the lengths of cut vegetable pieces were uneven.Source: Zhidongxi source

DeepRobotics Response to Regulatory Inquiry: Four Core Technical Staff Came from the Same Disbanded Quadruped Team · industrial

A first-round regulatory inquiry response drafted by CITIC Securities shows that DeepRobotics CTO Li Chao, deputy director of advanced research Chu Zhen, and deputy director of R&D Mo Xiaobo all previously worked on Nanjiang Robotics' "Chitu" quadruped pre-research project, joining DeepRobotics after that project team was disbanded in 2017; hardware team lead Li Feng also came from Nanjiang. The response argues that since "Nanjiang's core business is AGVs and warehouse robots," there is no technical dependency between the two companies. Analysis outlet Jibendian Jiema questions this reasoning, noting it compares the two companies' main products without addressing where the core technology originated. Source: Sina Finance source

Wonik Robotics Tests Wheeled Humanoid Robots at Samsung Chip Fabs · humanoid

Wonik executive director Kim Min-cheol said the company is running a wheeled-robot proof of concept (PoC) with major semiconductor manufacturers including Samsung Electronics, for transporting materials such as wafers. Samsung's evaluation involves three stages, testing the equipment's tolerance for micro-vibration and dust; even if it passes, initial supply volumes will be limited, with neither timeline nor quantity disclosed. Source: The Elec source

Horizon Robotics' Yu Kai: Global Core Autonomous Driving Suppliers May Ultimately Number Only Two or Three · autonomy

Yu Kai said on September 15 that he expects the industry's end state to see 80% of automakers partnering with suppliers, while the top 20% of automakers develop vertically in-house, with third-party suppliers themselves consolidating. Source: Guandian.cn source