Skip to content
#Robot models and learning
Sep 17Thu
  1. Figure AI79

    Figure AI releases Helix 2.5, achieving zero-shot full-body autonomy in 30 unfamiliar homes

    Figure AI released Helix 2.5, which completed three long-horizon full-body behaviors—tidying a living room, folding towels, and making a bed—zero-shot in 30 Bay Area homes where no data had ever been collected, without collecting data or fine-tuning in the evaluation environments.

    Why it matters: The original article provides controlled comparison data on Index pretraining for zero-shot generalization and evaluation details, helping readers understand the actual level of humanoid robot pretraining transfer.

Sep 15Tue
  1. Leiphone Robotics53

    Xspark AI's Ding Wenbo on tactile intelligence: Robots need their own "spinal cord"

    In an interview, Xspark AI co-founder and chief scientist Ding Wenbo proposed that vision has already solved about 99% of robot perception problems, and that touch should form a low-latency, body-adjacent "spinal cord"-level model rather than being directly integrated into vision large models. He proposed a three-layer architecture of "slow brain VLM—fast brain VTLA+TWAM—spinal cord VTA," arguing that touch mainly serves collision-avoidance safety and post-training for dexterous manipulation, and noted that current bottlenecks lie in sensor consistency, tactile data representation, and unification across sensors, modalities, and embodiments.

  2. OFweek Robotics45

    EBKernel releases brain-inspired cognitive world model Cog-WM 1.0, exploring a new "root technology" path for China's embodied intelligence

    EBKernel released its brain-inspired cognitive world model Cog-WM 1.0 on September 14 at the Brain-Inspired Embodied Intelligence sub-forum of the Shanghai Pujiang Innovation Forum. It includes a navigation branch, Cog-WM Nav 1.0, and a manipulation branch, Cog-WM Manip 1.0, and uses memory, prediction, and active exploration to help robots understand the physical world.

  3. New Strategy Robotics54

    JoyIn releases Aether embodied intelligence model and questions OpenAI distillation

    On September 10, 2026, Suzhou-based JoyIn released the Aether embodied intelligence model, aimed at humanoid robot body control and real-world task execution. JoyIn says the model was trained on about 200 hours of human video and zero real-machine data, has roughly 4 billion parameters, exceeds 90% success rate on first attempts at some tasks, can run on bodies from Unitree Robotics, AgiBot and others, and that it plans to partially open up the model's capabilities.

  4. 小鹏机器人66

    XPeng XOS 6.3.0 OTA Deploys Second-Generation VLA Upgrade, Adds InfiniVLA and X-Foresight World Model

    XPeng XOS 6.3.0 deploys a second-generation VLA upgrade via OTA, restructuring the model architecture around the time dimension: InfiniVLA lets historical visual information continuously participate in current judgments, with the vehicle side selecting 30 seconds of historical context alongside streaming autoregressive inference.

    Why it matters: The article explains concretely the trade-offs in co-designing model architecture and deployment under onboard compute constraints, making it a useful engineering reference for deploying large models on limited hardware.

  5. Chelsea Finn56

    OM-1 Releases First Robot Foundation Model, Zero-Shot Generalizes to Desktop Arms, Industrial Arms, and Humanoid Robots

    OM-1 has released its first robot foundation model, which can zero-shot generalize to desktop robotic arms, industrial robotic arms, and humanoid robots. The model learns directly from human operation data without using teleoperation or robot data, claims near-human-level dexterity and efficiency, and supports multi-robot collaboration. Chelsea Finn reposted and congratulated the relevant authors on the release.

Sep 14Mon
  2. Dexmal 原力灵机 · 微信公众号73

    Force Spirit DM0.5 Tops All Four RoboColiseum Leaderboards

    Force Spirit announced that DM0.5 has topped all four RoboColiseum leaderboards (instruction following, spatial understanding, perturbation adaptation, and general manipulation), making it the only model on the leaderboards to rank first in every evaluation.

    Why it matters: The original article lists DM0.5's specific scores on multiple leaderboards and its inference latency reduction data, which can be used to assess the practical effectiveness of its pretraining approach.

Sep 11Fri
  1. PKU EPIC Lab64

    PKU EPIC Lab proposes UniPart for zero-shot language-driven 3D part segmentation

    PKU EPIC Lab proposes UniPart, a feed-forward cross-modal 3D Transformer conditioned on CLIP text embeddings that uses free-text phrases to select functional parts on point clouds, achieving zero-shot language-driven 3D part segmentation.

    Why it matters: The paper presents a text-driven 3D part segmentation method and a self-built dataset, and validates transfer to real-world language-conditioned grasping.

  3. 机器人daily · 微信公众号56

    Tashi Intelligence launches general embodied AI model AWE3.7, demonstrating cross-embodiment and cross-scenario generalization

    Tashi Intelligence released AWE3.7, an iterative version of its AWE series general embodied AI model, claiming cross-environment, cross-embodiment, and cross-task generalization, with the same model deployable on dual-arm collaborative robots, wheeled chassis, and humanoid robots. The materials show its performance in industrial scenarios, including parts sorting, tape wrapping, network cable insertion, phone boxing, screw tightening, drive belt grasping, and block sorting, as well as household tasks such as clothes folding, whiteboard wiping, and smoothie making, and claim that model-level changeover time has been compressed to the hour level.

Sep 10Thu
  1. 云启资本60

    Yunqi Podcast Roundtable: Three Embodied AI Founders Discuss World Models, Scaling, and Deployment Bottlenecks

    During the World Robot Conference, Yunqi Capital co-hosted a themed roundtable with SEE Fund, which was included in the podcast Attent!on. Three founders—Xu Huazhe of PokeBot, Li Qiang of Quantum Dynamics, and Liu Nianqiu of Feimo Yi Technology—shared their views on world models, scaling, and deployment.

    Why it matters: The three entrepreneurs approached the topic from models, hardware systems, and mobility respectively, specifically breaking down where the bottlenecks in embodied intelligence deployment lie and where they diverge on scaling efficiency.

  2. Skild AI79

    Skild AI annual recurring revenue tops $100 million, outlines deployment-first robotics strategy

    Skild AI announced that its annual recurring revenue (ARR) has surpassed $100 million, of which $50 million was confirmed within 10 months of starting deployments. It has more than 60 paying customers, covering material handling, delivery, inspection, security, food service, and warehouse, factory, and data center scenarios; about 90% of revenue comes from manipulation tasks, with mobility accounting for 10% (of which AMR solutions account for 4%).

    Why it matters: Beyond the company's self-reported revenue milestone, it also offers concrete lessons on how deployment in turn shapes model research, which is useful for understanding the relationship between real-world deployment and research.

Sep 9Wed
  1. 帕西尼 · 微信公众号46

    PaXini x BeingBeyond: Co-building a New Ecosystem for Embodied Intelligence Data and Models

    PaXini AI has entered into a partnership with BeingBeyond, a general-purpose embodied foundation model company. BeingBeyond becomes a PaXini Inside ecosystem partner, and the two parties will work around the Being-H series of models to advance embodied foundation model training, full-modality data production, and POC validation in real-world scenarios.

  2. Leiphone Robotics42

    Wanxun Technology releases NOVA2.0 flexible embodied brain, already supporting over 10 million real-world operations across 40+ industry scenarios

    Wanxun Technology has released NOVA2.0, its second-generation flexible embodied brain, built on a touch-native multimodal hierarchical architecture based on the concept of "body intelligence." It supports native one-brain-multiple-forms, 1B+ dynamic parameter combinations, and 20ms ultra-fast trajectory generation. The architecture has completed over 10 million extreme-condition operations across 40+ real industry scenarios, including real estate and construction, power and energy, and transportation. It can adapt to flexible single-arm, dual-arm, dexterous hand/bionic human and other body configurations, and can generalize to new scenarios with as little as 1% of the startup data volume required by conventional approaches.

Sep 8Tue
  1. 开普勒人形机器人 · 微信公众号44

    Kepler Bumblebee humanoid robot smart coffee cabin opens co-construction partnership slots

    In the Kepler Bumblebee smart coffee cabin, the Bumblebee humanoid robot handles the entire coffee-making process. Powered by a self-developed VTLA (vision-tactile-language-action) omni-modal large model that coordinates and controls a professional coffee machine, it produces a cup in about 45 seconds, runs continuously for 8 hours on a 1-hour charge, and supports 24/7 unattended operation.

Sep 7Mon
  1. RLCN强化学习研究 · 微信公众号44

    Roundtable Discussion | RLChina Paper Seminar Episode 168 Live Preview (2026.09.08)

    Episode 168 of the RLChina Paper Seminar will be livestreamed on Bilibili and Tencent Meeting on September 8, 2026, from 16:00 to 16:30, hosted by Hu Zican, a doctoral student at Nanjing University, with master's student Xie Xinyi presenting the paper "Look Before You Leap: Distilling Tree Search into Action Evaluation for Frozen VLA Models."

Sep 6Sun
  1. 银河通用机器人 · X38

    Galbot releases humanoid robot gesture model RoboGesture

    What if humanoid robots could gesture as naturally as they speak? 🤖 Introducing RoboGesture — enabling humanoid robots to listen, respond, and make meaningful gestures. RoboGesture completes the "listen → respond → gesture" loop in about 2 seconds, bringing us closer to natural human-robot interaction. Semantic alignment. Real-time. Safety-aware. 📍 #ECCV2026 · Poster Session 2 · ExHall #178 🗓️ Sep 10 · 16:30–18:30 CEST 🔗 https://robogesture.github.io #Galbot #EmbodiedAI #HumanoidRobot

Sep 5Sat
  1. PKU EPIC Lab68

    GIF: An agentic generation framework for interactive and functional object compositions for robot learning

    PKU EPIC Lab proposes GIF, an agentic generation framework for robot learning that reframes the problem as disentangled reconstruction followed by relative pose recovery. CoGen combines 2D and 3D generative models to produce instance-disentangled meshes and coarse initial poses; GPRM refines the relative poses under joint geometric and physical guidance; and a VLM verifier selects the candidate that best matches the structured specification.

    Why it matters: The paper splits interactive and functional object composition generation into disentangled reconstruction and relative pose recovery, and provides a benchmark of eight contact-geometry classes along with collision-rate figures, which can help readers assess this route to simulation data generation.

Sep 4Fri
  1. 中科慧灵 · 微信公众号65

    CASIA and others propose World4RL: diffusion world model combined with reinforcement learning to optimize robot manipulation policies, real-world success rate rises to 93.3%

    The Institute of Automation of the Chinese Academy of Sciences, Zhongguancun Academy, and the CASBOT team propose World4RL, a policy optimization framework for robot manipulation that combines a diffusion world model with reinforcement learning, enabling robot policies to be optimized end-to-end in imagined trajectories generated by the model without additional real-robot interaction.

    Why it matters: The paper turns the world model from a trajectory-prediction tool used at test time into a reinforcement learning training environment, giving readers a concrete path to understanding how the cost of real-robot interaction can be reduced.

  2. 智动未来49

    Zhido Future Launches "Zhizi World" World Model, Enabling Xiaozhi S3 to Understand the Real World

    At the 2026 XAIR Expo Embodied Intelligence Industry Expo, Zhido Future unveiled a technical preview of its differentiable world model "Zhizi World" (PGD World Model) and demonstrated validation progress on its embodied intelligent robot Xiaozhi S3. "Zhizi World" reconstructs particle-based scenes from short multi-view videos and can edit parameters such as object pose, material, mass, and friction to predict the physical outcomes of actions. S3 Xiaozhi features bionic dual arms, lidar-vision fusion navigation, and multimodal voice interaction; on-site demonstrations included scooping popcorn, filling cups and delivering them, and self-service retail shopping.

Sep 3Thu
  1. ETH Zurich Robotic Systems Lab38

    euROBIN final event showcases 15 AI robots at the European Parliament in Brussels

    The euROBIN project's final event was held in Brussels, where 15 AI robots joined European researchers, industry, and policymakers to showcase Europe's robotics strength and discuss how to turn research leadership into global industrial competitiveness. The ETH Zurich Robotic Systems Lab team, as a project member, presented its robot Transformer.

Sep 2Wed
  1. PKU EPIC Lab79

    ZETA: A Controlled Study of Zero-Shot Cross-Embodiment VLA Transfer for Tabletop Manipulation

    PKU EPIC Lab published the ZETA study on arXiv (arXiv:2609.02546), proposing a unified definition of zero-shot cross-embodiment VLA transfer and a controlled benchmark covering 14 held-out target embodiments, distinguishing strict zero-shot transfer from pretrain-exposed transfer. Experiments show that local end-effector state-action representations, source embodiment diversity, and auxiliary co-training improve transfer by about 15, 18, and 7 percentage points, respectively, and that adding only 5% target-embodiment data during pretraining raises average target-embodiment progress by 13.4 percentage points.

    Why it matters: The paper provides a unified definition of zero-shot cross-embodiment transfer and a controlled benchmark, allowing readers to understand the actual benefits of different representation and data strategies.

  2. RLCN强化学习研究 · 微信公众号72

    Seven questions on DreamerV3: general reinforcement learning with fixed hyperparameters via a world model

    The RLChina "Seven Questions" column examines DreamerV3, a general reinforcement learning algorithm proposed by Google DeepMind and the University of Toronto. Based on a world model, the algorithm first trains an actor-critic on imagined trajectories, and with a single set of fixed hyperparameters it matches or exceeds domain-specific algorithms across 8 major categories and more than 150 tasks.

  3. Stial/斯帝尔科技33

    Stial Uses AI Flexible Wet Grinding Solution to Repair Complex Defects in Titanium Alloy Plates

    Stial has released an AI flexible wet grinding solution for repairing scratches and pits in titanium alloy plates. The process includes the SeeV AI perception system to identify defects, the NextBrain™ AI brain to plan grinding paths and process parameters online, and PolishX flexible force control to stabilize contact force, with wet cooling simultaneously carrying away grinding heat. Under the specified range and test conditions, the system's contact force control accuracy can reach ±0.5% FS. The solution transforms experience-driven grinding into a controllable process driven by data and algorithms, and the related technology has been applied in multiple high-end manufacturing fields such as shipbuilding, automotive parts, and 3C electronics.

  4. 机器科学48

    RoboScience launches REX G1 wheeled humanoid general-purpose manipulation robot

    RoboScience has launched REX G1, a wheeled humanoid general-purpose manipulation robot equipped with its self-developed Visics general embodied large model built on the VLOA architecture, targeting scenarios including logistics, factories, retail, and homes. REX G1 has 22 degrees of freedom, can carry 10 kg with both arms, supports ±0.1 mm repeat positioning accuracy, is only 540 mm wide and can enter 0.75-meter narrow passages, and relies on 2070 TFLOPS of on-device compute to complete the perception, decision, and control loop.

Sep 1Tue
  1. 小鹏机器人63

    XPeng's second-generation VLA large model gets its first major upgrade, XOS 6.3.0 to debut globally on the G9L

    XPeng held a sharing and experience day themed "TIME," announcing the first major upgrade to its second-generation VLA large model. The new XOS 6.3.0 version will make its global debut on the XPeng G9L, with the core aim of moving AI from 3D spatial understanding to 4D spatiotemporal understanding.

    Why it matters: The original breaks the upgrade into three specific changes: contextual memory, reasoning approach, and world model. Readers can use these to understand how the shift from spatial understanding to spatiotemporal understanding is realized.

  2. Keenon Robotics/擎朗智能46

    KEENON Robotics showcases KOM 3.0 at the 2026 World Robot Conference, letting robots "think before they act"

    At the 2026 World Robot Conference, held August 19–23, KEENON Robotics showcased KOM 3.0, a VLA architecture for the service industry, which it says integrates a latent-space world model. On site, the XMAN-R1 autonomously completed the full process from picking up a cup to extracting coffee; two humanoid robots collaborated under unified scheduling to wash and fold clothes; and together with the T10 delivery robot and C40 cleaning robot, they demonstrated "general-purpose + specialized" collaborative service. In the demonstrations, the robots completed long-horizon tasks end-to-end autonomously, with no teleoperation.

Aug 28Fri
  2. 机器人daily · 微信公众号60

    Lingxi Zhihong's ROSS Harness Wins National Top Three in Humanoid Robot Games Industrial Scenario Competition with Fully Autonomous Mode

    In the "Assembly and Loading Station" event of the 2026 World Humanoid Robot Games Industrial Scenario Competition, Lingxi Zhihong's robot completed three tasks fully autonomously: transporting and shelving, picking and loading, and sub-millimeter assembly of 18 engine valves, scoring 84, 65, and 21 points respectively. After a 10-point special deduction, it finished with a total score of 160 points, ranking second only to two teams under a leading industry robotics company in the enterprise rankings, making it the only startup to break into the national top three on its first participation.

Aug 27Thu
  1. Pollen Robotics · YouTube64

    Pollen Robotics opens pre-orders for Microduck, a 25 cm reinforcement learning robot

    Pollen Robotics has opened pre-orders for Microduck at $399 (excluding taxes and shipping), with first deliveries targeted before Christmas 2026. The 25 cm-tall robot is built for reinforcement learning, with 15 degrees of freedom, a front camera, an 8x8 LiDAR, two IMUs and a full sensor suite, and supports designing behaviors in simulation and transferring them to the real robot.

    Why it matters: It packs the full loop from simulation training to real-world deployment into a small, pre-orderable complete robot, letting readers gauge the barrier to building their own reinforcement learning workflow.

  2. Pollen Robotics · YouTube62

    Pollen Robotics launches Microduck, a $399 tiny robot that can be trained in simulation

    Pollen Robotics has launched Microduck, a 25 cm tall, 780 g, 15-degree-of-freedom tiny robot, priced at $399 for pre-order (excluding taxes and shipping), with first deliveries targeted before Christmas 2026. It comes in four colorways: Cream, Graphite, Lavender, and Sky.

    Why it matters: The original article provides hardware specs, pricing, and an open-source simulation training entry point, which together indicate the learning value of this small robot for developers.

  3. 自变量机器人 · 微信公众号75

    X Square Robot releases WALL-SS next-scale autoregressive world model

    X Square Robot has released WALL-SS, a next-scale autoregressive world model that uses a coarse-to-fine, layer-by-layer refinement generation architecture, breaking away from the traditional diffusion-model approach of generating video in a single segment. In testing, its action-following score was 0.29, the only nonzero score among the compared models; the overlap between generated trajectories and action instructions rose from 0.251 to 0.539; streaming inference can reach 60 seconds of footage without collapse; and in closed-loop policy consistency tests across 6 tasks and 5 training stages, the success-rate agreement between virtual and real-world tasks reached 93%.

    Why it matters: The source gives three quantitative metrics and the architecture design, which readers can use to assess how feasible the world model is as a policy testbed.

  4. PsiBot/灵初智能69

    Lingchu Intelligence completes over $100 million new funding round, with industrial capital including Tuopu Group and Chery Ruicheng Fund participating

    Lingchu Intelligence announced the completion of a new funding round of over $100 million. Investors include Tuopu Group, Ruicheng Fund under Chery Holding, Lens Technology, 37 Interactive Entertainment, Wuhu Municipal Investment Holding Group, and Fosun Chuangfu, while existing shareholder Zhuhai Technology Industry Group made an excess additional investment. The proceeds will focus on embodied world model R&D and commercial deployment in logistics and advanced manufacturing scenarios. The company's self-developed dual-system architecture comprises the Psi-R2 manipulation policy model and the Psi-W0 action-conditioned world model, and it plans to expand the scale of human data collection and accelerate the layout of data collection facilities.

    Why it matters: Lingchu Intelligence officially announced the completion of the funding and stated that the funds will go toward world model R&D and commercialization in logistics and manufacturing scenarios, giving readers insight into its industry-finance resource layout.

  5. 北京智在无界科技有限公司33

    BeingBeyond Wins First Prize at HICOOL 2026 Global Entrepreneurship Competition

    At the opening ceremony of the HICOOL 2026 Global Entrepreneur Summit on August 26, BeingBeyond stood out from 10,209 participating projects to win the competition's first prize, with founder Lu Zongqing accepting the award on stage. Founded in May 2025, the company focuses on training embodied general foundation models using large-scale human video data. In July 2026, it released the Being-H0.8 general dexterous manipulation model, which unifies vision, touch, action, and future state changes into a single latent space.

Aug 26Wed
  2. OFweek Robotics52

    Leju Hosts Three Dialogues at WRC: Robotics Industry Shifts from Making Movements to Doing Things Reliably

    The author reviews three dialogues hosted jointly by Leju and Ant Lingbo at WRC, covering whether embodiments are easy to learn, whether data is easy to use, and whether models are easy to evaluate, with 12 guests in total. Key viewpoints include that hardware is becoming a precondition for model capability, that Ant Lingbo's real-world data exceeds 100,000 hours and covers more than 20 configurations, that Noitom says about 20% of data plays a core role in model behavior, and that evaluation should guide adjustments to data distribution; on this basis, the author argues that industry attention is shifting from making a single movement to doing a task reliably and consistently.