OneHBC: A Unified Framework for Humanoid Body Control

MjOneHBC

A Unified Framework for Humanoid Body Control Training Based on Mjlab

OneHBC is a dedicated research framework for training humanoid robot locomotion and whole-body control policies using Mjlab with the mujoco physics engine. It supports high-performance end-to-end reinforcement learning and motion imitation for humanoid robots, with a focus on speed tracking, AMP-based motion imitation, and whole-body trajectory tracking. Future extensions will support general whole-body VLA (Vision-Language-Action) control.

---

Features

• Velocity Tracking Control: Omnidirectional speed command tracking for robust locomotion

• AMP (Adversarial Motion Priors): High-quality natural motion imitation learning

• Whole-Body Trajectory Tracking: Accurate task-space and joint-space trajectory following

• Under Development: General whole-body VLA (Vision-Language-Action) control pipeline

---

Environment Requirements

• OS: Ubuntu 22.04 / 24.04

• Mjlab: >=1.3.0

• Python: 3.12

• CUDA: 12.8 or higher

---

Installation

1. Set Up Conda Environment and Dependencies

conda create -n onehbc python=3.12
conda activate onehbc

2. Install mjlab and rsl-rl Library

pip install mjlab
pip install rsl-rl

3. Clone MjOneHBC

git clone https://github.com/HongtuZ/MjOneHBC.git
cd MjOneHBC
pip install -e source/OneHBC

---

Training Commands

Velocity Tracking Control

# Train
python scripts/rsl_rl/train.py Velocity-Flat-THS23DOF --num_envs 4096

# Train with video recording
python scripts/rsl_rl/train.py Velocity-Flat-THS23DOF --num_envs 4096 --video

# Play
python scripts/rsl_rl/play.py Velocity-Flat-THS23DOF --num_envs 16

### AMP Imitation Learning
TODO: Add AMP training and evaluation commands

### Whole\-Body Trajectory Tracking
TODO: Add whole-body trajectory tracking training and evaluation commands