Open Artifacts Latest coverage

FineART: Fine-Grained Annotated Robotic Trajectory Dataset

FineART is a bimanual manipulation dataset co-released by Scale AI and Hugging Face LeRobot. The data was collected and annotated by the Physical AI unit at Scale AI. FineART is a bimanual manipulation dataset where every episode comes with dense, human-written subtask annotations. It has 40,543 teleoperated episodes (1,718 hours), 533,913 subtask labels, and 151 tasks, all collected on a Trossen Stationary AI (ALOHA-style) bimanual robot.

Global First Robot Task Native 3D World Model: Cross-Dimensional UNITAS Enables Robots to Learn Physical Intuition, Sweeping Multiple Benchmarks

UNITAS is the world's first robot task-native 3D world model, achieving cross-perspective and cross-body generalization capabilities by unifying observation, action, and environmental changes within a three-dimensional physical space. The model performs exceptionally well in multiple benchmark tests, including LIBERO, LIBERO-Plus, and RoboTwin 2.0. With just 1.7B parameters, it outperforms larger parameter models such as Fast-WAM (6B) and Motus (8B).

Embodied Intelligence Begins to Evolve Systems: RoboFoundry Enables Robot Autonomous Evolution

RoboFoundry is a new framework for embodied agents that treats the entire system as a self-evolving policy. By freezing the base model and allowing the surrounding system to evolve, it achieves significant improvements in task success rates across benchmarks like EmbodiedBench, RoboMemArena, and LIBERO-PRO. The system enables zero-shot transfer and online evolution on real robots, with results showing substantial performance gains compared to baselines.

NovaSplat: Model-Free 6D Object Pose Estimation via 3D Gaussian Splatting Onboarding and Refinement

NovaSplat is a model-free pipeline for 6D object pose estimation that trains a 3D Gaussian Splatting model from a short smartphone video. It serves as both a template for detection and a differentiable backbone for pose refinement, outperforming model-based methods on a household-object benchmark and model-free competitors on an industrial benchmark with textureless and specular surfaces. A faster inference variant, NovaSplat-Lite, is also introduced.

PartiGen: Part-Aware Representations for Language-Conditioned Humanoid Motion Generation

This paper introduces PartiGen, a part-aware latent diffusion framework that integrates anatomical structure into whole-body motion representations.

When a Robot Reaches for a Cup, It's Moved Away: LIBERO-MAX Uses 224,000 Simulations to Test Dynamic Reliability

LIBERO-MAX is a study that evaluates the adaptability of robots in dynamic environments by designing paired tasks and mid-task changes to assess the performance of different strategies under environmental changes. The study found that the success rates of all tested strategies decreased, and the impact of different types of changes on the models varied significantly, indicating that robots need more robust perception and planning mechanisms to handle uncertainties in real-world scenarios.

USTC Team Puts a 'Reverse Thinking' World Model into Robots, Secures Millions in Funding in Just Three Months

Bai Ze Tong Jing is a less than six-month-old company based in Hefei, and its developed Causal World Model AWM enables robots to perform causal reasoning about environmental changes. AWM significantly improves data efficiency and task success rates by learning object, motion, and interaction patterns from unlabeled videos.

Agnes AI Opens Sources Agnes Harness, Enabling AI Agents to Connect with the Real World

Agnes AI officially opens sources its agent execution infrastructure Agnes Harness (AGH) and device interaction protocol MHS (Model Hardware Standard), providing developers with a unified toolchain that supports the continuous execution of AI Agents in real-world tasks and physical devices.

NVIDIA Releases Cosmos3-Nano Policy Trained on LIBERO-10 Data

NVIDIA has released a trained policy model for the Cosmos3-Nano robot, fine-tuned on LIBERO-10 data. The model achieves 94.4% success in standard scenes and 89.2% in perturbed scenes. It is part of a synthetic-data study and is loadable via the cosmos-framework policy server.

NVIDIA Releases Cosmos3-Nano Policy Model Trained on LIBERO-10 Tasks

NVIDIA has released a trained policy model for the Cosmos3-Nano robot, fine-tuned on LIBERO-10 tasks. The model demonstrates success rates of 93.6% in standard scenes and 87.1% in perturbed scenes, with training data and evaluation details provided.

BYD May Not Fully Believe in Humanoids, But Made One

The article discusses BYD's development of a humanoid robot, analyzing its design, technical specifications, and strategic positioning in the market. It highlights the company's focus on service-oriented robots for retail environments, emphasizing the balance between mechanical design, functional capabilities, and cost efficiency. The patent details and technical breakdown offer insight into the engineering challenges and design choices made for the robot's deployment in 4S stores.

RoboSpeak — WeChat (mp_account, tier=T2)

The article discusses the '觅蜂派' (MeiFeng Pai) platform, which allows ordinary people to earn money by performing everyday tasks like cleaning and cooking。

NVIDIA Releases Cosmos3-Nano Policy Model Trained on LIBERO-10

NVIDIA has released a robotics policy model, Cosmos3-Nano LIBERO-10, trained on the LIBERO-10 environment. The model demonstrates success rates of 93.4% in standard scenes and 82.9% in perturbed scenes, with performance evaluated across 10 tasks and 50 initial states. The model is available under the NVIDIA Open Model License.

LeRobot@LeRobotHF· 13:34SignalEditorial score7272

LeRobot's new streaming loader is designed for training on large-scale robot datasets, allowing efficient processing of 100 TB+ data without full local video copies. It reads episodes directly from Hugging Face or Buckets, prepares samples on-the-fly, and balances memory and throughput for varied training batches.

Tsinghua's Embodied Model Tops Global Rankings. Outperforming GPT-6 and NVIDIA Without Hacks or Extra Data

Xingdong Jiyuan's self-developed world action model VPP2 has topped the RoboDojo simulation leaderboard with an average success rate of 32.26% and a score of 39.26, outperforming models like GPT-6-Astra. VPP2 improves prediction and action generalization by decoupling video prediction from action learning through a phased training strategy.

LM-X: Explainable Vision--Language--Action Modeling via Progress, Event, and Uncertainty Prediction

LM-X is a new VLA foundation model that learns three directly supervised online signals—return-to-go (RTG), event-to-go (ETG), and heteroscedastic action-flow variance—to provide explainable control. Pretrained on over 20,000 hours of real-robot trajectories, LM-X achieves 74.1% success on RoboTwin2.0 tasks and outperforms GR00T N1.7 in both simulated and real-world settings. The model's signals track progress, anticipate motion, and detect high-error actions, establishing it as a transparent and effective VLA framework.

VioLA: Learning Generalist Humanoid Control Policies from Human Data

This paper presents VioLA, a generalist humanoid policy that predicts body and hand motion latents instead of joint commands. By mapping human and robot motion into the same latent space, VioLA enables zero-shot task execution on real robots, achieving 100% success in locomotion and 88.6% in manipulation without task-specific fine-tuning. The method is validated across multiple models and datasets.

A Practical Guide to Data Collection and Utilization in Embodied Intelligence

The article outlines the operations of a humanoid robot training center that generates high-quality data for embodied intelligence applications. It details the data collection process, the role of open-source datasets like RoboMIND, and the importance of standardized, high-value data in driving the development of autonomous robotic systems.

UniCross: Unified Cross-Skill Dexterous Manipulation Synthesis

This research introduces UniCross, a unified framework for dexterous manipulation that integrates four canonical skills—grasping, relocation, in-hand rotation, and in-hand translation—into a single relational task formulation. By modeling these skills with shared constraints and objectives, the approach achieves cross-skill compatibility, enabling distillation of a single policy without skill identifiers. The method supports long-horizon skill chaining and generalizes to unseen objects and hand morphologies, demonstrating that distinct manipulation skills can be instantiated from a common relational formulation.

Multi-Agent Coordination via Support-Preserving Distillation

This research paper addresses failure modes in offline multi-agent reinforcement learning (MARL) where standard flow-based teachers route noise samples to conflicting coordination modes. The authors propose Mode-Support Semi-Discrete Optimal Transport (MoSDOT), a method that summarizes multimodal replay into finite mode support with prescribed capacities and uses conditional semi-discrete optimal transport to assign each noise sample to a single mode before teacher training. Additionally, a shared-randomness variant is studied using a shared noise component at execution to expose the residual gap intrinsic to strict-product execution. On controlled diagnostics and offline MARL benchmarks, MoSDOT improves endpoint quality and routing consistency, particularly on datasets exhibiting multimodal joint behavior.