Skip to content
#Robot models and learning
Nov 3Tue
  1. 量子位65

    Xiaomi wins two Wuzhen Summit awards; humanoid robot completes 8-hour validation at automotive plant

    Xiaomi Group received the 2026 World Internet Conference Outstanding Contribution Award, and its "key technologies for a general-purpose robot manipulation foundation model for scaled deployment" received the Leading Technology Award. Xiaomi's humanoid robot took on an 8-hour shift for the first time at the center console side cover sequencing station in an automotive plant, achieving autonomous anomaly correction and system-level coordination with the production line; an Agent system built on the Xiaomi-Robotics-1 pretrained model is in trial operation at Xiaomi Youth Apartments, completing long-horizon task validation of over 30 consecutive minutes, covering 3 rooms and 50+ tasks.

  3. Ken Goldberg30

    KEC combines Fable and Astra for parameter tuning to achieve real2sim

    KEC combines Fable with Astra to rapidly tune complex simulation parameters via RSS, including geometry, dynamics, deformation, materials, lighting, textures, and camera extrinsics, so that the simulation matches a given input video. This work targets the capability gap in agentic robotics for extending to new environments, new objects, and new tasks.

  4. Guyuehome40

    Deep Vision Reinforcement Learning for Autonomous Quadrotor Flight: AirGym Control Hierarchies and Depth Observation Design Overview

    For quadrotor obstacle avoidance, AirGym incorporates depth perception, low-level control, and sensor noise into a single training pipeline, and outlines five control hierarchies, depth observation designs, and reward structures. The article also records the engineering finding that Blackwell GPUs cannot run Isaac Gym.

  5. Guyuehome52

    G1 Brunson: Humanoid Robot Between-the-Legs Dribbling Control Based on Deep Visual Reinforcement Learning

    G1 Brunson uses event-based rewards and asymmetric actor-critic training, bringing the depth camera directly into the observation loop during training to form a G1 whole-body dribbling policy. Reproducing the official weights locally in MuJoCo gave an average of 43.5 crossovers over 12 episodes with zero falls, matching the official results.

  6. Guyuehome37

    Dexterous Hand in Practice 10: Coordinate Transforms and Collisions—First Put Every Finger on the Same Map

    The 10th installment of the Dexterous Hand in Practice series focuses on coordinate transforms and collision detection, with the goal of unifying each finger onto the same map. The content covers a unified representation of multi-finger coordinate frames and collision handling, and is one part of this hands-on tutorial series.

  8. Guyuehome57

    WAM: Training a world model with quadruped robot locomotion data for navigation planning

    WAM trains a forward-dynamics world model on quadruped robot locomotion data collected under convex model predictive control, replaces hand-crafted steering laws with receding-horizon sampling planning, and combines A* global paths with dynamic obstacle extrapolation to achieve navigation and obstacle avoidance. The material states that all numerical values come from local reproduction experiments.

  9. 钛媒体60

    Stardust Intelligence's Lai Jie: Using AOS to Connect General Large Models with Embodied Models

    Stardust Intelligence founder Lai Jie said the company does not compete with general large models; instead, through Stardust OS (AOS), it packages perception, mobility, manipulation, and interaction into callable Meta capability packages and Skills, enabling models such as GPT and DeepSeek to divide the work and collaborate with its self-developed Lumo embodied foundation model.

  10. OFweek Robotics38

    Geek+ announces "24/7 Autonomous Intelligent Warehouse Plan" to build self-evolving unmanned warehouses in the physical AI era

    Geek+ announced the "24/7 Autonomous Intelligent Warehouse Plan," establishing embodied intelligence and full-warehouse intelligence as its next strategic direction, driving warehouses toward 7×24-hour unmanned autonomous operation and continuous evolution. The plan relies on end-to-end robotic solutions and a warehouse brain software platform comprising RMS, WES, and IOP, focusing on three areas: end-to-end process integration, strengthening intelligent closed loops, and large-scale deployment. Geek+ says it already has a global large-scale commercial deployment and localized operations system, but the autonomous capabilities in the plan remain development goals.

  11. 36Kr Embodied Intelligence69

    Sage Intelligent completes hundreds-of-millions-yuan Pre-B round to advance six-arm robots and industrial world model

    Shanghai Sage Intelligent Technology has completed a Pre-B funding round worth hundreds of millions of yuan, with investors including Jiantou Investment, Nantong Investment Management, and Yunqi Capital. The funds will be used for large-model R&D, iteration of multi-arm robot bodies, production line expansion, and overseas market development. The company has built a body matrix spanning single-arm to six-arm robots; its six-arm robot "Xuanjia" uses a spatially heterogeneous design, and the company has released an industrial world model, Xuanshu·SageNexus (WASM). Its customers cover four major industries—3C and semiconductors, automotive, energy, and food and pharmaceuticals—with more than 50 Global Fortune 500 or listed-company customers and overseas intended orders exceeding 500 million yuan.

  12. 36Kr Embodied Intelligence60

    Yicheng Innovation closes RMB 200 million Series C, plans to raise unmanned transfer vehicle capacity to 10,000 units

    36Kr has learned that Shenzhen Yicheng Autonomous Driving Technology Co., Ltd. (Yicheng Innovation) recently closed a RMB 200 million Series C round. Investors include Yueke Group, Shanda Investment, Anhui Guotou, and Foshan Venture Capital. The funds will be used for factory capacity upgrades, core technology iteration, and commercialization deployment.

  13. HuggingFace Daily Papers67

    H2RB co-trains with human demos and robot data to adapt VLA policies

    The study proposes H2RB, a human-robot co-training framework that adapts pretrained VLA policies on unpaired human and robot demonstrations through geometry-anchored alignment and a decoupled-decoder tokenizer. In real-world tasks, object transfer performance improved from 8.0 to 51.8, and the skill transfer score reached 62.7, while the co-training baseline completely failed to transfer human demonstration skills. Out-of-distribution transfer under object placement changes also improved from 6.0 to 45.3.

  15. Tech Xplore Robotics71

    Seoul National University proposes RISE framework to complete multi-stage humanoid tasks using separately collected locomotion and manipulation data

    A research team at Seoul National University has proposed RISE (Responsibility-Induced Specialized Experts), a VLA framework that can collect demonstration data for locomotion and object manipulation separately and then learn jointly from them, enabling a robot to execute multi-stage tasks consecutively without full multi-stage demonstrations.

  17. 机器人全球资讯 · 微信公众号70

    AgiBot releases embodied Agent framework Genie Master; Genie G2 works and evolves in a convenience store

    AgiBot has released its embodied Agent framework Genie Master. A Genie G2 running the framework, when deployed in a real convenience store, does not rely on human demonstrations or debugging; it can identify the causes of failures on its own, adjust strategies, and accumulate experience, improving continuously on tasks such as packaged-food packing, picking, and restocking. AgiBot says this is the industry's first validation of Physical RSI (embodied recursive self-improvement) outside simulation in a real business scenario, and it plans to arrange 6 Genie G2 units on November 2 for a continuous 24-hour livestreamed convenience-store operation test.

  18. APPSO · 微信公众号12

    New iPad Air reportedly to switch to OLED display; two Chinese astronauts to reach the Moon before 2030; Claude restricts Opus 5.5 from recreating games

    Bloomberg reporter Mark Gurman says Apple plans to update the entry-level iPad, iPad Air, and iPad Pro in the first half of 2027, with the iPad Air moving to an OLED display and high-end iPad Pro models possibly adding a vapor chamber to improve cooling.

  20. 机器人前瞻 · 微信公众号55

    Jiben Intelligence completes two consecutive funding rounds, raising nearly $10 million to build dexterous manipulation cerebellum

    Jiben Intelligence announced that it has completed two consecutive funding rounds totaling nearly $10 million (approximately RMB 67 million). The rounds were led by Alphaist Partners and Yaotu Capital respectively, with Jieli Angel and the actual controller of a components industry player participating as follow-on investors.

  21. 图灵人工智能 · 微信公众号40

    Tsinghua Professor Deng Zhidong: Physical AI Has Not Yet Had Its ChatGPT Moment; World Models Are Key

    Tsinghua University professor Deng Zhidong believes physical AI as a whole is still in an early exploratory stage and has not yet had its own ChatGPT moment; robots being able to run and dance does not mean they can enter homes to care for the elderly. He proposes that the key to a breakthrough lies in world models: letting AI rehearse consequences and costs before executing actions, combined with tasks and human intent, with evaluation criteria being generality, task success rate, and generalization ability. He also suggests connecting cloud, edge, and devices through mobile communications and deploying shared computing power at the edge to ease cost and power consumption pressures on device-side computing.

  22. 移动机器人产业联盟 · 微信公众号38

    Geek+ announces "24/7 Autonomous Intelligent Warehouse Plan" to build self-evolving unmanned warehouses in the physical AI era

    Geek+ announced the "24/7 Autonomous Intelligent Warehouse Plan," with embodied intelligence and whole-warehouse intelligence as its two main technical thrusts, driving warehouses toward 7×24-hour unmanned autonomous operation and continuous evolution. The plan relies on end-to-end robotic solutions, a warehouse brain software platform (including RMS, WES, and IOP), and a global large-scale delivery system, focusing on three directions: connecting the entire process, strengthening the intelligent closed loop, and achieving large-scale implementation. Its Gravity unified embodied intelligence framework and Gravity 4D embodied world model are used to improve autonomous execution capabilities in complex physical tasks and abnormal scenarios.

  23. 甲子光年 · 微信公众号38

    In conversation with Zhipu, Galbot, Guanglun Intelligence, and Deepexi: Will large language models eventually eat embodied models?

    At the "Qinghe Night Talk" roundtable, Zhipu CEO Zhang Peng, Galbot CFO Sun Dengke, Guanglun Intelligence's Yang Haibo, and Deepexi Technology's Zhao Jiehui discussed whether large language models will eat embodied models. Sun Dengke drew an analogy to fast and slow systems, arguing that embodied models handle S1/S0-level manipulation, planning, and environmental understanding, and may require a smaller S2 parameter scale than language models; Zhang Peng argued that intelligence will ultimately converge. Guanglun Intelligence positions itself as simulation-based physical AI data and evaluation infrastructure, while Deepexi Technology has launched DeepWorks, which can be tried for free.

Nov 2Mon
  1. 高工咨询 · 微信公众号40

    Geek+ Announces "24/7 Autonomous Intelligent Warehouse Plan" to Build Self-Evolving Unmanned Warehouses in the Physical AI Era

    Geek+ has released the "24/7 Autonomous Intelligent Warehouse Plan," advancing 24/7 unmanned autonomous operation and continuous evolution of warehouses along two main technical lines: embodied intelligence and whole-warehouse intelligence. Its end-to-end robotic solution relies on the self-developed Gravity unified embodied intelligence framework and the Gravity 4D embodied world model, while the warehouse brain platform consists of three major modules: RMS, WES, and IOP. The plan focuses on three directions: connecting the entire process, strengthening the intelligent closed loop, and achieving large-scale deployment.

  2. 机器之心 · 微信公众号74

    Zhejiang University and TeleAI propose GR-BC, using graph search to select effective actions from robot trial and error

    The team led by Chunhua Shen at Zhejiang University, together with the China Telecom Artificial Intelligence Research Institute (TeleAI) and Chenjia Bai's team at Gamma Intelligence, proposed GR-BC, which builds multiple attempts at the same task into a graph, uses reverse Dijkstra search to estimate the remaining steps from each state to success, and then weights action chunks accordingly to train a vision-language-action model.

  4. 机器人产业应用 · 微信公众号58

    LLMs Enter Embodied AI: Competing for Physical Data or the Next-Generation Intelligent Gateway

    The article analyzes the different paths taken by large-model companies such as OpenAI, Google, Alibaba, and DeepSeek in entering embodied AI, arguing that their primary driver is their own growth and narrative needs, the second is defensive positioning, and that data and gateway are long-term goals not yet realized.

  6. CVer · 微信公众号60

    NTU and other teams present MemSkill at NeurIPS 2026, letting agents learn and evolve memory skills

    A research team from Nanyang Technological University and other institutions presented MemSkill at NeurIPS 2026, packaging memory operations into learnable, reusable, and evolvable Memory Skills, optimizing skill selection through task feedback, and revising and expanding the skill library based on failure cases.

  7. 人形大讲堂 · 微信公众号62

    AgiBot proposes Worken robot energy-efficiency unit, discloses CEL collective experience learning validation results

    At its 2026 Innovation Day, AgiBot's Peng Zhihui proposed the robot energy-efficiency unit Worken, defined as the number of active robots × autonomous working hours × output efficiency ratio, to measure how much work robots actually do. At the Longqi factory, 8 robots worked a cumulative 528 hours with an output efficiency ratio of 0.7, for a total of 2,956.8 Worken; at Guangzhou Metro, 120 robots worked 240 hours with an efficiency ratio of about 0.5, for a total of 14,400 Worken.

  8. 深蓝AI · 微信公众号57

    DeepBlue Embodied Intelligence reviews six embodied AI approaches: modular planning, skill orchestration, imitation learning, reinforcement learning, VLA, and world models

    DeepBlue Embodied Intelligence examines the debate over embodied intelligence approaches along six main technical lines: modular planning, skill orchestration, imitation learning, reinforcement learning, VLA, and world models, drawing on 12 highly cited surveys from the past decade. The article notes that each approach answers the question of what robots should learn from, that the classic problems have not disappeared, and that what has changed is that the technical units for solving them are growing larger. It also mentions an industry estimate that general-purpose robots would need 100 million hours of data to reach a 70%–80% baseline success rate, and that more than 100 open-source embodied foundation models appeared worldwide in the past 18 months.

  9. 雷科技58

    Six AgiBot Genie G2 Robots Enter Shanghai Convenience Store for 24-Hour Continuous Operation Challenge

    On November 2, six AgiBot Genie G2 humanoid robots entered an operating convenience store in Xuhui, Shanghai, working 24 consecutive hours in roles including greeter, barista, stocker, and cleaner, under online supervision from viewers across the internet. The robots relied on real-time visual inventory checks to return items to their places on disorganized shelves, stopping before people and then going around them, and routing around boxes. Through Genie Master and the embodied execution protocol EEP, they stored successful and failed task trajectories in long-term memory to feed back into motion strategy optimization.

  10. 北京理工大学 · 微信公众号36

    CCTV features BIT swarm intelligence research group: "BIT Zhixingzhe" humanoid robot and quadruped robot dog unveiled

    The swarm control research group at the School of Mechatronical Engineering, Beijing Institute of Technology, presented platforms including the "BIT Zhixingzhe" humanoid robot and a quadruped robot dog on CCTV's program "Amazing, We Are 'Science' People." At the 2026 Beijing Yizhuang Humanoid Robot Half Marathon, "Zhixingzhe" completed the full 21-kilometer course, achieving zero falls and zero interruptions across more than ten types of complex terrain including urban roads, wetland parks, steep slopes, and near-90-degree sharp turns, and won second place for "Best Gait." The research group also independently developed Sigma, a simulation system for single-drone and swarm algorithms, which supports multi-scenario, multi-weather simulation training.

  11. Guyuehome60

    Diffusion Policy: Learning Robot Manipulation Policies via Action Diffusion

    Using the Push-T tabletop pushing task as a vehicle, this piece fully reconstructs the mathematical formulation of action diffusion policies, covering cosine-schedule noise addition, the noise-prediction training objective, FiLM conditioning injection, and sliding-window execution. The author also uses public pretrained weights to complete a real inference evaluation across ten seeds.

  13. Guyuehome19

    SoftCon: Octopus swimming control based on bionic muscle actuation and soft-body dynamics

    The SoftCon framework builds an octopus swimming control pipeline from tetrahedral soft-body dynamics, muscle constraints that contract along fibers, and traveling activation waves generated by keyframe matrices and propagation filtering, using PPO to learn coordination rather than reconstruction. The framework originates from SIGGRAPH Asia 2019.

  14. Guyuehome63

    InsertAnything Evaluated in Isaac Lab for Precision Peg-in-Hole Insertion Policies with Fingertip Force Feedback

    Official weights of InsertAnything's force-conditioned insertion policy were evaluated in Isaac Lab: at 0.1 mm clearance, baseline success rates for two hole types were 96–100%. With a 5 mm initial hand offset, success rates were unaffected, but with a 3 mm hole-position observation deviation, the hexagonal hole success rate dropped to 60%.

  15. Guyuehome57

    G1 Dribble: End-to-End Reinforcement Learning for Humanoid Robot Soccer Dribbling Control

    G1 Dribble uses end-to-end reinforcement learning to achieve humanoid robot soccer dribbling; the walking task's weights are not adjusted, and velocity commands drive both the body and the ball. The author tested it locally in MuJoCo-Warp and recorded eighteen cloud sessions, reporting touch rhythm, the distance band for re-touching the ball, and failure modes, including that a centered start steps over the ball without touching it, a first touch occurs only when offset to the side, and at 1.5 meters per second it knocks the ball away directly.

  16. Guyuehome58

    HACMan++: Spatially Anchored Motion Primitives for Long-Horizon Robotic Manipulation

    HACMan++ targets long-horizon manipulation for six-dimensional object rearrangement, decomposing actions into three parts: selecting a primitive, anchoring a placement point on the point cloud, and continuous parameters, with unified decision-making via a per-point, per-primitive evaluation map. On this basis, the framework achieves closed-loop chaining of pushing, grasping, moving, and placing.

  17. Guyuehome57

    OpenCR-MuJoCo: Discretized MuJoCo simulation of tendon-driven continuum robots

    OpenCR-MuJoCo presents a discretized simulation method for tendon-driven continuum robots in MuJoCo, modeling the continuous backbone as a rigid chain with elastic joints, where bending is produced by tendon length differences. The work uses SoRoSim reference data and the official demo to validate static accuracy, release oscillation, and closed-loop tracking.