Key Takeaways Today

· Rokae (Chinese robot-arm maker) delivered its 10,000th humanoid force-control arm in Zoucheng; more than 70 embodied-AI companies are integrating it into their robots

· D.R. Horton, the largest US homebuilder, signed a multi-year agreement to build more than 1,250 homes on Promise Robotics' robotic production lines

· Omron is acquiring US industrial PC maker OnLogic for $675 million, with closing expected around January 2027

· South Korea's physical AI budget for next year is 3.1 trillion won, more than triple this year's

· Xinjian Transmission (Chinese roller-screw maker) said in its inquiry response that sales of lead screws to its top customer, "Company A," fell to just 224,400 yuan in July–August this year

· Stardust Intelligence (Astribot, Chinese humanoid robot startup) raised another 1 billion yuan three months after its Series B, at a valuation of 15 billion yuan

RoboSPA: at maximum difficulty, all four leading VLA models score zero on memory tasks · benchmark

Most common benchmarks use clean scenes and short procedures, so they don't reveal how hard a task must be before a model starts to fail. RoboSPA uses 56 base tasks, 280 variants and 527,000 trajectories to test spatial reasoning and long-horizon planning across five difficulty levels. The best performer, π0.5, saw its spatial-reasoning success rate fall from 39.2% at level one to 21.4% at level five. On the hardest memory-intensive tasks, all four models — RDT, GO-1, π0.5 and X-VLA — scored 0%.

Zhejiang University and others · Coverage: Keji Xingzhe (Tech Traveler) source

UNITAS: putting observation, action and scene change in the same 3D coordinate system · world-model

Most policy models convert back and forth between pixel and action representations. UNITAS instead uses 3D point trajectories, modeling in a single coordinate system how the gripper moves (action flow) and how objects move in response (scene flow), so one model both generates actions and predicts how the world changes. With only 1.7B parameters, it reaches an average success rate of 99.8% on LIBERO, 92.94% and 93.04% on the RoboTwin 2.0 Clean and Randomized settings respectively, and an average success rate of 85% on real-robot tasks.

DexForce (Chinese robotics startup) · Coverage: Robot Lecture Hall source; Chinese tech news source

Success Guided Sampling: train only on tasks that "sometimes succeed, sometimes fail," and million-scale parallel simulation finally yields learning · manipulation

In large-scale parallel reinforcement learning, most randomly generated task configurations are either too easy or too hard, wasting compute. The University of Washington's method trains only on configurations where the current policy's success rate is in the middle range, and scales to more than one million simulated environments, working for both legged locomotion and contact-rich manipulation. In simulation, the success rate for threading a nut onto a bolt rose from 6% under random sampling to 70%. On a UR5e real robot trained with zero real data and cameras only, gear assembly succeeded 47 times out of 50, while real nuts and bolts managed only 18/48.

University of Washington · Coverage: Interesting Engineering source

Omron acquires OnLogic (US) ｜ Full acquisition ｜ Enterprise value $675 million · industrial

The price is about ¥110 billion, paid in cash, mainly from the company's own funds, with closing expected around January 2027, pending regulatory approval. OnLogic is headquartered in Vermont, makes ruggedized industrial PCs, has 301 employees, and earns over 90% of its revenue from the Americas and Europe. After closing, the brand will be kept and CEO Sean Larkin will stay on, running it as an independent business. Omron plans to combine its own sensors and controllers with OnLogic's edge industrial PCs to build a physical AI platform. At the announcement, President Junta Tsujinaga said: "We hope to steadily capture the growing demand in North America." EE Times China reported that Omron's control-equipment business is currently concentrated in Japan and China, with the Americas accounting for about 10%. Sources: Reuters source; EE Times China source

Stardust Intelligence (Astribot) ｜ New round ｜ 1 billion yuan ｜ Valuation 15 billion yuan · humanoid

Led by Touhong Donghai (Shenzhen Investment Holdings), with participation from Hongtai Fund (Chinese VC firm), Ninghai Xingtaihe Fund and Shanheng Capital; existing shareholder Xianmu followed on. It comes just three months after a Series B of over 1 billion yuan, and the company has completed 4 rounds within six months. Stardust builds rope-driven AI robots; its new-generation humanoid T1 began shipping in June this year, starting at 89,900 yuan. The company says total capacity across its S and T series has reached the ten-thousand-unit scale. CEO Lai Jie said: "After waiting so long, investors are finally asking about commercialization."Source: PEdaily source

AgiBot (Chinese humanoid robot maker) ｜ New financing launched ｜ Valuation 500 billion yuan · humanoid ⚠️ Unconfirmed report

36Kr (Chinese tech media) exclusively reported that AgiBot has started a new funding round at a valuation of 500 billion yuan, with its IPO on hold. At the end of July, Yicai (Chinese financial media) learned from AgiBot that the company had begun the process of listing in Hong Kong. AgiBot has not yet responded publicly.Sources: 36Kr source; Yicai

Sankai Electromechanical (Zhejiang) ｜ Series B ｜ Several hundred million yuan · hardware

Co-led by the Industrial Mother Machine Industry Investment Fund and Wuzhou Zhongda Investment, with Caitong Capital and Wenling Haihe Rongzhi following. At the end of last year it had just closed a Series A of over 100 million yuan led by Dachen Caizhi. Sankai was founded in 2007, began mass-producing precision planetary reducers in 2012, and since 2022 has made planetary and harmonic reducers for robots. The new funds will mainly go to humanoid-robot joint reducers and integrated drive units.Source: 36Kr source

Ubitium (Germany) ｜ Seed round ｜ $14 million · hardware

Led by Ibex Investors with participation from OpenOcean, bringing total seed funding to $18 million. In robots, sensor-signal processing, AI inference and real-time control typically run on three different types of chips; Ubitium wants one reconfigurable processor to handle all three. Its prototype chip UB108 has 256 processing elements and was taped out on Samsung's 8nm process. The first commercial chip, UB410, is planned to have 4,096 processing elements, with tape-out in 2027 and evaluation samples sent to customers.Source: AI Insider source

Shita Robotics (Shenzhen) ｜ Angel round ｜ Amount undisclosed · hardware

Huawei Hubble, a fund under Shenzhen Hi-Tech Investment, and Gaoxuan Investment participated. Founded in 2025, the company makes electronic skin from flexible photonic crystals plus photosensitive devices, targeting a thickness under 1 millimeter and a sensing density above 10,000 points per square centimeter. Founder Zhang Wei was born in 1998, and the team averages 23 years old.Source: PEdaily source

Rokae delivers its 10,000th humanoid force-control arm · embodied

At Rokae, which is listed in Hong Kong, the 10,000th humanoid force-control arm came off the line and was delivered at its Zoucheng Future Factory. China News Service said it is the first company in China to reach ten-thousand-unit delivery in humanoid force-control arms. Rokae already works with more than 80 companies, and more than 70 embodied-AI companies are integrating this arm into their own robots. Zhu Bainan, technical director of the humanoid arm business unit, used wiping a table as an example: "Press too lightly and it doesn't get clean; press too hard and it doesn't move smoothly." He argues that cameras can only see position, not "force." Selling arms to robot makers reached ten thousand units earlier than selling complete robots would have; this industrial robot maker has tapped embodied-AI demand in a different form.Source: China News Service source

D.R. Horton signs multi-year agreement to build more than 1,250 homes on Promise Robotics' robotic production lines · industrial

D.R. Horton is the largest US homebuilder by number of homes built, and it will use Promise Robotics' "factory as a service" (FaaS) to produce more than 1,250 homes. It is the first major US homebuilder to adopt the platform. Promise's system reads existing floor plans, generates production plans and robot instructions, then connects to on-site installation, so builders don't need to build their own factories. D.R. Horton's senior vice president of strategic growth and innovation, Brad Conlon, said the appeal is shorter build times, higher capacity and more predictable delivery.Source: PR Newswire source

Bot Auto picks Lenovo AD1 as its mass-production compute platform, says the full autonomous-driving hardware costs about $50,000 per vehicle · autonomy ⚠️ Vendor claim

Bot Auto, a Houston driverless-truck company, will begin fitting trucks with Lenovo AD1 domain controllers from Q1 2027. The company estimates total hardware cost for compute plus sensors at $50,000 per vehicle and says this is three to eight times lower than other approaches. The denominator in that comparison comes from an industry cost guide, not from competitors' public quotes.Source: FreightWaves source

Kodiak and Charger open a Dallas–Laredo autonomous freight lane · autonomy

The route is 435 miles long, hauls refrigerated and dry goods, and completed its first load on September 8, with a safety driver still in the driver's seat. The Port of Laredo handled over $350 billion in trade in 2025, and the two companies say it accounts for nearly 40% of US–Mexico trade. Kodiak must complete a safety case before removing the driver; its internal "autonomy readiness" metric was 96% at the end of September, versus 93% at the end of August.Source: Hoodline source

South Korea's physical AI budget next year is 3.1 trillion won, more than triple this year's · adjacent

At Seoul's Physical AI Technology Week on Monday, South Korea's Ministry of Science and ICT announced the 2027 cross-ministry plan, raising the physical AI budget from 900 billion won this year to 3.1 trillion won. Fiscal funds go to core technologies and training infrastructure. The Financial Services Commission will additionally use the National Growth Fund, which totals 150 trillion won over five years, to support commercialization and large-scale equipment investment. Within it, the "Ultra-Long-Term Technology Investment Fund" is planned at 880 billion won this year, with a term of up to 16 years.Source: Digital Today source

Portland transportation bureau proposes robotaxi caps: 75 vehicles per company, 225 citywide in the first year · autonomy

Waymo and others have not yet launched in Portland. A draft that the Portland Bureau of Transportation (PBOT) presented to the City Council's public works committee this week requires 50,000 miles in the city before paid service can begin, 25,000 with a human driver and 25,000 without. Whether the council will also require paid safety attendants on board appears only in social-media discussion so far; there is no formal proposal.Source: Townhall source

Musk calls Waymo a "pet killer," offers no Tesla detection data · autonomy

Following last week's mention of gray cats on gray roads that are hard to recognize at night, Musk said in a Saturday livestream: "Waymo is a pet-killing machine... they really do run over a lot of pets." He said Tesla has solved the problem but gave neither a launch date nor detection data. Citing NHTSA data, Business Insider reports that between July 2025 and July 2026 Waymo reported at least 9 collisions involving animals, 6 of them domestic animals.Source: Fox Business source

Liang Wenfeng reportedly meeting embodied-AI founders intensively; DeepSeek weighing building its own team or investing · adjacent ⚠️ Unconfirmed report

Citing people familiar with the matter, Phoenix Tech reports that DeepSeek may form its own R&D team or enter through investment. It led a round in Winverse AI, founded by Hu Han (already reported), though Kuai Technology, citing several industry sources, says Winverse focuses more on underlying model acceleration and is not strictly an embodied-AI company. DeepSeek did not respond.Sources: MyDrivers source; Sina source; 36Kr source

Liao Hongtao, former co-CTO of UBTech (Chinese humanoid robot maker), joins Tencent's Futian Lab · embodied ⚠️ Unconfirmed report

Liao Hongtao will head the embodied-intelligence effort, and the team has begun hiring; neither company has made an announcement. Futian Lab's full name is the Guangdong Provincial Laboratory of Embodied Intelligence, jointly built by the IDEA Research Institute and Tencent and launched in November 2024, with Tencent chief scientist Zhang Zhengyou as director.Source: Guandian source

Hello Robotaxi (Chinese ride-hailing company's robotaxi unit) releases driving world model HelloWorld · world-model

Developers can describe scenes in natural language or directly edit vehicle and pedestrian trajectories to generate multi-view video and lidar observations, inserting crossing pedestrians or switching between rain, snow, day and night, to fill in long-tail scenarios that road data collection struggles to cover.Source: Guandian source

AgiBot launches the "Yuansheng Ecosystem Development Plan": over 2 billion yuan across five years, more than 1,000 robots to universities · humanoid ⚠️ Stated plan

The first batch partners with more than 20 universities including Peking, Tsinghua, Zhejiang, Shanghai Jiao Tong and Xi'an Jiaotong. The robots deployed span the Jingling, Lingxi and Yuanzheng humanoids and the Kuotuo quadruped, along with the AIMA development platform.Source: Sohu source

Pony.ai (Chinese robotaxi company) and Uber plan road tests in London within weeks · autonomy

Following Zagreb, the two companies announced on October 8 that they will test Pony's seventh-generation robotaxi in London, as part of the earlier plan to deploy over 2,000 vehicles in five European cities. The UK's Automated Vehicles Act is expected to be fully in force in 2027; until then, pilots can proceed through a separate permitting route.Source: KrASIA source