Skip to content
#Robot models and learning
Jun 17Wed
  1. Sanctuary AI57

    Sanctuary AI extends Physical AI strategy to industrial robots, achieving 99.5%+ success rate on wire plugging task

    Sanctuary AI announced a task success rate of over 99.5% on a wire plugging production task at a global Tier 1 automotive supplier, with a cycle time of 2.54 seconds, validated against the customer's real-time production benchmarks. The company has shifted its strategy to deploying Physical AI on existing commercial robot hardware rather than waiting for humanoid hardware to reach mass production; the system handles flexible cables on a moving conveyor, with throughput matching the customer's production line level.

Jun 15Mon
  1. 石麻笔记 · 微信公众号67

    In Conversation with Jianyuan Wang, First Author of VGGT: The Future of 3D Reconstruction Foundation Models and Robotic Spatial Perception

    This article is an interview with Dr. Jianyuan Wang, first author of VGGT, which won the CVPR 2025 Best Paper Award, and of VGGT-Ω, a CVPR 2026 Best Paper finalist. He is currently pursuing a PhD through a joint program between the Oxford Visual Geometry Group and Meta AI.

  2. New Scientist Robots56

    Reporter visits Physical Intelligence: progress and controversy on the general-purpose robotics path

    Physical Intelligence is developing a general-purpose control brain that can work across many kinds of robots, using a vision-language-action (VLA) model to teach robots tasks such as making coffee, folding clothes, and peeling vegetables, rather than focusing on a single machine as Tesla, Boston Dynamics, or Amazon do.

Jun 13Sat
  2. Chelsea Finn52

    CHORUS uses a single VLA policy to control multi-robot teams for decentralized collaboration

    Chelsea Finn's team fine-tuned a VLA so that a single policy can control any robot in a robot team, enabling multi-robot collaboration without explicit communication (the system described in the quoted tweet is named CHORUS). Results show that this method matches or outperforms training separate models for each robot or training a unified centralized model, and it scales easily to large teams. Paper: https://arxiv.org/abs/2606.12352

Jun 12Fri
  2. 石麻笔记 · 微信公众号62

    In conversation with Huang Wenlong: PointWorld and 3D world models

    Investor Shi Ma speaks with Stanford PhD student Huang Wenlong (advisor Li Feifei) about his PointWorld work and 3D world models. PointWorld uses 3D point flows to unify state and action representations, mapping different robots' joint actions into point cloud changes via URDF to address inconsistent actions across embodiments; training used the DROID and BEHAVIOR datasets, with annotations relying on FoundationStereo, VGGT, and CoTracker3.

  3. Sergey Levine38

    Flow reversal steering guides high-level decision-making in diffusion-based VLAs

    Flow reversal steering can "steer" diffusion-based vision-language-action models (VLAs) through high-level actions, such as high-level actions inferred from vision-language models (VLMs). It also lets us run reinforcement learning (RL) in the diffusion noise space, with high-level reasoning guiding exploration: think about the task first, then practice! 👇

Jun 10Wed
  1. Sergey Levine37

    A New Trick for RL Training of Flow Matching Policies

    Diffusion (or flow) makes for excellent policies, but training them with RL is notoriously hard: BPTT is unstable, and RL over diffusion blows up the horizon. In our new paper, we show how to optimize a flow matching actor with a "weird little trick"—using the identity matrix to "approximate" the Jacobian of the flow denoising process. 👇

Jun 9Tue
  1. Holiday Robotics60

    South Korea's Ministry of Science and ICT Launches Physical AI Leading Technology R&D Project, Investing 34 Billion Won Over Two Years to Advance World Model Localization

    South Korea's Ministry of Science and ICT held a launch briefing for the Physical AI Leading Technology R&D Project in Seoul on June 9, planning to invest 34 billion won over two years to jointly develop a Korean-style world model and domestic simulator with 10 industry-academia-research institutions including LG Electronics, aiming to raise robots' final action success rate by more than 20 percentage points compared with not using a world model, and planning to achieve an MVP within one year and eight months and conduct validation at manufacturing and logistics sites.

    Why it matters: The South Korean government has launched a project to localize core physical AI technologies; readers can learn about its investment scale, target metrics, and industry-academia-research division of labor.

Jun 8Mon
  1. PKU EPIC Lab64

    KPGrasp: A Scalable Keypoint Flow Matching Method for Dexterous Grasp Generation

    PKU EPIC Lab proposes KPGrasp, a dexterous grasp generation method that combines an all-Euclidean 3D hand-keypoint parameterization with a Transformer flow-matching model to learn grasp priors from large-scale data, without contact losses or contact-based test-time optimization.

    Why it matters: The paper presents a new contact-loss-free grasp generation method and specific results on two benchmarks, letting readers compare it with existing methods and assess its approach.

Jun 7Sun
Jun 5Fri
  1. PKU EPIC Lab61

    PKU EPIC Lab proposes LIMMT, a data-centric study for humanoid motion tracking, accepted to ICML 2026

    PKU EPIC Lab released LIMMT (Less Is More for Motion Tracking), which it describes as the first data-centric study for physics-based simulated humanoid motion tracking. The method defines motion data quality along three dimensions: physics feasibility, diversity, and complexity. Training on less than 3% of AMASS achieves better tracking performance than training on the full dataset, and the team also cleaned web-sourced mocap data. The paper has been accepted to ICML 2026; see arXiv:2606.06953.

    Why it matters: The paper presents a three-dimensional data quality framework and results from training on a small amount of data, offering a methodological reference for researchers working on humanoid motion tracking.

Jun 4Thu
  1. 1X Technologies61

    1X Establishes World Model Lab, Sam Sinha Appointed Head of World Models

    1X announced the formal establishment of the 1X World Model Lab, focused on large-scale pretraining of embodied world models, with the goal of accelerating the realization of fully autonomous humanoid robots. Sam Sinha, formerly a founding research scientist at Luma AI, has joined as Founding AI Researcher and Head of World Models.

    Why it matters: 1X officially outlined the World Model Lab's data approach and its full-stack in-house development stance, giving readers insight into its technical path toward autonomous humanoids.

Jun 3Wed
Jun 1Mon
  1. Agile Robots/思灵机器人28

    Agile Robots gains early access to NVIDIA Cosmos 3 for physical AI simulation

    Agile Robots has obtained early access to NVIDIA Cosmos 3, with early simulation tests showing strong capabilities. The company will integrate Cosmos 3 for customers, offering new options for simulating, testing, training, and deploying physical AI, and is exploring the Cosmos-based NVIDIA Video Augmentation Skill for automated data curation, generation, and evaluation.

May 29Fri
  1. 达纳灵动 · X16

    Dyna Robotics co-founder to speak on robot learning at CVPR

    Robotics is shifting from impressive demos to scalable learning systems. 🤖 DYNA co-founder and chief scientist @JasonMa2020 will speak at the #CVPR workshop on scalable robot learning systems. 🙌 If you work on embodied AI, data engines, VLM/VLA, sim-to-real transfer, ML systems, or evaluation, join the discussion! We will share how to build robots that do more than a single task—robots that keep learning, adapting, and scaling. https://scalebot-workshop.github.io/

  2. ROBOTIS · X61

    ROBOTIS AI Sapiens humanoid robot integrates NVIDIA Kimodo text-to-motion model

    ROBOTIS has integrated NVIDIA's Kimodo, a text-to-motion diffusion model, into the open humanoid robot AI Sapiens, allowing natural language prompts to directly generate full-body humanoid motions that are executed on real hardware. The motions are driven by DYNAMIXEL Q (QDD) actuators, with no motion capture suit and no frame-by-frame programming required. The full development source code will be released as open source, and the full video is available at https://youtu.be/2xAJzQ8V9p8 .

May 28Thu
  1. Chelsea Finn55

    EXPO-FT uses reinforcement learning post-training to bring π0.5 to a 30/30 success rate on all 8 manipulation tasks

    Through reinforcement learning post-training, EXPO-FT brought π0.5 to a 30/30 success rate on all 8 manipulation tasks tested, using an average of only 19 minutes of reinforcement learning data per task. The method targets challenging manipulation tasks, including routing string lights and inserting a power connector to light the string, potting a billiard ball, and placing a flower into a wine bottle. It is released as an efficient, reliable, open-source VLA fine-tuning method; the paper and video are at https://pd-perry.github.io/expo-ft/.

May 18Mon
May 15Fri
  1. 富唯智能26

    2025 Guangzhou AI Innovation List | Fuwei Intelligence's GRID large model selected as an outstanding AI+Manufacturing case

    Fuwei Intelligence's embodied-intelligence industrial robot solution, built on its self-developed GRID large model, was selected for the 2025 Guangzhou Artificial Intelligence Innovation and Development List as an "Outstanding AI+Manufacturing Case" and was also named to the "Startup Potential List." The list is assessed under the leadership of the Guangzhou Municipal Science and Technology Bureau, with review covering five major dimensions including technological originality and implementation feasibility. The solution features multimodal perception, autonomous environment recognition, real-time intelligent decision-making, and adaptive motion control, and targets industrial scenarios such as intelligent sorting, flexible assembly, and material transport.

May 13Wed
  1. Noble Machines51

    Noble Machines Breaks Down Its Industrial Autonomous Learning Stack: Whole-Body Control, Task-Level Learning, and Runtime Error Correction

    Noble Machines published an article introducing its three-layer robot learning stack for industrial autonomy, along with on-site test results from GTC 2026: over four days, the same model simultaneously drove multiple Moby robots to complete two autonomous tasks at the NVIDIA booth and its own booth.

May 11Mon
  1. Lumos Robotics/鹿明机器人60

    Lumos Robotics completes several hundred million yuan A1 and A2 funding rounds, with Mitsubishi Electric leading both

    Lumos Robotics has completed several hundred million yuan in A1 and A2 funding. The A1 round was led by Mitsubishi Electric, with participation from Puhua Capital and Wuzhong Financial Holdings, among others. Mitsubishi Electric continued to lead the A2 round, with participation from Hundsun Electronics, Haigao Group, and Kunshi Investment, among others. The funding will be used for embodied large-model R&D and scenario expansion. After the funding, the company's cash reserves reached several hundred million yuan. Its flexible quality-inspection embodied intelligence solution, co-created with Mitsubishi Electric, has been put into trial use at Mitsubishi Electric's factory in China. Its self-developed Lumos Nexcore physical AI engine and bodyless data-collection system Lumos FastUMI Pro support industrial deployment and data feedback.

    Why it matters: The original text provides details on the funding structure and industrial partnerships, allowing readers to understand its industrial deployment path and data-collection technology layout.

May 8Fri
  1. 石麻笔记 · 微信公众号74

    Interview with Jianlan Luo: How LWD Enables Robots to Keep Learning After Deployment

    This article is an interview with Dr. Jianlan Luo about his new work LWD (Learning While Deploying), which enables general-purpose robot policies to continuously perform online reinforcement learning from their own experience during real-world deployment. LWD combines offline RL pretraining with online RL fine-tuning to unify offline and online optimization objectives; the value function is trained separately with a VLM and uses both positive and negative samples (roughly over 60% positive and over 30% negative), and deployment data flows back into the training pipeline in about 40 seconds on average.

  2. Figure AI62

    Figure AI shows two Helix-02 humanoid robots collaborating to tidy a bedroom

    Figure AI demonstrated two humanoid robots equipped with Helix-02 collaborating to tidy a bedroom within two minutes, completing tasks such as opening a door, hanging clothes, putting away headphones, closing a book, taking out trash, pushing a chair under a desk, and making a bed together. The two robots run the same learned vision-language-action (VLA) policy, with no shared planner, message passing, or central coordinator; each infers the other's intent from its actions via its own cameras. The company says that to its knowledge this is the first demonstration of multi-humanoid collaborative mobile manipulation completed by a single learned neural network directly from pixels to actions.

    Why it matters: The official demonstration shows two humanoid robots driven by the same VLA system collaborating to tidy a bedroom, giving readers insight into an approach to multi-robot collaboration without a shared planner.

May 6Wed
  1. Genesis AI · X50

    Genesis AI releases GENE-26.5 robot foundation model

    Genesis AI has released GENE-26.5, calling it the first "robot brain" on the path toward human-level capability. It is a full-stack approach combining a robot-native foundation model, a 1:1 humanoid robotic hand, non-invasive data-collection gloves that capture motion, force, and touch, and a simulator. The model is trained jointly on language, vision, proprioception, touch, and action, and was tested on a set of self-designed tasks; the demonstration was fully autonomous, at 1x speed, using a single model with the same weights.

  2. Genesis AI · YouTube32

    Genesis AI's GENE-26.5 Autonomously Performs Straw Flip at 1x Speed

    Genesis AI's robotic foundation model GENE-26.5 autonomously completed a continuous one-handed fine manipulation sequence: unwrapping a straw, flipping the straw with one hand, dropping it into a cup, and then handing the drink to a person. The entire sequence was executed autonomously and filmed at 1x speed, not teleoperated. The demo showcases fluid one-handed dexterous manipulation across multiple small, precise actions.

May 2Sat
  1. Singularity Hub62

    EPFL team lets robots with different designs share skills, paper published in Science Robotics

    Researchers at EPFL published a method in Science Robotics that lets multiple robots with different designs complete the same task from a single human demonstration. The method does not rely on AI; instead, it analyzes the physical properties of three-revolute-joint robotic arms, marks safe regions within each robot's range of motion, and classifies three-joint robots into six categories based on shared physical constraints, writing those constraints into each robot's program. The authors call this kinematic intelligence.

Apr 30Thu
  1. LimX Dynamics67

    LimX Dynamics Open-Sources FluxVLA Engine, a Standardized VLA Engineering Foundation

    LimX Dynamics has officially open-sourced FluxVLA Engine, a standardized VLA engineering foundation for embodied intelligence that covers the full pipeline from data and training to simulation evaluation and real-robot deployment.

    Why it matters: LimX Dynamics has officially open-sourced its VLA engineering foundation, which lets readers understand how unified configuration and real-robot deployment workflows lower the barrier to research and development.

Apr 29Wed
  1. Figure AI64

    Figure 03 production ramps up: BotQ capacity rises 24-fold in 120 days to 1 unit per hour

    Figure announced that its BotQ factory has delivered more than 350 third-generation humanoid robots, Figure 03, with takt time improving from 1 unit per day to 1 unit per hour, a 24-fold throughput increase in 120 days. The line is driven by self-developed manufacturing execution software and covers more than 150 networked workstations. First-pass yield for completed units exceeds 80%, the battery line's first-pass yield is 99.3%, more than 9,000 actuators have been produced, and each robot undergoes more than 80 functional verification tests before leaving the factory.

    Why it matters: The official figures give specific production and quality-control numbers and explain how fleet scale supports the deployment of perception-conditioned whole-body control.

Apr 26Sun
Apr 22Wed
  1. New Scientist Robots76

    Sony AI table tennis robot Ace defeats professional players; research published in Nature

    Sony AI's autonomous table tennis robot Ace competed under the rules of Japan's professional table tennis league, defeating multiple elite and professional players; the research was published in Nature (DOI: 10.1038/s41586-026-10338-5).

    Why it matters: The article explains Ace's competitive record and three key technologies, and places it alongside the milestones of Deep Blue and AlphaGo, helping readers judge where physical sports competition currently stands.

  2. PsiBot/灵初智能75

    Lingchu Intelligence releases Psi-R2 and Psi-W0, open-sources 1,000-hour full-modality human hand manipulation dataset

    On April 10, Lingchu Intelligence released Psi-R2, pretrained on human data, and the Action-Conditioned World Model Psi-W0, and open-sourced the first 1,000 hours of a full-modality human hand manipulation dataset, with total data reserves reaching 100,000 hours.

    Why it matters: The original text provides the division of labor between the two new models and the open-source data entry point, so readers can assess the feasibility of using human data for robot training.