Today's Highlights

· 27 autonomous mining trucks at Shougang Mining's Shuichang Iron Mine have moved into mixed human-machine operation, with effective operating time 8% higher than human-driven trucks.

· Deep Blue Robotics (Chinese robotics firm 96%-owned by DeepBlue Technology), is undergoing bankruptcy liquidation, owing more than 23 million yuan to over 100 employees.

· In Robocurve's RoboHarm evaluation, GPT-6 Astra controlled a robotic arm to complete 60 dangerous tasks, refusing on safety grounds only twice.

· SoftBank has agreed to acquire RAI Institute, founded by Marc Raibert, from Hyundai Motor Group; the deal is under CFIUS review.

A 9-gram, $5 pneumatic module packs pinching, twisting, hooking, and grasping into one structure · manipulation

Dexterous hands modeled on elephant trunks and octopus tentacles keep piling on parts and control complexity as they go. Four institutions including the Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences took a different approach, replicating only the core functions of a human hand: a dual pneumatic chamber plus a flexible coil spring as the skeleton, with two pneumatic input channels producing three deformation modes — extension, expansion, and contraction. The same 9-gram, roughly $5 module can both stably pinch a strand of hair 90 microns in diameter and grip objects with an inner cavity of 30mm to 210mm — a graspable size range over 2,300 times its own smallest scale, 12.9 times higher than existing systems with the same number of actuators, with a maximum single rotation of 690°. Two modules connected in series into a self-growing endoscopic robot completed inspection of six checkpoints on a full-scale (1:1) aircraft engine blade array, even capturing images of the blade trailing edges that traditional methods struggle to reach.

Li Yingtian and Yang Yang's team (Shenzhen Institute of Advanced Technology, CAS / Nanjing University of Information Science and Technology / Tongji University / CUHK-Shenzhen) · Advanced Science cover · analysis source

Qianjue Robotics (Chinese tactile-sensor maker) ｜ two consecutive strategic funding rounds ｜ hundreds of millions of yuan · hardware ⚠️ company-sourced figures

BlueRun Ventures, a fund under CICC Capital, and Vertex Capital participated. Following an earlier round this month worth over 100 million yuan, this Shanghai-based visual-tactile sensor company's shareholder list now includes Hillhouse Ventures, Yuanhe Origin Capital, and Fortune Capital, alongside industry players such as Li Auto, Zhiyuan Robotics (Chinese humanoid startup), and Chigo. The company says its self-developed tri-color visual-tactile sensor has reached ten-thousand-unit-scale mass production, with over 300 paying customers; its products are used by manufacturers including Zhiyuan, Google, AgiBot (Chinese embodied-AI startup), Galbot, and Unitree, with industrial-side embodied model validation done with CATL and a logistics solution being developed with L'Oréal. Founder Ma Daolin is a professor at Shanghai Jiao Tong University and winner of the ICRA 2021 Best Paper Award; during his postdoctoral work at MIT he studied under Alberto Rodriguez, who now heads Robot Behavior at Boston Dynamics. Source: zhidxcom source

Vantora (US) ｜ new round ｜ $100 million · adjacent

Silversmith Capital Partners contributed $100 million. The company was called UP.Labs four years ago, incubating startups for enterprise clients such as Porsche, Alaska Airlines, and J.B. Hunt; after rebranding, it now builds companies exclusively for enterprise clients, who can also fold the resulting startups directly back into their core business. Founder and CEO John Kuolt said that in the past, some ideas were too sensitive to a given partner to sell externally, forcing the company to abandon them — "what we were missing out on were exactly the highest-value problems." His example: a Fortune 100 industrial company needing to convert all of its hardware and machinery to autonomous operation, saying "that layer of intelligence has to be owned in-house — you can't outsource it to a third party." Physical AI is exactly where this kind of demand concentrates, and Vantora has signed new, unnamed clients in industrial manufacturing and oil and gas. Source: TechCrunch source

Treble Technologies (Iceland) ｜ Series A-2 ｜ $18 million · adjacent

Led by Paladin Capital Group, with participation from KOMPAS VC, Frumtak Ventures, and the European Innovation Council Fund, bringing total funding to around €36 million. Founded in Reykjavik in 2020, the company builds physics-based acoustic modeling: factoring in spatial geometry, material properties, and sound source position to generate acoustic data close to real-world spaces, allowing evaluation before a prototype is even built. Amazon and Logitech already use it to test audio products; Amazon senior audio research manager Wontak Kim said using virtual acoustic environments to test scenarios that physical testing can't reproduce is a key part of developing Alexa. Paladin's François Ruether offered a different angle: today's robots rely almost entirely on vision — they can see what's ahead, but can't pick up on sounds such as someone falling in the next room or a collision around a corner. Source: Robot Lecture Hall source

Zhenghe Industrial ｜ targeted share issuance registration approved ｜ up to 698 million yuan · hardware

The China Securities Regulatory Commission approved the registration for the issuance of up to 24.525 million shares. 148 million yuan will go toward a project integrating hardware and software for miniature chain systems; once at capacity, it will produce 4,800 three-fingered dexterous hands, 3,500 five-fingered dexterous hands, and 12,000 flexible actuators annually, along with 380 mobile work stations and 190 fixed work stations, with a projected comprehensive gross margin of 30.28% over a three-year construction period. Expansion of agricultural machinery components received 431 million yuan, remaining the largest single allocation. The company launched its chain-driven dexterous hand, CHOHO Hand, in January this year, saying it has already secured actual orders. First-half revenue was 978 million yuan, up 8.10% year-on-year, while net profit attributable to shareholders was 70.0787 million yuan, down 21.26% year-on-year, which the company attributed to increased foreign exchange losses and reduced government subsidies. Source: Haibao News source

27 autonomous mining trucks at Shougang's Shuichang Iron Mine move from dedicated-route pilot to full mixed operation · autonomy

Shuichang Iron Mine currently operates 27 autonomous mining trucks, with their operating range expanding from fixed-route supply for building materials production lines to the main pit, where crewed and uncrewed vehicles now run interspersed. The mine operator's comparison: hourly output is on par with human driving, exceeding it in some scenarios, with effective operating time up 8% without changing production organization. As of November 2025, autonomous trucks on the building materials line had accumulated over 350,000 km of safe driving and hauled over 11 million tonnes. Before mixed operation began, hardware was modified first — the pneumatic, hydraulic, and control mechanisms of 17 trucks were adjusted over nearly four months; because heavy-load downhill driving relies frequently on mechanical braking, causing brake pads to wear quickly and braking performance to degrade, the mine operator teamed up with Easy Control (Chinese autonomous mining tech firm) and Yutong Mining Trucks over two months to develop electric-retarding intervention technology, using electric braking to replace mechanical deceleration and recover kinetic energy for reverse charging. Under network-outage conditions, multi-source fusion positioning maintains centimeter-level accuracy. Wang Xinbing, who has 14 years of mining truck driving experience, transferred to autonomous driving operations and maintenance: "In the past it was year-round, high-intensity outdoor driving, exhausting and with safety risks — now I sit in the command center." Source: Workers' Daily source

Hai Robotics' HaiPick cuts picking staff at BEAMS' Tokyo warehouse to 6 people · industrial

BEAMS' Fukagawa fulfillment center went live with the HaiPick system in late September 2024, with 57 ACRs paired with 4.7-meter-high shelving and 32,500 storage locations, covering roughly 30,000 square meters across four floors, 75,000 SKUs, and 1.2 to 1.3 million units of inventory. Hiroki Hatano, section chief of logistics division strategy, gave these figures: previously, shipping 3,000 e-commerce orders in 2 hours required scheduling 20 people (typically 25 in practice); now 6 picking stations with 6 people handle the same volume, a 70% reduction in picking labor. The system ships 1,500 totes per hour, and the 4.7-meter vertical shelving raises storage capacity to 1.8 times the original. Hanging garments, footwear, fragile items, and very small accessories remain on manual shelves; totes allow mixing of 3 to 4 SKU categories, with error prevention handled via RFID verification plus live photos and color swatches on the picking interface. Asked about the difficulty of scaling up logistics robotics deployments gradually, Hatano said the project set scale-up targets from the outset and is currently proceeding fully on plan. Source: AGVwang source

Tesla's Cybercab registrations in Texas rise to 58 · autonomy ⚠️ compiled figures

The Texas Department of Motor Vehicles' autonomous vehicle registry added 9 Cybercabs on September 18, registered under Tesla Robotaxi, LLC, bringing the Cybercab count from 49 earlier this month to 58; including Model Y, Tesla's total approved Robotaxi fleet stands at 448 vehicles. Registration does not mean a vehicle is in active service — the registry does not indicate how many are within Austin's operating area versus parked at a facility. Tesla VP of AI Ashok Elluswamy cited 1 million miles of unsupervised driving, while the figure previously disclosed on the earnings call was 380,000 miles. NHTSA has required Tesla to explain by September 30 how the steering-wheel-less Cybercab meets federal safety certification (previously reported). Source: TeslaNorth source, The Cool Down source

JR East trials 3 self-moving trash cans at Shinjuku Station · adjacent

The robots stand just over one meter tall and move at 20 cm per second — deliberately slow, so passengers can approach to throw away trash and so the robot can stop at any moment. Shinjuku Station sees over 2.7 million passengers daily, the busiest station in the world per Guinness World Records, yet the section operated by JR East has only 7 sets of trash and recycling bins, requiring manual collection more than 10 times a day. The robots deliver collected trash to cleaning staff. JR East researcher Shinya Kawamichi said: "We wanted to test whether something like this could operate automatically and safely in an environment like Shinjuku Station," adding that the team also wants to know "how much difference a moving trash can makes, and whether people will actually be willing to throw trash into it." The trial runs through late November. Source: AFP source

AccelZoom's (Chinese humanoid robotics startup) tools now in over 30 middle schools in Beijing · humanoid ⚠️ conference remarks

Founder Cheng Hao said at the 2026 Tsinghua AI Youth Conference that over 30 middle schools in Beijing now offer embodied AI courses using AccelZoom's tools, and that the company will also open its embodied AI development platform to Tsinghua faculty and students. Source: Beijing Daily source

Unitree fully open-sources its next-generation general-purpose humanoid foundation model, UnifoLM-WLA-1.0 · world-model ⚠️ company-sourced figures

The model has 6B parameters, trained on roughly 2,500 hours of real-robot data, with a single model coordinating 64 tasks — 54 tabletop manipulation tasks and 10 whole-body mobile manipulation tasks — compatible with two-finger grippers and two types of dexterous hands. WLA stands for World-Language-Action, integrating world modeling into action generation: the reasoning backbone, UnifoLM-ER-1-4B, is built on Qwen3-VL-4B and trained on over 5 million pieces of embodied reasoning data, leading among participating open-source models on 7 of 16 multimodal perception and understanding benchmarks, scoring 82.0 on Where2Place — 13 points higher than GPT-6 Astra — and 88.9 versus 83.3 on EmbSpatial. In the prediction stage, optical flow is used to extract regions that genuinely change before and after interaction as a supervisory signal, so the model only judges "what will change" rather than repainting the entire frame; on the action side, end-effector pose, end-effector joints, and lower-body joints are discretized via RVQ across three separate channels, then decoded back into continuous actions by an MMDiT Action Expert. Benchmark results are self-reported by the company, and code, weights, and datasets are being released progressively according to the company. Unitree has previously open-sourced UnifoLM-VLA-0 and UnifoLM-WMA-0. Source: Caijing source

Figure rents 30 homes in the Bay Area to put Helix 2.5 straight to work · embodied ⚠️ company-sourced figures

With no new data collection, no additional training, and no environment adaptation, Figure 03 walked into homes it had never entered before and successfully completed three long-horizon whole-body tasks — tidying the living room, making the bed, and folding towels — with the company reporting a zero-shot success rate rising from 9% to 56%. Founder and CEO Brett Adcock said the company rented 30 homes in the San Francisco Bay Area just to answer one question: can a robot walk into an unfamiliar home and start doing chores right away? Bed-making was the hardest task, since the bed, bedding, and room were all new each time, requiring the robot to grab pillows, pull sheets, put items back in place, and smooth everything out; towel-folding was used to test handling of deformable objects. 56% is still a capability-demonstration figure, short of production readiness, but it specifically measures the exact thing past demos have most avoided — whether a system still works in an environment it was never adapted to. Source: 36Kr source

SoftBank agrees to acquire RAI Institute from Hyundai Motor Group · adjacent

Announced September 18, with the deal amount undisclosed; the transaction is under review by the Committee on Foreign Investment in the United States (CFIUS). RAI Institute, based in Cambridge, Massachusetts, was founded in 2022 by Marc Raibert as part of the arrangement when Hyundai acquired Boston Dynamics; Raibert also founded Boston Dynamics, which SoftBank owned from 2017 to 2021. This second institute founded by the same person is now returning to SoftBank, coming after Boston Dynamics' IPO was pushed back to after 2027 (previously reported) and after SoftBank's acquisition of ABB's robotics business last year (for roughly $5.38 billion). Source: The Robot Report source

Deep Blue Robotics enters bankruptcy liquidation, owing over 100 employees more than 23 million yuan · adjacent

Publicly disclosed employee claims show Deep Blue Robotics (Shanghai) owes more than 100 employees over 23 million yuan, of which more than 21 million yuan is wages and severance, with housing fund and individual social security contributions calculated separately. Shanghai's Pudong New Area Court accepted its bankruptcy liquidation case five months ago, and in August the Changzhou Tianning District Court also accepted the case for its 94%-owned subsidiary, Deep Blue Robotics (Changzhou). Founded in 2018, the company's parent, DeepBlue Technology (valued at 14 billion yuan, holding 96%), had spread its product line broadly, including the "Little Blue Whale" indoor cleaning robot, UV disinfection robots, delivery robots, lawn-mowing robots, and passenger AGVs. Around 2021, it discussed a disinfection robot partnership with LG, publicly citing procurement interest for tens of thousands of units worth over 700 million yuan in orders, then built a factory in Changzhou and, in 2023, a base in Shangqiu, Henan, planning for an annual capacity of 20,000 units and 1 billion yuan in annual output value. Pencil News, citing an insider, said the cause of bankruptcy was "insufficient cash flow." Source: 36Kr source

Huarui Zhipu (Chinese robotics chip maker) launches Praxis One, packing a robot's "brain" and "cerebellum" into a single 3nm chip · hardware ⚠️ company-sourced figures

At an industry conference in Hangzhou on September 17, Guangzhou-based Huarui Zhipu launched its Praxis platform and its first two products. Praxis One, based on MediaTek's Genio Pro 5100 3nm SoC, combines the "brain" (perception, reasoning, decision-making) and the "cerebellum" (motion control) on a single chip, with a combined 120 TOPS of compute, wide-voltage 12V-to-60V power supply, and power draw of 3.6W idle and roughly 7.2W to 9.5W at peak load. Wu Naishen, MediaTek's IoT business marketing director, said the SoC uses an all-big-core CPU paired with an 8th-generation NPU 890, delivering 106 TOPS and supporting 16-channel vision sensing and Wi-Fi 7. Praxis Nano, based on the Genio 720, takes a lightweight "cerebellum" approach, with the company claiming it can cut costs by over 40% compared to similar solutions. COO Zhao Zhihao said robotics teams typically spend 9 to 12 months on chip adaptation, PCBA design, system porting, and stability validation, and the platform aims to eliminate that step; in a live demo, the same YOLOv8 model went from 3.1 frames per second on GPU to 114 frames per second on the NPU using the company's compiler toolkit. The same event also unveiled a data collection system and three scenario-specific robots, including the 1.85-meter industrial heavy-load humanoid S1 Zilong, which has Praxis One built in. Source source

Jijia Vision signs strategic cooperation framework agreement with Yangtze Automotive Electronics · world-model

The cooperation remains at the framework and joint-evaluation stage, planned in three phases: jointly building an industrial data collection and training platform, with Jijia Vision providing world models and embodied foundation models to be trained and deployed on real production lines at Yangtze Automotive Electronics; once the first scenario is validated, it would be replicated across Wenzhou's auto parts supply chain. Yangtze Automotive Electronics' business spans 10 countries, serving over 60 OEMs and having participated in the development of over 500 vehicle models. Jijia Vision has previously trialed logistics sorting with China Post's Wuhan branch and signed cooperation agreements with FAW Mould and Longsheng Technology. Source: Netcarshow source