Skip to content
#Robot models and learning
Oct 13Tue
  1. OFweek Robotics54

    Fangqi Technology completes eight-figure angel round led by TUSStar

    Fangqi Technology, a Tsinghua-affiliated robot brain company, has completed an eight-figure angel funding round led by TUSStar, just four months after its registration and establishment. The company does not build hardware bodies; instead, it trains robot skills using ordinary human video footage combined with world models, and claims that overall processing cost is 80% lower than traditional teleoperation, time consumption is reduced to 30%, and action execution success rates are on par with traditional approaches. It has begun joint research and real-world testing with a five-star chain hotel management group in Jiangsu, and is supplying its robot brain system to a robot hardware manufacturer in Shenzhen.

  2. 国科投资 · 微信公众号56

    Astribot completes RMB 1 billion new funding round at RMB 15 billion valuation, four rounds in six months

    Astribot has completed a new funding round of RMB 1 billion at a valuation of RMB 15 billion, led by Shenzhen Investment Holdings' Donghai Investment. The funds are planned for model iteration, talent expansion, commercialization, and production capacity buildout. The company claims to be the world's first to achieve mass production of rope-driven AI robots, with capacity raised to the 10,000-unit level; its next-generation humanoid robot T1 began shipping in June 2026, bringing high-precision manipulation robots into the RMB 90,000 price range, and its foundation model Lumo-2, released in July, ranked first among participating models in the combined performance across three dimensions in more than 22 real-robot evaluations.

  4. Dexmal 原力灵机 · 微信公众号56

    Force Machine DM0.5 showcases three task types at the CIIF: connector insertion quality inspection, industrial insertion, and flexible cutting

    Force Machine presented its self-developed DM0.5 general-purpose embodied brain at the AXERA booth at the 2026 China International Industry Fair, completing three task types: connector insertion quality inspection, Insert Anything general-purpose industrial insertion, and dexterous-hand flexible cutting, with model inference performed on the AXERA high-compute platform.

  5. 云启资本66

    Stardust Intelligence completes new RMB 1 billion funding round at RMB 15 billion valuation

    Stardust Intelligence recently completed a new funding round of RMB 1 billion, reaching a valuation of RMB 15 billion. The round was led by Toushe Donghai (Shenzhen Investment Holdings), with participation from Hongtai Fund, Ninghai Xingtaaihe Fund, and Shanheng Capital, among others, and existing shareholder Xianmu increased its investment. This is its fourth funding round in the past six months; its Series B valuation first exceeded RMB 10 billion in June 2026 and rose to RMB 15 billion three months later.

    Why it matters: It reviews the pace of the company's four funding rounds in six months and details of its full-stack architecture, which can be used to understand the commercialization progress of the rope-driven approach.

  6. OFweek Robotics68

    Wang Xingxing raises embodied intelligence forecast; Liang Wenfeng reassesses robotics strategy

    The article analyzes how Wang Xingxing, at the China International Industry Fair, shortened his most optimistic estimate for the tipping point of embodied intelligence from two to three years to one to two years, and admitted he had not believed in AI enough. Reports from Phoenix News Technology and others say Liang Wenfeng has been meeting intensively with embodied intelligence entrepreneurs; DeepSeek has already made a strategic investment in Unitree and plans to lead a funding round for Winverse AI, raising tens of millions of dollars at a post-money valuation of about 400 million dollars.

  7. OFweek Robotics63

    Oula Wanxiang Completes Four Funding Rounds in Half a Year as Zhou Shunbo Bets on Home Embodied-Intelligence Robots

    On October 12, 2026, home embodied-intelligence company Oula Wanxiang completed a angel-plus round of several hundred million yuan, jointly invested by Chengyuanxiang Fund, Xinneng Venture Capital, Xizhou International, Topband, Hony Capital, and others. Since its founding in March 2026, the company has completed four consecutive funding rounds. Founder Zhou Shunbo was the first head of Huawei's Embodied Intelligence Innovation Lab. The company has chosen a mobile base plus dual-arm approach, with products focused on being "cultivable"; it is initially serving Maker users, has reserved more than 1,000 seed users, has a team of about 70 people, and plans to advance small-scale validation of its first-generation product in the first quarter of 2027.

  9. New Strategy Robotics28

    SIASUN exhibits at the 26th China International Industry Fair, showcasing DUCO embodied intelligent robots and intelligent manufacturing solutions

    SIASUN is exhibiting intelligent robot products and intelligent manufacturing system solutions at the 26th China International Industry Fair, held in Shanghai from October 12 to 16, at Booth A239 in Hall 7.1H. The embodied intelligence exhibition area features the BR-MR73A full-body wheeled humanoid robot for die-bonding equipment loading and unloading and multi-view vehicle lock inspection, the BR-DR73 half-body humanoid robot coffee latte art workstation, as well as the OneHub multi-robot collaboration system and the Songyi wheeled humanoid robots Agile and Power.

  10. OFweek Robotics58

    Unitree's Wang Xingxing Says He Underestimated AI's Ability to Build Robots Over the Past Decade

    At the 26th China International Industry Fair, Unitree founder Wang Xingxing said his biggest mistake over the past decade was not believing in AI enough. He said the release of GPT-6 greatly improved embodied intelligence practitioners' understanding, and that Unitree tested it on various robots with good results. He believes a breakthrough could come within one to two years if things move quickly, and that GPT-7 next year could directly solve the embodied intelligence problem.

  12. Physical Intelligence62

    Physical Intelligence opens Pi API beta, offering π model training and inference

    Physical Intelligence has opened the Pi API beta, allowing developers to upload data, fine-tune π models on the company-managed compute and infrastructure, and then run inference through a hosted gateway; the current interface already offers more data-efficient fine-tuning and improvement capabilities.

    Why it matters: Pi API brings data upload, fine-tuning, and hosted inference into one ready-made pipeline, helping readers weigh building their own machine learning infrastructure against calling the API directly.

Oct 12Mon
  1. OFweek Robotics67

    Oula Wanxiang completes hundreds of millions of yuan angel+ round, four funding rounds in half a year

    On October 12, Oula Wanxiang announced the completion of an angel+ round of hundreds of millions of yuan, with participation from Chengyuanxiang Fund, Xinneng Venture Capital, Topband, Hony Capital and others; the company has completed four funding rounds within half a year of its founding. The company proposes a cultivable approach, using a mobile chassis plus dual robotic arms form factor, initially iterating for maker developers, and partnering with Haokang Home to obtain home-service demonstration data; in September it released two self-developed AI models, OLASem and OLAGeo, focused respectively on semantic understanding and 3D spatial perception, attempting to build a data loop.

  2. OFweek Robotics62

    Geek+ Takes 20% Stake in Beijing Shenyi via 50 Million Yuan Strategic Investment

    On October 9, 2026, Hong Kong-listed Geek+ announced a 50 million yuan strategic investment in Beijing Shenyi, a nine-month-old company with no revenue yet, acquiring a 20% stake at a post-investment valuation of 250 million yuan. The transaction is a related-party transaction and is exempt from independent shareholder voting. Beijing Shenyi proposes an L4E+E4L self-evolution paradigm, emphasizing continuous parameter iteration through interaction with the real environment after robot deployment. It has completed 0.1-millimeter-level industrial flexible insertion work on site at the World Robot Conference in China, with the key learning process taking only half a day.

  3. Guyuehome50

    MyoLeg: Measured Biomimetic Walking Control Based on Musculoskeletal Dynamics and Neural Feedback

    The author tests MyoLeg's official pretrained policy and a spinal reflex baseline, shifting the walking control target from torque to 80 neural excitation channels, involving activation dynamics, force-length-velocity relationships, and moment arms. The article presents quantitative results on gait mechanics, impact recovery, muscle weakness, and fatigue, and summarizes lessons learned in reproducing the reflex circuit.

  4. Guyuehome53

    Any2Track: Multi-Expert Policy Distillation for General Motion Tracking in Humanoid Robots

    Any2Track addresses the difficulty of a single reinforcement learning policy handling vastly different motions such as dancing, fighting, and getting up from a fall. It uses a two-stage approach: first training specialized teachers by cluster, then using DAgger to aggregate the teachers' behaviors into a single student network, and it presents G1 simulation comparison results.

  5. Guyuehome34

    TinyWAM Source Code Walkthrough: Data Generation, Visual Encoding, and the Full Flow Matching Pipeline

    This TinyWAM source code walkthrough reviews the core implementation of this lightweight World Action Model, covering data generation, visual encoding, Flow Matching, model training, and the closed-loop execution pipeline. The article explains each stage alongside the source code to help readers understand the technical principles and engineering implementation of world action models.

  8. Guyuehome41

    Daily Project Hands-on Sharing: ST-WAM Source Code Walkthrough

    ST-WAM is a model for robot action generation. This article presents a hands-on analysis based on its source code, introducing the implementation of dual-space future modeling, CAIR history retrieval, and cross-branch attention, and outlining the training and inference pipelines to explain how semantic information and temporal information jointly serve robot action generation.

  10. 松灵机器人55

    What is Mobile ALOHA? How a dual-arm mobile robot learns tasks through teleoperation

    Mobile ALOHA is a low-cost, open-source dual-arm mobile manipulation system proposed by a Stanford research team. It adds a mobile base and a whole-body teleoperation interface to the ALOHA dual-arm teleoperation platform, synchronously collecting vision, robotic arm state, and base motion data through teleoperation, then training policies with imitation learning methods such as behavioral cloning. Among these, ACT predicts a segment of an action sequence at a time.

  11. Pollen Robotics · YouTube64

    Pollen Robotics launches Microduck Arena virtual competition arena

    Pollen Robotics has released Microduck Arena, a virtual arena where Microduck robots race, train, and compete without the need for physical robots. Users train policies with reinforcement learning and publish them to their own Hugging Face accounts to enter challenges; each submission is evaluated uniformly by the Arena, and results can be compared across players. Published policies can be downloaded as .onnx files to watch replays and learn from the best runs.

    Why it matters: It connects policy training, unified evaluation, and replay-based learning into a closed loop, letting readers learn how to train and compare reinforcement learning policies without physical robots.

  12. Robohub54

    Humanoid robots can run and box, but folding clothes is the real test

    Wai-keung Fung of Cardiff Metropolitan University writes that at the World Humanoid Robot Games 2026, Tiangong Ultra completed the 100 meters in 8.64 seconds, faster than Usain Bolt's 9.58-second world record, but soft manipulation tasks such as folding clothes and doing laundry are a better test of robot intelligence. The article notes that fabric shapes vary widely and skills are hard to transfer directly between robots with different configurations and sensors, and that Open X-Embodiment is exploring this problem by gathering over a million real trajectories from 22 kinds of robots.

Oct 11Sun
  1. IEEE RAS51

    XFlowMP Wins CASE 2026 Best Conference Paper Award; Khang Nguyen Discusses Learning for Robot Motion Planning

    Khang Nguyen and co-author Minh Nhat Vu received the CASE 2026 Best Conference Paper Award for XFlowMP, a framework that uses Schrödinger bridges to transform random noise into trajectory distributions, generating feasible motions with velocity and acceleration based on start and end configurations.

Oct 10Sat
  1. GitHub Robotics Topic54

    TorchRL v0.14.1 released, covering environment resets, collectors, and LLM workflow fixes

    PyTorch's official reinforcement learning library TorchRL has released the v0.14.1 maintenance version, fixing multiple issues in environment resets, collectors, replay buffers, planning, training, logging, and LLM workflows, and introducing an advance notice of device deprecation for v0.17.

  2. Leiphone Robotics72

    Columbia's Yunzhu Li team presents progress and ecosystem of open-source tactile sensor FlexiTac at IROS 2026

    At the tactile-themed workshop at IROS 2026, Columbia University's Yunzhu Li team presented progress roughly a year and a half after the open-sourcing of FlexiTac, an open-source flexible piezoresistive tactile sensor. The sensor uses a layered structure of FPC and force-sensitive film and can be assembled in 3 minutes after parts are prepared; the team open-sourced the complete BOM, assembly instructions, and reading software and integrated it into the LeRobot framework. Simulation adaptation requires calibrating only 3 hyperparameters of a spring-damper system.

  3. Xbotics具身智能实验室 · 微信公众号69

    Discrete Forcing: Two-Step Action Generation via Discrete Action-Guided Continuous Denoising

    Discrete Forcing proposes first using a single discrete forward pass to determine the overall structure of an action, then mapping it back to continuous space as the generation starting point for Flow Matching, completing fine correction through a single continuous forward pass, so that the entire action generation requires only 2 NFE.

  4. Chelsea Finn51

    Chelsea Finn shares Mulligan paper: the more reliable the robot, the less efficient DAgger data collection becomes

    Chelsea Finn points out that as robots become more reliable, DAgger-style data collection efficiency drops sharply; for example, a 90% success rate means a human spends more than 90% of the time as a passive observer. The Mulligan paper she cites proposes the key idea that failures often come from similar initial states, so learning can be focused on these states; in deployment, the robot performs the task and the human intervenes when needed, with each collection round focusing on the initial states where the robot fails. Across 3 real tasks and 2,550 blind tests, it outperformed uniform collection under the same budget. Paper: https://arxiv.org/abs/2610.05882 .

  6. The Robot Report59

    Boston Dynamics details new four-finger dexterous hand design for Atlas

    Boston Dynamics introduced its redesigned four-finger dexterous hand for the humanoid robot Atlas, featuring 13 degrees of freedom, direct drive, the ability to handle objects over 100 pounds, and a design intended for mass production. Director Alberto Rodriguez said the hand drops the pinky finger, uses 13 fully enclosed identical actuators with no fragile cables spanning joints, and is designed for high-fidelity simulation and sim-to-real reinforcement learning; the team has tested industrial tasks such as power tool use, cable handling, and screw finding, but has not yet deployed it across the entire robot fleet.

Oct 9Fri
  4. Robotics 24/762

    Sereact dual-arm robotic system handles Zalando returns in Germany and Poland

    Sereact, Zalando, and CEVA Logistics announced the launch of AI-automated returns handling at fulfillment centers in Greven, Germany, and Świebodzin, Poland. Powered by Sereact's physical AI platform Cortex, the dual-arm robotic system can autonomously identify, grasp, and sort previously unseen fashion returns without product-specific training.

  5. Vulcan Robotics · X27

    Multimodal large models may replace robot-specific policies

    No training needed, no sim-to-real gap, just a prompt and a vision. Sentdex points out that every task in robotics today is a complex combination of locomotion, manipulation, vision, planning, and more, and fails once it deviates from training data or simulation characteristics. As a result, many demos are conducted only under perfect lighting and carefully placed objects, and never land under unknown real-world conditions—every deployment is a unique edge case. In the example, a general-purpose multimodal large language model (DSV4.1F) completed a robot task it was never specifically trained for, using an unseen camera, unlearned servo motors/motors, an environment it had never handled, and a morphology it had never been tested on. The video is at 10x speed and is currently slow, but this is already the slowest state at this level of intelligence. Sentdex says that the more he uses general-purpose multimodal large models, the more he realizes that specialized methods such as ACT and VLA may be mere supporting acts and distractions, and will eventually be swallowed by general models that do robotics incidentally. He does not think it will always be large language models, but the fact that large models can complete the most challenging robot tasks shows that... —Sourccey functionality is continuously improving.

Oct 8Thu
  1. The Robot Report61

    Helm.ai signs $70 million in commercial contracts for physical AI foundation models within 12 months

    Helm.ai announced it has signed $70 million in commercial contracts over the past 12 months, with customers including global automotive OEMs, Tier 1 suppliers, and industrial automation companies, and said the company is nearing break-even. Its foundation models use an unsupervised "deep teaching" method that hierarchically decouples perception and decision-making to improve data efficiency, and the technology has been used in SAE L2 to L4 autonomous driving, mining, construction, and delivery drone projects.

  2. IEEE Spectrum Robotics63

    ETH Zurich turns off-the-shelf robotic hand into a small robot that walks on its fingertips and operates autonomously

    Amirhossein Kazemipour's team at ETH Zurich's Soft Robotics Lab fitted a commercial robotic hand from Wuji Technology with an 80-gram backpack (containing a battery, IMU, and Raspberry Pi Zero) and used reinforcement learning in simulation to train a policy that lets it crawl on its fingertips.

  3. 机器人daily · 微信公众号61

    The 4D world model direction backed by Fei-Fei Li is now running successfully in factories through Yingshen Intelligence

    The article uses AMD's announcement of an all-stock acquisition of Fei-Fei Li's World Labs for about $8.2 billion as an entry point to compare Yingshen Intelligence's approach of betting on native 4D world models since its founding. Yingshen Intelligence's self-developed "Yingshen 360" capture system uses 4 to 6 consumer-grade RGB cameras to achieve millimeter-level, 25 frames per second real-time 4D data generation, reducing overall hardware costs by more than 90%; in the 4D foundation world model "Pojie S1", the GAF Gaussian Action Field achieves an average success rate more than 10 percentage points higher than the previous best method on dynamic manipulation tasks.