Key Points Today

· The US Department of Transportation approved a five-year exemption for autonomous trucking companies including Aurora, allowing cab-mounted warning beacons to replace manually placed warning triangles when stopped

· The US Treasury issued the first-ever penalty of $200,000 under its outbound investment review, stemming from an investment of about $92,000 in Chinese embodied-AI startup Flexiv-incubated Qiongche Intelligence

· RoboJEPA proposes a scaling law for multi-embodiment robot world models as compute grows, and has released its 8B model and all code

· Chinese autonomous delivery vehicle company Rino.ai (White Rhino) closed a C2 round, bringing its cumulative Series C to $100 million

· Sequoia led robot-data company Mecka's $60 million Series B, with Nvidia and Samsung as new investors

· When the instruction says "stove," π0.5 turns on the burner with a 100% success rate; with "hot plate," it drops to just 2%

RoboJEPA: A Scaling Law for Multi-Embodiment Robot Latent World Models · world-model

RoboJEPA trains a JEPA world model on data from 12 robot embodiments. The authors find that its latent-space rollout error follows a second-order power law in compute, and downstream planning performance also improves predictably with compute, so rollout error can substitute for part of real-robot evaluation. The largest model has 8B parameters, which the authors call the largest JEPA predictor to date. On real robots, given only a single goal image, it completes long-horizon tasks zero-shot. All checkpoints and the training and deployment code have been released.

Artem Zholus et al. · arXiv 2610.10515 source

Long-WAM: World Action Models Need an Autoregressive Backbone to Use Long History · world-model

Long-WAM has 83↑ on HF, and concludes that being able to see history does not mean being able to use it. On RoboCasa GR-1, as context grows from 0 to 19.2 seconds, the success rate of an autoregressively pretrained video backbone rises from 63.3% to 78.7%, while a bidirectionally pretrained backbone shows no net gain. On real Unitree G1 and YAM robots, the dynamic cup-stacking success rate is 95%, while π0.5 and Fast-WAM did not succeed even once in 20 trials.

Wei Huang et al. · arXiv 2610.10528 source

OpenWAM: Stanford Builds a Controlled-Variable Testbed for World Action Models · world-model

The team of Fei-Fei Li and Jiajun Wu fixes a backbone of a 5B video expert plus a 2B action expert, changes only the ordering and mutual visibility between video prediction and action generation, and compares the variants one by one. The transfer experiment is the most useful: an inverse-dynamics module that sees only local imagery and is also trained with counterfactual data reaches an average success rate of 84.0% on new LIBERO-90 tasks, while the version that receives full context reaches only 47.0%. The authors note that the video predictor on the target tasks was fine-tuned with action-labeled demonstrations, so this does not count as zero-shot. The code is open source.

Fei-Fei Li, Jiajun Wu et al. (Stanford) · arXiv 2610.07922 source · Commentary: Machine Heart (Chinese AI media outlet) source

Change One Word in the Instruction, and VLA Success Rate Can Swing by Dozens of Points · vla

On LIBERO, when π0.5 receives "switch on the stove," it turns on the burner with a 100% success rate; with "switch on the hot plate," it succeeds only 2% of the time. The authors first score multiple phrasings of a small number of training tasks, then have a large model distill 10 to 20 rewriting rules from them, and at deployment each instruction is rewritten once according to the rules. The frozen π0 gains a relative 16% to 27% on 12 held-out tasks, without retraining.

Mikey Watts et al. (independent researcher; UCLA) · arXiv 2610.10526 source

PhysEvo: Letting a Frozen Model Self-Evolve by Modifying Tools, Without Changing Weights · manipulation

PhysEvo has 22↑ on HF. It has a meta-agent read execution traces, diagnose failures, and then rewrite the tools and skills that the frozen model Astra calls, while the model weights stay untouched. On 42 RoboDojo tasks its success rate is 62.00%, versus 47.17% for the previous strongest reference, RoboDawn's single-pass Astra agent. Moving the harness evolved in simulation onto a real AgileX PiPER arm, across 5 tasks and 25 trials in total, the success rate is 84.00%.

Wenqing Tian et al. · arXiv 2610.08995 source

RobotWorld: What Still Stops General-Purpose Agents from Taking Over Robots Directly · benchmark

RobotWorld has 26↑ on HF, and uses 84 simulated tasks to test how well multimodal agents complete physical tasks through robot interfaces, covering manipulation, mobile manipulation, legged locomotion, driving, and flight control. The agents build pipelines such as image segmentation and camera calibration on their own, but the combinations often fail, for example reaching the commanded pose while losing the object's state, or treating an unfinished task as complete. Astra is stronger on spatial and constrained-contact objectives, while Opus 5.5 is stronger on continuous-balance and timed-interaction objectives.

Zhiqin Yang et al. · arXiv 2610.10409 source

CureWM: Robot World Models Keep Predicting Failure as Success · world-model

The authors examined four published checkpoints from two architecture families and found that all of them are insensitive to changes in action, and that actions already confirmed to fail are often predicted as successes. CureWM starts from successful demonstrations, constructs substitute actions of varying severity, executes them in simulation or on real robots to verify the outcomes, and then fine-tunes on those samples. On 484 LIBERO failure counterfactuals, the share of failures the model predicts as successes falls from 80% to an average of 38%. The code is open source.

Jiuyi Xu et al. · arXiv 2610.09134 source

TouchScale: 500 Hours of Same-Source Human Visuo-Tactile Data · benchmark

TouchScale used a single wearable setup to collect 500 hours of human interaction data with about 2,000 task descriptions, synchronously recording egocentric RGB-D, wrist video, and full-hand tactile signals from both hands. Using it for visuo-tactile mid-training of robot policies raises the average success rate on four contact-rich real-robot tasks from 22.5% to 57.5%. The dataset will be made public.

Dayou Li et al. · arXiv 2610.10288 source

FlashNeRD: Learned Robot Dynamics Faster Than the Simulator · world-model

The predecessor NeRD replaced a simulator's numerical dynamics with a learned model, but was not much faster. FlashNeRD switches to a streaming parallel architecture, lets contact points appear anywhere, and can couple bidirectionally with objects simulated by analytical solvers. Its dynamics model is up to 55 times faster than the optimized NeRD, and training an ANYmal locomotion policy with PPO takes only 34 seconds.

Mohammadmehdi Ataei et al. · arXiv 2610.09130 source

Workhorse: Learning Humanoid Whole-Body Carrying from Human Demonstrations Only · locomotion

Workhorse is trained on human demonstrations collected without robots and without retargeting. A visual planner predicts target poses for the torso, both wrists, and both feet, and a reinforcement-learning whole-body tracker then executes them on the robot. A real Unitree G1 can sort boxes using both hands and feet and can catch a thrown box; it recovers when someone shoves it or snatches the box away. The sorting success rate in simulation is 77%.

Songbo Hu et al. · arXiv 2610.09117 source

Rino.ai (白犀牛, White Rhino) ｜ Series C2 ｜ $100 million cumulative Series C ｜ Valuation undisclosed · autonomy

The round was led by Yinshan Capital (Chinese investment firm), with participation from Xiangtan state capital, Shenzhen Heavy Investment (Shenzhen state-backed investor), Hunan Caixin, Woori Financial Group of South Korea, and Bozheng Capital, roughly five months after the Series C1 in May. Including this round, Rino.ai has raised more than RMB 1.42 billion in total. The company was founded in 2019 by Zhu Lei and Xia Tian, former members of Baidu's autonomous driving team, and builds L4 autonomous delivery vehicles for urban public roads; it says it has delivered more than 5,000 autonomous logistics vehicles in total, covering over 200 cities. SF Express (Chinese logistics giant) has invested in Rino.ai three times within a year. Source: Cheche Xi (Chinese auto media) source; Gasgoo source

Mecka ｜ Series B ｜ $60 million ｜ Valuation undisclosed · adjacent ⚠️ Vendor claim

Sequoia Capital led, with Nvidia, Microsoft's M12, Qualcomm Ventures, and Samsung as new investors. Individual investors include DoorDash CEO Tony Xu and Milan Kovac, former head of Tesla's Optimus program. Mecka pays people to wear sensors and record everyday actions such as making coffee and repairing cars on phones, converts the data into robot training formats, and also handles deployment integration for companies without robotics teams. The company says its annualized revenue exceeded $100 million in June and expects to reach $300 million by year-end. TechCrunch previously reported that the valuation when the round was under negotiation was about $500 million. Source: TechCrunch source; AI Insider source

Figure ｜ Nvidia in talks on additional investment ｜ $1 billion ｜ Pre-money about $38 billion · humanoid ⚠️ Rumor

According to The Information, citing a person familiar with the matter, Nvidia has discussed investing another $1 billion in Figure. Figure is seeking new funding at a pre-money valuation of about $38 billion. A year ago it raised more than $1 billion at the same valuation, and Nvidia was already a shareholder at the time. Source: ITHome (Chinese tech news site) source

Qingli Intelligence (厘清智能) ｜ Angel round + Angel+ round ｜ Hundreds of millions of RMB ｜ Valuation undisclosed · adjacent

Ant Group invested in both rounds and led the Angel+ round; CMC Capital, SinoAuto Investment, and a fund under CICC Capital (Chinese investment bank's investment arm) participated, with existing shareholders such as Shunwei Capital and Fengrui Capital adding to their positions. The company was founded in April this year and already secured a seed round of hundreds of millions of RMB in June. Its products are the T1 series data-collection gloves and the EgoDepth capture device, used to record hand pose, touch, and depth. Founder Li Yiming is currently an assistant professor at Tsinghua University's School of Artificial Intelligence and was formerly a research scientist at Nvidia. Source: Zhidx (Chinese tech media) source

Tian Keyu's world-model company (unnamed) ｜ New round ｜ Nearly $30 million ｜ $200 million post-money · world-model

According to Bloomberg, Wuyuan Capital and IDG Capital participated. Tian Keyu received his PhD from Peking University last summer and is the first author of a NeurIPS 2024 best paper. In the same year he was dismissed by ByteDance for serious disciplinary violations, and a court ordered him to pay RMB 500,000 in compensation. The company has not yet been named and has no product. Its technical approach is to build an AI-specific token language for video; it has so far produced a "dictionary" of 200,000 symbols and plans to feed in 100 million hours of video. Tian says this can cut the cost of generating each second of video by at least an order of magnitude, with a formal launch planned for 2027. Source: Robot Qianzhan (Chinese robotics media) source; Sina Tech source

Lingsheng Technology (灵生科技) ｜ Series A ｜ Hundreds of millions of RMB ｜ Valuation undisclosed · embodied ⚠️ Vendor claim

Huafang Capital, Daohe Yuanqi, Suwen Electric Power, and an industry player in embodied AI invested jointly. Lingsheng builds RUDA, a cross-embodiment general-purpose foundation, with the goal that newly connected robots can reuse existing data and skills. Founder Yang Hongbing was formerly CTO of a leading humanoid robot company, and co-founder and CTO Jiang Yuhua is a Tsinghua PhD born after 2000. The company says its optical-module plug-in task reaches a test success rate close to 100% under specific conditions. Source: Pedaily (Chinese investment media) source

ScanArk (景立构) ｜ Pre-Angel round ｜ Tens of millions of RMB cumulatively across three rounds in 4 months ｜ Valuation undisclosed · world-model

Within four months, ScanArk completed a seed round, a seed+ round, and a Pre-Angel round, with investors including Xiaomiao Langcheng, Guohai Innovation Capital, and a leading embodied-robotics company. This team of Tongji University PhDs reconstructs real environments into native 3DGS scenes with physical properties and interaction states; robots first try and fail in the digital replica, then verify on real hardware. Source: 36Kr (Chinese tech media) source

Linggan Robotics (灵感机器人) ｜ Angel round ｜ Tens of millions of RMB ｜ Valuation undisclosed · hardware

Lihe Capital and Hefei Innovation Investment invested jointly. Founded in 2025, the company makes fingertip visuo-tactile sensors, force sensors, and electronic skin, and the funds will go to R&D, team expansion, and building a production line. Source: DoNews source; Ifeng Tech source

XPeng's Robotaxi Is Named "XPENG YOYO," Hailed by Invitation Code in Guangzhou · autonomy

To ride XPeng's Robotaxi, you currently need an invitation code. On October 8, XPeng (Chinese EV maker) named its Robotaxi business "XPENG YOYO," and the ride-hailing mini-program went live at the same time. In Gasgoo's hands-on test, hailing a car required entering an invitation code, and the service covers only designated areas of Guangzhou. The vehicle is based on the XPeng GX, carries 4 in-house Turing chips with 3,000 TOPS of compute, uses a second-generation VLA model, and does not rely on lidar or high-definition maps. By the August earnings call, internal test rides in Guangzhou had exceeded 2,000. According to ITHome, XPeng targets carrying passengers without safety operators in 2027 and per-vehicle break-even for Guangzhou operations in the second half of 2027. Fleet operations are handed to partners, and XPeng earns revenue from vehicle sales, technical services, and operating revenue share. Source: Gasgoo source; paultan.org source

AgiBot Places 100 Lingxi X2 Units as Shopping Guides for Three Days; Store Claims 39% Sales Increase · humanoid ⚠️ Single-source claim

On September 28, 100 Lingxi X2 units (from AgiBot, Chinese humanoid robot maker) were stationed in Asd (Chinese cookware brand) stores. Shanghai Observer reported that the 3-day trial operation brought cumulative sales of about RMB 2.23 million, with average sales performance across the hundred stores up 39%. The robots work ten hours a day, first asking customers how many people are in their household and whether they prefer stir-frying or simmering soup, and then recommending cookware. To check promotions, customers must scan a code themselves, and checkout is handed to a human clerk. The report also mentions that many customers made a special trip to see the robots; the 39% increase does not exclude this foot traffic, and no comparison baseline period was given. Source: World Journal source

FF Robotics Ships 265 Units in September, 817 Cumulatively · embodied ⚠️ Vendor claim

FF had previously disclosed cumulative shipments of 552 units. In September, FF and FFR robot sales and shipments reached 265 units, a single-month record; the third quarter totaled 575 units, and cumulative shipments from the end of February to the end of September reached 817 units. The fourth-quarter focus is security patrol, and FF is developing a solution with North American enterprise developers that pairs robot dogs. Source: Gasgoo source

Pony.ai and Uber to Test Seventh-Generation Robotaxi in London · autonomy

Pony.ai's (Chinese robotaxi company) seventh-generation Robotaxi will begin road testing in London in the coming weeks; the two companies announced an expanded partnership on October 8, aiming to launch a Robotaxi service locally. Their plan in Europe is to deploy 2,000 Robotaxis, with the first stop, Zagreb, where Verne owns and operates the fleet. Neither side has said who will own or operate the London fleet. Uber has another partner in London, Wayve. Uber's additional $300 million investment in Wayve is conditional on Wayve deploying Robotaxis, with London first. Uber expects to offer autonomous rides in up to 15 cities worldwide by the end of 2026. Source: TechCrunch source

RoboSense Sells Its Complete-Robot Business to a New Company of Its Founder for RMB 59.878 Million · embodied

RoboSense (Chinese lidar maker) is selling its embodied-robot complete-machine R&D, manufacturing, and sales business, together with related patents and inventory, to Xiyuan Robotics for RMB 59.878 million in cash. As of the end of June, the unaudited book net value of this business was only RMB 2.384 million, and RoboSense expects to book a net gain of about RMB 37.9 million as a result. Xiyuan was founded in July 2026, and its actual controller is RoboSense founder and board chairman Qiu Chunxin. The transaction constitutes a connected transaction under Hong Kong Stock Exchange rules. Tianyancha shows that Xiyuan completed an angel round in late September, with investors including Xin Capital, Matrix Partners China, and Yuanma Lidong. RoboSense will be a component supplier to Xiyuan going forward. 21st Century Business Herald (Chinese business newspaper) explains that RoboSense building complete machines itself would put it in competition with its humanoid-robot customers. Source: 21st Century Business Herald source

UBTech: Teaming Up with FAW-Volkswagen, Building First Overseas Factory in Almaty · humanoid

FAW-Volkswagen (Chinese joint venture of FAW and Volkswagen) and UBTech (Chinese humanoid robot maker) reached a strategic partnership to jointly develop and test humanoid robots in automotive logistics scenarios and build demonstration scenarios; the announcement did not disclose the number of units or the amount. Separately, according to Gasgoo, Kazakhstan's President Tokayev met UBTech Vice President Zhong Yong in Almaty and said he supports UBTech building a robot equipment factory locally. This will be UBTech's first overseas production base, with products mainly aimed at education and services. UBTech also says it recently signed overseas orders in Europe, Japan, and South Korea worth more than RMB 50 million. Source: Gasgoo source; source

US DOT Grants Aurora and Other Autonomous Trucking Companies a Five-Year Exemption, Switching Stop Warnings to Cab Beacons · autonomy

The US Department of Transportation said late on October 7 that it approved a five-year exemption for Aurora and other autonomous trucking companies. Under current rules, when a truck stops, the driver must get out and place warning devices; with the exemption, cab-mounted warning beacons can be used instead. Waymo and Aurora jointly applied for a similar exemption in 2023, the FMCSA rejected it in December 2024, and Aurora then filed a lawsuit. According to FreightWaves, from October 2025 the FMCSA switched to granting temporary three-month exemptions each time, the latest of which expires on October 9. In its application, Aurora projects that its L4 truck fleet will exceed 200 vehicles by the end of 2026. Opponents are also pushing back: an Illinois freight carrier has asked the Seventh Circuit to stay the current exemption, and as of October 2 the court had not ruled. Source: Reuters source; FreightWaves

First US Treasury Outbound Investment Penalty Points to a Small Investment in Qiongche Intelligence · adjacent

On October 7 the US Treasury announced the first civil penalty under the Outbound Investment Security Program (OISP), in the amount of $200,000. The penalized party, Amidi, is the parent company of Plug and Play Tech Center; a China-based fund it controls invested about $92,000 in Shanghai-based Qiongche Intelligence (Chinese embodied-AI startup) on April 19, 2025, without making the required notification to the Treasury. The Treasury characterized it as a procedural violation, did not say the investment was a prohibited transaction, and did not allege any wrongdoing by Qiongche. The fine is more than twice the investment amount. Qiongche was incubated by Flexiv (Chinese robotics company) and builds software and hardware "brains" for embodied AI. The OISP took effect on January 2, 2025, and covers artificial intelligence, semiconductors, and quantum computing. Source: US Treasury; South China Morning Post source

China Has More Than 70 Embodied-AI Training Grounds in Operation · embodied

China Central Television (CCTV) reported that more than 70 embodied-AI training grounds have been built and put into operation across China, with over 40 more under construction or planning, concentrated in the Yangtze River Delta, Beijing-Tianjin-Hebei, and the Pearl River Delta. China's Ministry of Industry and Information Technology and other ministries recently launched a special campaign covering the industrial, service, and special-purpose domains, pushing training data into product iteration and large-scale deployment. Du Xiqi, head of Lightwheel (Chinese simulation and robot-data company), said its data platform has delivered 1.5 million hours of human video data. Dong Dianbiao, head of Shenzhen Cyborg Robotics, said that after training on tens of thousands of data samples, robots achieve a success rate of 80% to 90% on a single action. Source: Cover News source

US Department of Energy Selects 4 National Laboratory Robotics Projects · adjacent

The US Department of Energy's Office of Science selected 4 national-laboratory-led robotics and automation projects for autonomous science at experimental facilities, including Argonne's MAESTRO, Brookhaven's DART, and Oak Ridge's TRACE. The DOE requires the projects to produce open software interfaces, reusable robot skills, digital-twin environments, benchmark tasks, and datasets; the announcement did not disclose funding amounts. Source: Newswise source

IFR: Global Operational Stock of Industrial Robots Reaches 5 Million · industrial

The IFR released World Robotics 2026 on September 24: the global operational stock of industrial robots grew 9% in 2025 to a record 5 million units. New installations that year were 603,307, up 11%. The IFR expects installations to grow 9% in 2026 to about 655,000 units, and reach 806,000 in 2029. Growth is mainly in Asia; Germany's 2025 sales fell 8%, to under 25,000. Source: IFR; Global Electronics China

IDC: AgiBot Shipped 8,600 Humanoids in H1; 2030 Forecast Raised to 750,000 Units · humanoid

The latest IDC report shows AgiBot shipped over 8,600 humanoid robots in the first half, about 35% of the global total and more than 45% of the Chinese market. The same report raises its 2030 global humanoid shipment forecast by about 50%, to more than 750,000 units. For the global total in the first half, IDC counts nearly 25,000, while Frost & Sullivan counts 36,000 (already reported); the two firms' counting methodologies still do not align. Source: Autohome (Chinese auto portal) source; China Securities Journal source

Component Makers Ramp Up Shipments: H1 Figures from Changying Precision and RoboSense · hardware

Changying Precision (Chinese precision-components maker) said on an investor interaction platform on October 8 that it delivered more than 1.1 million precision humanoid-robot components in January–August this year, that third- and fourth-quarter order growth is faster than in the first half, and that it is expanding capacity to meet customer demand. Its products include core transmission parts for dexterous hands, rotary actuators, and lead screws, and its affiliate Guosenke makes harmonic reducers. Changying's "embodied intelligence and emerging technology hardware" business had first-half revenue of RMB 243 million, up 537.64% year on year, but this category also includes non-robot products such as cables and liquid cooling. RoboSense's first-half shipments of robotics and other lidar were 282,600 units, up about 510% year on year, while it spun its complete-machine business out of the listed company (see Industry News). Components are counted by piece and complete machines by unit, so the two growth rates cannot be directly converted into per-unit content. Source: China Securities Journal source; 21st Century Business Herald source

Upstream Materials: Direct-Drive Magnetic Materials and Reducer Flexspline Steel · hardware

At the end of September, Benmo Technology signed a strategic cooperation framework agreement with Zhongxi Kechuang, a unit of China Rare Earth Group, to jointly develop magnetic materials suited to direct-drive motors, such as neodymium-iron-boron, magnetic-concentrating rings, and radial magnetic rings. The specific projects and investment amounts await later implementation agreements. In South Korea, Asia Business Daily reported that Sea-Ah Steel has achieved localized production of the specialty steel at the core of robot reducers. Source: 10jqka Finance (Chinese financial portal); Asia Business Daily