Jinrui Han^{* 1,2}, Dewei Wang^{* 1,3}, Chenyun Zhang¹, Xinzhe Liu^1,4, Ping Luo⁵, Chenjia Bai^†1, Xuelong Li^†1
* Equal Contribution  † Corresponding Author
¹Institute of Artificial Intelligence (TeleAI), China Telecom   ²Shanghai Jiao Tong University   ³University of Science and Technology of China   ⁴ShanghaiTech University   ⁵The University of Hong Kong

This repository contains the official implementation of our paper: HUSKY: Humanoid Skateboarding System via Physics-Aware Whole-Body Control. In this work, we propose a learning-based whole-body control framework that empowers humanoid robots to perform dynamic skateboarding.

This repository contains:

• The mjlab training framework (src/mjlab_husky)
• Customized RL implementations (rsl_rl/)
• Motion data for AMP and trajectory planning (dataset/)
• Lightweight MuJoCo evaluation scripts (test_scene/)
• Tested checkpoints (ckpts/)

This code has been tested on Ubuntu 22.04 with CUDA 13.0. To install this repository, please follow these steps:

1. Install the uv package manager (if you don't have it yet):

   curl -LsSf https://astral.sh/uv/install.sh | sh

2. Clone the repository:

   git clone https://github.com/TeleHuman/humanoid_skateboarding.git
   cd humanoid_skateboarding

3. Sync dependencies:

   uv sync
   uv pip install -e .

uv run train Mjlab-Skater-Flat-Unitree-G1 --env.scene.num-envs 4096

uv run play Mjlab-Skater-Flat-Unitree-G1 --checkpoint_file your-ckpt-path

We also provide a lite MuJoCo simulation script for evaluation:

bash test_scene/sim.sh your-onnx-path

The test_scene/mjlab_scene.xml file is automatically generated from the mjlab scene_cfg. In the simulation, you can control the skateboard using the keyboard arrow keys. Visualization examples are shown below, rendered from test.pt and test.onnx:

Viser	MuJoCo

If you find our work helpful, please consider citing us:

@article{han2026husky,
    title={HUSKY: Humanoid Skateboarding System via Physics-Aware Whole-Body Control},
    author={Jinrui Han and Dewei Wang and Chenyun Zhang and Xinzhe Liu and Ping Luo and Chenjia Bai and Xuelong Li},
    journal={arXiv preprint arXiv:2602.03205},
    year={2026}
  }

This codebase is under CC BY-NC 4.0 license. You may not use the material for commercial purposes, e.g., to make demos to advertise your commercial products.

• mjlab: Our training framework is based on mjlab by MuJoCo Lab.
• rsl_rl: The reinforcement learning algorithm is built upon the rsl_rl library.
• mujoco_warp: GPU-accelerated interface for rendering and physics simulation.
• mujoco: High-fidelity rigid-body physics engine.
• AMP: We build on Adversarial Motion Priors for pushing behaviors.
• DHAL: We drew inspiration from the quadrupedal robot skateboarding project.

For further collaborations or discussions, please feel free to reach out to:

• Jinrui Han: jrhan82@sjtu.edu.cn or Wechat: Bw_rooneY
• Chenjia Bai (Corresponding Author): baicj@chinatelecom.cn