RoboTwin is a comprehensive platform for bimanual robotic manipulation, offering a scalable data generator and benchmark with strong domain randomization. It targets researchers and developers in robotics, computer vision, and AI, providing a robust environment for training and evaluating dual-arm robot policies.

How It Works

RoboTwin leverages generative digital twins to create diverse and realistic simulation environments. Its core strength lies in extensive domain randomization across task configurations, embodiments, and visual properties, enabling the development of highly generalizable robotic manipulation policies. The platform supports multiple policy baselines and facilitates custom policy deployment.

Quick Start & Requirements

Installation involves cloning the repository and following instructions in the RoboTwin 2.0 Document, estimated to take around 20 minutes. Specific hardware requirements, such as GPU and CUDA versions, are detailed in the documentation.

Highlighted Details

• Supports multiple policy baselines including DP, ACT, DP3, RDT, PI0, TinyVLA, DexVLA, and LLaVA-VLA.
• Offers over 100,000 pre-collected trajectories in the RoboTwin Dataset.
• Features a highly configurable data collection pipeline with domain randomization.
• Recognized with a Best Paper Award at an ECCV Workshop and selected as a CVPR Highlight paper.

Maintenance & Community

The project has multiple branches for different versions and challenges, indicating active development. Links to documentation, papers, and community resources are provided.

Licensing & Compatibility

Released under the MIT license, allowing for commercial use and integration with closed-source projects.

Limitations & Caveats

The platform is under active development, with several planned features and policy integrations listed in the TODO section. Users are encouraged to perform their own data collection due to the high configurability.