QinnyuLan / SONIC_unitree_h2

# SONIC_unitree_h2
Official Unitree H2 deployment implementation for TRACE: Efficient Transfer Learning for Neural Whole-Body Control Across Humanoid Embodiments.
## Overview
TRACE transfers the pretrained SONIC whole-body policy to Unitree H2 with low additional training cost. This deployment provides unified command interfaces for G1, H2, and SMPL reference motions, supporting motion replay, motion planning, and VR teleoperation. A direct token interface is reserved for future VLA/VLN integration.
## Workflow
| Workflow | Description | Guide |
|---|---|---|
| Sim2Sim | MuJoCo simulation with the C++ controller and DDS communication | Simulation |
| Sim2Real | Whole-body motion tracking on Unitree H2 | Hardware deployment |
| VR teleoperation | PICO head-and-hand tracking with XRoboToolkit | Teleoperation |
Source code is organized as `src/runtime/` (C++ policy), `src/interface/` (Python command producers), and `src/sim/` (simulator). Asset preparation, robot services and validation tools are organized under `tools/`.
## Getting Started
The runtime initializes a standing pose, waits for **]** or **s** to enable the policy, then accepts commands from independent motion replay, planner, and teleoperation processes. Press **o** to stop with damping. Simulation and hardware use the same runtime. Head pitch/yaw and dexterous hands have separate command channels. Input services and the planner use Python; the real-time policy runtime uses C++. Policy and planner inference use TensorRT.
 [code] 
Follow the deployment guide to install dependencies, build the controller, and run simulation. H2 checkpoints are provided in `models/sonic_h2_released/`. The SONIC planner is optional and runs as a separate service.
## Documentation
- Architecture and command protocol: process boundaries, sparse references, and independent head/hand control.
- Repository layout: C++ runtime, Python services, models, configuration and motions.
- Installation and deployment: environment setup, Sim2Sim, Sim2Real, and VR teleoperation.
- Troubleshooting: installation, GPU startup, and connection issues.
- Robot hand service: Inspire DDS/RS-485 deployment, autostart and watchdog.
- Released model: input/output interfaces and checksums.
- Reproducibility: model validation, inference checks, and test results.
- Documentation index: configuration, troubleshooting, and development notes.
## Citation
If you use this project in your research, please cite:
 [code] 
Bibliographic metadata is also available in `CITATION.cff` and `references.bib`.
## Acknowledgments
- SONIC / GR00T Whole-Body Control: pretrained controller, motion representation, and deployment framework.
- MuJoCo: physics simulation and visualization.
- Unitree SDK2: robot DDS communication.
- DFX Inspire Service: Inspire hand DDS service and serial protocol implementation.
- XRoboToolkit: PICO VR device and head-and-hand tracking interfaces.
See upstream dependencies and references for the full dependency list.
## License