RoboTwin-Platform/RoboTwin

[ICML 2026] RoboTwin 2.0 Offical Code Repo

What it solves

RoboTwin provides a scalable platform for developing and benchmarking bimanual (dual-arm) robotic manipulation. It addresses the need for high-quality training data and robust evaluation environments to improve how robots handle complex tasks involving two arms, focusing on visual robustness, language diversity, and cross-embodiment performance.

How it works

RoboTwin functions as a data generator and benchmark. It provides a large-scale dataset of over 100,000 pre-collected trajectories and tools for users to collect their own data with domain randomization. The platform integrates with XPolicyLab for policy serving and evaluation, allowing users to run evaluations locally across multiple GPUs or via a split deployment where a policy server is remote and the simulator is local.

Who it’s for

It is designed for robotics researchers and developers working on bimanual manipulation, Vision-Language-Action (VLA) models, and generalizable robotic policies.

Highlights

  • Large-scale Dataset: Includes over 100,000 pre-collected trajectories for immediate training.
  • Strong Domain Randomization: Supports custom task configurations and domain randomization to improve model robustness.
  • Flexible Evaluation: Supports single-task, multi-task multi-GPU scheduling, and remote policy-server deployment.
  • Broad Compatibility: Provides conversion scripts for LeRobot datasets and integrates with StarVLA for VLA development.

Related

  • Project
    NVlabs/RoboLab

    RoboLab is a high-fidelity simulation benchmark built on NVIDIA Isaac Lab that provides over 100 manipulation tasks to evaluate and compare generalist robot policies.

  • Project
    Robbyant/lingbot-vla

    A pragmatic Vision-Language-Action (VLA) foundation model pre-trained on 20,000 hours of real-world robot data to enable high-performance robotic control across multiple platforms.

  • Project
    RoboDojo-Benchmark/RoboDojo

    RoboDojo is an open‑source benchmark that evaluates robot‑manipulation policies in both Isaac Sim and on real robots. It provides 42 simulated and 18 real tasks, parallel simulation, and a one‑command workflow that talks to policies hosted via XPolicyLab, producing scores for a public leaderboard.

  • Project
    Jianghanxiao/PhysTwin

    PhysTwin is a research codebase that reconstructs physics‑informed digital twins of deformable objects from RGB‑D video, provides interactive simulation, analysis tools, and example pipelines for robot planning and batched simulation.

  • Project
    RLinf/RLinf

    A scalable reinforcement learning infrastructure designed for Embodied and Agentic AI, providing a unified backbone for training VLA models and deploying them across simulators and real-world robots.