RoboTwin-Platform/RoboTwin

[ICML 2026] RoboTwin 2.0 Offical Code Repo

What it solves

RoboTwin provides a scalable platform for developing and evaluating bimanual (dual-arm) robotic manipulation policies. It addresses the challenge of obtaining high-quality training data and testing the robustness of robotic agents across different environments, embodiments, and tasks.

How it works

RoboTwin functions as a data generator and benchmark. It provides a simulation environment where users can collect trajectories or download over 100,000 pre-collected trajectories. The platform integrates with XPolicyLab for policy serving and evaluation, allowing for single-task or multi-task evaluation across multiple GPUs or via a remote policy server. It also supports domain randomization to improve the visual robustness of trained policies and provides tools to convert data into LeRobot format for compatibility with various ML policies.

Who it’s for

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

Highlights

  • Large-scale Dataset: Includes over 100,000 pre-collected trajectories for immediate training.
  • Robust Evaluation: Supports multi-task, multi-GPU scheduling and remote server deployment for policy testing.
  • Domain Randomization: Built-in support for generating diverse data to enhance visual and environmental robustness.
  • Broad Compatibility: Integrated with XPolicyLab and provides conversion scripts for LeRobot datasets.

Related

  • Project
    RoboDojo-Benchmark/RoboDojo

    A unified sim-and-real benchmark for evaluating generalist robot manipulation policies across 60 tasks and five capability dimensions.

  • Project
    RoboVerseOrg/RoboVerse

    A unified platform and benchmark for scalable robot learning that integrates multiple simulation frameworks and datasets into a single interface.

  • Project
    NVlabs/RoboLab

    A task-based evaluation benchmark for robot manipulation policies built on NVIDIA Isaac Lab, featuring 120+ tasks with automated success detection.

  • Project
    open-gigaai/giga-world-1

    An open-source framework for training and deploying robot world models to evaluate robot policies through controllable video generation.

  • Project
    RLinf/RLinf

    RLinf is an open‑source, PyTorch‑compatible reinforcement‑learning infrastructure for large‑scale embodied and agentic AI. It ties together many simulators and real‑world robot platforms, supports a wide range of accelerators, and provides ready‑made pipelines for fine‑tuning huge vision‑language, diffusion, and MoE models with RL algorithms such as PPO, GRPO, SAC, and newer flow‑transformation methods. The library includes extensive system optimizations (e.g., FUSCO) and dozens of example recipes, making it suitable for researchers and engineers who need a unified, scalable backbone for training and deploying RL agents in both simulation and the real world.