Key Points Today

· Hyundai Kia is handing body-welding teaching to AI, with as many as 1,700 weld points per vehicle, and plans to use it in its South Korean plants next year

· Tesla's Austin Robotaxi has pushed its end-of-service time to midnight, still 6 hours short of round-the-clock operation

· Waymo has launched Nashville airport pickups and drop-offs, open to some riders first

· Hong Kong's first embodied AI testing center has opened, with 7 companies taking part, including Unitree (Chinese humanoid robot maker) and Galbot (Chinese embodied AI startup)

· Zhejiang University analyzed 100 failed embodied-task trajectories: 91% were memory-related

Paper Progress

Seeing Is Not Remembering: EMem-Bench Tests Embodied Memory with Four Task Types · benchmark

Robots on long-horizon tasks keep searching the same room or keep pulling at the same drawer that won't open, and the problem often lies in memory. A Zhejiang University team analyzed 100 failed trajectories of Gemini-3-Flash and Qwen3-VL-32B on EmbodiedBench, and 91% were linked to four types of memory problems. The benchmark built from this contains 2554 executable samples across four task types: passive observation, dynamic tracking, interaction failure, and experience generalization. The accompanying external memory system EMem raises Qwen3-VL-8B's average success rate from 25.4% to 36.2%. About three weeks after release, the benchmark's Hugging Face downloads are approaching 30,000.

Zhejiang University OmniAI team (Zhejiang University, Central South University, Institute of Software of the Chinese Academy of Sciences) · arXiv 2609.28236 source · Commentary: Xinzhiyuan (Chinese AI media outlet) source

Open Source · Tools · Evaluation

· Manifold AI WorldScape R²SI: Manifold AI (流形空间) says this recursive self-evolving robot framework topped the overall RoboDojo Agent Track leaderboard, ahead of GPT-6-Astra and Nvidia GR00T-N1.7, but has not published scores. The framework has two layers: an inner Robot Agent executes tasks, and an outer Recursion Agent reviews, revises skills, and trains a reactive model. On the push_T task, the first round took 723 steps, over the 600-step budget; after revision it dropped to 596 steps, then to 389 after further iteration. On September 23, Chuguang Yueqian's (初光跃迁) Simate-Beta also claimed the top of this leaderboard with an average score of 33.95; whether the two compete in the same track and whether their scores are comparable cannot be determined from public information. ⚠️ Vendor's own claim source; source

· LevelField-2 (沃土大模型 2.0): Lanwo AI (蓝沃AI), incubated by Shanghai Jingzhi (上海精智), open-sourced this machining-process model at the China International Industry Fair. It is built on an open-source Qwen base model with 9 billion parameters, and the Apache 2.0 license permits free commercial use. Based on a factory's own equipment, tooling, and material inventory, the model generates a process card for each operation, which process engineers then review. Since version 1.0 it has been trained on more than 100,000 sets of drawings covering 43 process types, with 36 veteran process masters taking part in annotation; version 1.0, open-sourced in July, has more than 16,000 individual users. source

Funding and Deals

Gewutong Zhi (格物通智) | Angel round | nearly RMB 100 million | valuation undisclosed · world-model

Investors include Peak (Chinese sportswear brand) Investment, Hongfu Capital, Hangzhou Capital, Hangzhou Talent Group, a provincial-level embodied-AI industry fund, and several leading listed companies. The company was registered only in August 2026. Founder and CEO Cai Jia previously led the foundation world model at Huawei's vehicle BU and took part in developing ADS 2.0 through 4.0. It builds an embodied brain using what it calls a "physical inversion world model": it infers contact and force from the motion in video, then calibrates with real tactile and force data. Peak has signed a strategic partnership, with footwear manufacturing as the setting. The company expects to deliver its first proof-of-concept (PoC) project by the end of 2026, in areas including textiles and apparel, optical modules, and auto parts. EE Times China (电子工程专辑) notes that at least 8 institutions co-invested less than RMB 100 million in this angel round, so the equity is fairly dispersed.Source: 36Kr source; EE Times China source

Xinji Wuxian (芯际无限) | Strategic investment from AgiBot (Zhiyuan Robotics, Chinese humanoid maker) | amount undisclosed · hardware ⚠️ Single-source claim

The investee is a Tsinghua-affiliated RISC-V AI chip company. Its "cerebellum chip" XC64AE uses 12 RISC-V cores plus a self-developed NPU; the first chip has been taped out and returned, and is being validated jointly with several embodied-AI companies. Its "brain chip" V1000 uses 2.5D chiplet packaging and is rated at 1000 TOPS. The company says its robot joint-control chip is already in mass production, and that its automotive-grade motor-control MCU XC32MC holds orders for more than 5 million units.Source: Sohu source

Wuyuan Technology (物原科技) | Multiple rounds | amount undisclosed | valuation several hundred million USD · world-model

Investors include Ant Group, Meituan (Chinese delivery platform), Capital Today-style Dachen Caizhi (达晨财智, Chinese VC firm), and Xinglian. Founder Wu Haixu won Tsinghua's Special Prize in 2022 and has been a postdoc at MIT CSAIL since 2025. In February this year he published GeoPT as first author, which uses synthetic dynamics to let a model learn the evolution of particle trajectories on unlabeled data; Kaiming He is among the collaborators. According to a person close to the company, Wuyuan is building a world model with physical modeling capability, with a technical approach partly continuing GeoPT.Source: QbitAI source

Yu Tao's new company | DeepSeek rumored to participate | amount undisclosed | two rounds valued at about USD 70 million and USD 200 million · world-model ⚠️ Rumor

After the report that it intends to lead an investment in Winverse AI, DeepSeek now also appears on the rumored investor list of the startup founded by Yu Tao. The information comes from Top Chinese Tech Society (Top华人科创社), and the deal has not been confirmed. Yu Tao did reinforcement learning on Tencent AI Lab's Juewu team, then became head of the humanoid robot foundation model at LimX Dynamics (逐际动力, Chinese humanoid robot maker), where he led the build of the whole-body control system LimX-WBT. The new company's direction is LLMs plus embodied intelligence.Source: Sohu source

Commercialization and Deployment

Hyundai Kia Hands Welding Teaching to AI, Plans to Use It in South Korean Plants Next Year · industrial

A single vehicle has as many as 1,700 weld points, which workers used to "teach" one by one to plan the robot's motion paths. On October 16 at the Uiwang Research Institute, a welding robot first identified weld-point positions and travel paths in a virtual space, adjusted its own path and posture when it found errors, and only then worked on the car body. Hyundai Motor plans to apply this technology in its plants in South Korea starting next year. Training is done in simulation, using Nvidia GPUs and Isaac Sim for sim2real, that is, transferring what was learned in simulation to real equipment. At the same event, another robot installed a door weatherstrip with a dedicated gripper, adjusting its pressing force in real time as it worked. According to The Asia Business Daily, Hyundai Kia plans to formally introduce autonomous-operation technology starting in 2032, mainly in newly built plants; digital-twin virtual commissioning developed with Nvidia will first be used to prepare new-model launches at the Ulsan EV plant. These technologies will be on show at E-FOREST TECH DAY 2026 on October 19–21. On the scale of the exhibition, Hyundai's official release says about 200 new technologies, while Chosunbiz says more than 200.Source: Chosunbiz source; The Asia Business Daily source; Hyundai Motor source

Tesla Austin Robotaxi Now Runs Until Midnight · autonomy

The new ride-accepting window for Tesla's Austin Robotaxi is 6 a.m. to midnight, and the app has already pushed a notification to riders. In August it ran until 10 p.m., in early October it was extended to 11 p.m., and this time one more hour has been added. Tesla AI vice president Ashok Elluswamy said earlier that round-the-clock operation is only possible once FSD v15 goes live, which is planned to reach all Robotaxi vehicles this month; v15 has about 10 billion parameters, roughly ten times the previous version. That leaves 6 night hours short of 24-hour operation.Source: Not a Tesla App source

Waymo Launches Nashville Airport Rides · autonomy

Waymo driverless cars have begun serving Nashville International Airport (BNA) for pickups and drop-offs, initially open to only some riders. According to EVwire, it is the first fully driverless ride-hailing service in Tennessee to offer airport rides. Earlier, Nashville's fleet depot, managed by Lyft's Flexdrive, went into operation, handling charging and maintenance for these vehicles.Source: FTN news source; EVwire

Industry News

Hong Kong's First Embodied AI Testing Center Opens, with Unitree and Galbot Participating · adjacent

The Hong Kong Productivity Council has opened the Embodied AI International Applications Hub at its Kowloon Tong headquarters, under the council's Industrial AI and Robotics Center. The first batch includes 7 mainland China and Hong Kong companies, among them robot makers Unitree and Galbot. The center simulates near-real working conditions so that manufacturers can test and optimize robot designs before overseas deployment. Council chairman Sunny Tan said Chinese manufacturers can use Hong Kong as a platform to train robots for overseas operations without leaving China. Secretary for Innovation, Technology and Industry Sun Dong said the center's opening is an important step in Hong Kong's embodied AI strategy.Source: ForkLog source

Altman: OpenAI Humanoid Robot Demo in 2027 · humanoid ⚠️ Statement from a speech

In September, Altman said on a podcast that OpenAI "will" build a humanoid; this time he gave Fortune a timeline: "I think we'll have a cool demo in 2027... it will be a few more years before robots are out on the streets." He did not reveal what the demo will show, what the machine will look like, or when outsiders will get access to it. OpenAI's careers page lists 27 robotics roles, including custom electromechanical actuators, manufacturing, and supply chain.Source: Mint source; ITHome source

Zhongke Guiji's Wang Peng: Dexterous Hands Are in a Bubble, and Many Customers Buy Them for Algorithm Validation · embodied

China Business Journal (中国经营报), citing IT Juzi (IT桔子, Chinese startup-funding database) data, reports that as of early August, funding in the dexterous-hand sector had passed RMB 28.5 billion, with the first 7 months already exceeding the whole of 2025. At the China International Industry Fair, Wang Peng, founder of Zhongke Guiji (中科硅纪, Chinese dexterous-hand maker), said that a demo lasting half an hour to an hour places different demands on a product than continuous operation for 8 or even 24 hours. "Many customers who buy dexterous hands use them as a platform for developing algorithms and validating technology." He also believes that if dexterous hands are made only to accompany humanoid robots, near-term demand is limited. Zhongke Guiji was incubated by the Chinese Academy of Sciences' Institute of Automation and says cumulative orders passed RMB 100 million as of the first half.Source: China Business Journal source

Data Collectors' Daily Target Is 160 Samples; Night Shift Pays RMB 135 a Day · adjacent

A data-collection studio in Beijing sets a daily target of 160 samples per collector; top performers can reach more than 260, and a large screen publicly shows real-time rankings for output and accuracy. The figures come from South China Morning Post reporting, relayed by Mint. Day shifts pay RMB 120 a day and night shifts RMB 135, and many people leave after a few weeks. Guo Hongjuan of Jiangsu collects data at home about 6 hours a day, earning roughly RMB 4,000 a month. Reuters reported in August that a novice needs about 300 attempts on average to produce one usable motion, and even experienced workers need 50. Three employees said this subsidized project still shows no path to profit.Source: Mint source

ICE Spends About USD 1.3 Million on 4 Modified Spots; Protesters Demonstrate Outside Boston Dynamics Headquarters · adjacent

The procurement plan revealed in September has been signed into a contract. Federal contract documents show that US Immigration and Customs Enforcement bought 4 Cerberus units for about USD 1.3 million; Connecticut company RADeCO builds them on Spot with an added robotic arm, thermal imaging camera, and explosives sensor. Groups including Don't Mourn Organize already protested outside Boston Dynamics headquarters on October 15, with another on the 20th. They are demanding that ICE's Spot software license be revoked and that Massachusetts claw back the USD 25 million tax credit it gave Boston Dynamics in June. A Boston Dynamics spokesperson said products sold through RADeCO are equally bound by the company's ethics terms, and that the robots are legally prohibited from being weaponized.Source: Waltham Times source

Aether AI Releases Causal Robot System CRIS-0 · world-model ⚠️ Vendor's own claim

CRIS-0, from San Diego startup Aether AI, first uses a world model to predict the consequences of each action, then checks them against actual results; when they don't match, it retries, steps back, or replans. The company says it replans in about 2 seconds after a disturbance and reacts to safety risks in about 0.2 seconds. Interesting Engineering points out that these figures come from the company's own controlled tests, which gave neither the number of trials nor the robot used. Founder Biwei Huang is an assistant professor at UCSD and has studied causal discovery for 12 years. The company raised USD 20 million in June.Source: Interesting Engineering source

4-Meter-Tall Hydraulic Humanoid Neo1 Debuts at the China International Industry Fair · humanoid ⚠️ Vendor's own claim

Neo1 from Beijing's NeodraDynamics weighs 1.2 tonnes, and 39 of its 41 active degrees of freedom are hydraulic joints. The company says its two arms can carry a combined 500 kg and that it targets hazardous work such as nuclear-facility maintenance, mine hauling, and rescue. It has been tested in the supply chain and industrial sites of JD Industrial (JD.com's industrial supply arm), but public materials do not say how it moves, whether it needs a human operator, or when it will be deployed.Source: Unboxholics source

Officials Say Over 10,000 Robots Have Settled in Zhangjiang AI Innovation Town · adjacent

This official figure does not specify whether "settled" means deployed, used for training, or owned by companies. The general-purpose brain from Ant Lingbo (蚂蚁灵波), a company in the town, was adapted during pre-training to 17 robot brands and 20 configurations. The National and Local Co-built center's "Linglong" robot has entered Yanfeng (Chinese auto-interiors supplier), doing data collection and autonomous-operation validation at stations such as flame treatment and kit sorting. From October 23 to 25, the International Embodied Intelligence Skills Competition will be held in Zhangjiang, with 9 tracks and 23 events.Source: IT Times source

Hardware · Supply Chain

· Intel Razor Lake AX: Intel vice president Cho Min-sung said in Seoul on October 16 that this robotics SoC will launch in 2028, targeting about 700 TOPS of compute. The current Panther Lake-based robotics development kit has a theoretical 180 TOPS, and the next-generation Nova Lake arrives in early next year. The Elec compares it with Nvidia's previous-generation Jetson AGX Orin at about 275 TOPS. He said South Korean customers' AMRs begin mass production next year, while most humanoids are slated for 2028. ⚠️ Planned, not delivered source

· Orbbec (奥比中光, Chinese 3D vision sensor maker): launched the 24-line 3D lidar Pulsar MH24 at the Canton Fair. Revenue from its vision solutions for the global robotics market in the first half was more than 2 times that of the same period last year, and it has more than 1,600 robotics customers. Its Bac Ninh, Vietnam site plans small-scale production by the end of 2026 and volume production by the end of 2027; together with the Shunde base, it plans annual capacity of more than 6 million units. ⚠️ Planned, not delivered source