RoboTwin is a dual-arm robot manipulation simulation platform designed for benchmarking and research, featuring generative digital twins. It targets researchers and engineers in robotics and AI, offering a standardized environment for developing and evaluating manipulation policies, with a focus on Sim2Real transfer.

How It Works

RoboTwin utilizes generative digital twins to create realistic simulation environments for dual-arm robot manipulation tasks. The platform supports hardware integration with AgileX Robotics and software with D-robotics, enabling direct transfer of learned policies to real-world robots. It provides a structured framework for data collection and policy evaluation, including baselines like RDT and Diffusion Policy.

Quick Start & Requirements

• Installation requires modifying the mplib library code and downloading assets.
• Estimated setup time is approximately 20 minutes.
• See INSTALLATION.md for detailed instructions.

Highlighted Details

• Accepted to CVPR 2025 (Highlight) and ECCV Workshop 2024 (Best Paper).
• Supports hardware from AgileX Robotics and software from D-robotics.
• Includes implementations and training scripts for RDT, Diffusion Policy, and 3D Diffusion Policy baselines.
• Offers a variety of dual-arm manipulation tasks, such as block manipulation, bottle handling, and tool adjustment.

Maintenance & Community

The project is actively updated, with recent changes including OpenPi deployment and RDT deployment updates. The code was migrated from an early version, and a gpt branch is available for generating expert data collectors. A WeChat group is available for user feedback and support.

Licensing & Compatibility

This repository is released under the MIT license, permitting commercial use and closed-source linking.

Limitations & Caveats

The README notes that Vulkan can be unstable in some off-screen devices. The benchmark leaderboard is considered provisional and will be comprehensively re-evaluated in Version 2.0.