HOVER WBC Initial Release
Release Notes
Overview
Initial release of HOVER WBC, a neural whole-body controller framework for humanoid robots implemented as an IsaacLab extension. This release implements the methodologies described in the OmniH2O and HOVER papers.
Key Features
Core Functionality
- Teacher-student policy training pipeline for humanoid motion control
- Support for both generalist and specialist policies
- Comprehensive motion tracking and evaluation metrics
- Sim-to-sim validation using Mujoco environment
Training Capabilities
- AMASS dataset integration with retargeting support for Unitree H1 robot
- Configurable training environments
- Resumable training from checkpoints
- Multiple tracking modes including OmniH2O and humanplus
Development Tools
- Unit testing framework
- Pre-commit hooks for code quality
- IDE integration support (VSCode)
- Docker container support for headless operation
Technical Requirements
- IsaacSim 4.5.0
- IsaacLab 2.0.0
- Python 3.10
- Ubuntu 22.04
Installation
- Automated dependency installation via
install_deps.sh - Environment configuration through YAML files
- Flexible configuration system with override capabilities
Documentation
- Comprehensive README with installation and usage instructions
- Detailed API documentation
- Example configurations and training scripts
Known Limitations
- AMASS dataset not included due to licensing restrictions
- Current implementation includes four pre-selected modes for generalist policies
- Mujoco wrapper limited to single environment operation
License
Released under Apache License 2.0