跳到主要内容

GBC Project Now Open Source!

· 阅读需 4 分钟
Yifei Yao
Yifei Yao
GBC Project Lead, SJTU MVASL Lab

We are thrilled to announce that the GBC (Generalized Behavior Cloning) framework is now officially open source! 🎉 Researchers and developers worldwide can now access and use our codebase for their humanoid researches!

📦 Open Source Repositories​

We have released our key codebases on GitHub:

  • GBC Core Framework: https://github.com/sjtu-mvasl-robotics/GBC

    • Contains complete motion retargeting pipeline
    • Provides comprehensive training environment and tools
    • Supports multiple humanoid robot platforms

  • Reinforcement Learning Algorithm Library: https://github.com/sjtu-mvasl-robotics/rsl_rl

    • Implements MMPPO, PPO, and other core algorithms
    • Provides efficient distributed training architecture
    • Includes MMTransformer network implementation

🌟 GBC Framework Highlights​

The GBC framework is a comprehensive, unified solution designed to solve the end-to-end challenge from human motion to robot action:

1. Adaptive Data Processing Pipeline​

  • Leverages a differentiable IK network to automatically retarget any human MoCap data to any humanoid robot
  • Provides real-time, efficient, and physically feasible kinematic retargeting
  • Comprehensive data post-processing capabilities

2. Innovative Network Architecture & Enhanced PPO Algorithm​

  • Employs MMTransformer to seamlessly integrate robot environment observations with reference actions
  • Enables single-action to multi-action imitation through our MMPPO framework
  • Supports switching freely between action tracking and reinforcement learning instruction following modes

3. Cross-Platform Testing and Validation​

  • Train policies in Isaac Sim
  • Validate policy transfer in MuJoCo (evaluation code coming soon)
  • Provides comprehensive evaluation framework

📋 Release Roadmap​

ComponentStatusDescription
Retargeting Pipeline✅ ReleasedComplete framework for retargeting human motion capture data to any humanoid robot morphology
RL+IL Repository✅ ReleasedCore reinforcement learning and imitation learning algorithms used in GBC
Reinforcement Learning Library✅ ReleasedAssociated RL algorithm library with implementations of PPO, MMPPO, and custom extensions
Documentation & Tutorials✅ ReleasedBasic usage documentation and getting-started tutorials
Robot-Specific Training Demos✅ ReleasedDemonstration examples for specific robot configurations
Training Scripts for All Test Robots⏳ Coming SoonComplete training scripts for all robots evaluated in the paper
Demo Videos🔜 ImminentComprehensive demonstration videos of the framework in action
Hugging Face Checkpoints🔜 ImminentPre-trained model checkpoints available on Hugging Face
Tutorial Videos⏳ PlannedStep-by-step video tutorials for using the GBC framework
Advanced Features Documentation⏳ PlannedDetailed tutorials for advanced features (AMP, DAgger, Symmetry Augmentation, etc.)
MuJoCo Sim2Sim Validation⏳ PlannedScripts for validating sim-to-sim transfer using MuJoCo
Deployment Code⏳ PlannedDeployment code for currently available robot configurations

🛠️ Getting Started​

Prerequisites​

Before starting, ensure you have all the necessary prerequisites installed on your system, including Python, CUDA, Isaac Lab/Sim (optional), and appropriate hardware.

Installation Steps​

  1. Set up your environment

    # Create and activate conda environment
    conda create -n GBC python=3.10
    conda activate GBC

  2. Install GBC and dependencies

    # Create and navigate to GBC directory
    mkdir GBC && cd GBC

    # Clone GBC repository with all submodules
    git clone https://github.com/sjtu-mvasl-robotics/GBC.git --recursive
    cd GBC

    # If you have Isaac Lab installed, remove its auto-installed rsl_rl first
    pip uninstall rsl_rl

    # Install GBC and its dependencies
    pip install -e .
    cd dependencies/rsl_rl
    pip install -e .

  3. Verify installation

    python verify_installation.py

    Successful installation should show:

    ✅ GBC Core: OK
    ✅ SMPL Support: OK
    ✅ Isaac Lab: OK (if installed)
    ✅ rsl_rl: OK (if installed, and has dev tag)

🎥 Coming Soon​

We are actively preparing the following content to be released soon:

  • Demo Videos - Showcasing the GBC framework in action on various humanoid robots
  • Tutorial Videos - Step-by-step video guides on how to use the GBC framework
  • Pre-trained Models - Pre-trained checkpoints available on Hugging Face

🤝 Collaboration Opportunity​

We are actively seeking to test and validate our algorithms on additional humanoid robot platforms. If you are a researcher or industry partner working with humanoid robots and are interested in collaboration, we'd love to discuss potential partnerships. Our framework is designed to be adaptable to various morphologies, and we're excited to expand our testing across diverse platforms.

🙏 Community Support​

We warmly welcome community feedback, suggestions, and improvements to the algorithm and its implementation! Your pull requests and contributions are greatly appreciated and vital to the continued development of this project.

📞 Contact Us​

If you have any questions or feedback, please feel free to reach out to us:

If you want to contact our laboratory for more information, it is recommended to reach out to the corresponding author. For technical inquiries, please contact the first author.

We look forward to seeing the amazing humanoid robot applications the community creates with the GBC framework!