MoRE: Humanoid Lifelike Gaits Learning on Complex Terrains

This is the official PyTorch implementation of the paper "MoRE: Mixture of Residual Experts for Humanoid Lifelike Gaits Learning on Complex Terrains".

🛠️ Setup Environment

Clone this repository:

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

Create a conda environment:

conda env create -f conda_env.yml 
conda activate more

Install pytorch 2.3.1 with cuda-12.1:

conda install pytorch==2.3.1 torchvision==0.18.1 torchaudio==2.3.1 pytorch-cuda=12.1 -c pytorch -c nvidia

Download and install Isaac Gym:

cd isaacgym/python && pip install -e .

Install rsl_rl and legged gym:

cd rsl_rl && pip install -e . && cd .. 
pip install -e .

🚀 Getting Started

1️⃣ Train base loco policy:
This stage trains a fundamental locomotion policy, which will later used in the residual learning phase.

python legged_gym/scripts/train.py --task g1_16dof_loco --headless

• Train for 30k–50k iterations (recommended: ≥40k).
• Use at least 3000 environments for stable learning.

2️⃣ Train residual policy:
In this stage, the model learns a residual policy built upon the pre-trained base policy.

Before training, set the checkpoint path of the base policy in g1_16dof_moe_residual_config.py

loco_expert_ckpt_path = "./logs/g1_16dof_loco/xxxxx/model_xxx.pt"

Then, start training:

python legged_gym/scripts/train.py --task g1_16dof_resi_moe --headless

• The body mask dataset required for training canbe downloaded from 📎 This link. Place it under ./body_mask_data before running trianing.
• Train for 40k iterations in total. (First 30k iterations: train the residual network. Last 10k iterations: enable body mask for deployment adaptation.）
• Use at least 6000 environments for optimal results.

The second training stage supports distributed multi-GPU training for faster convergence.

torchrun --nproc_per_node=2 legged_gym/scripts/train.py --task=g1_16dof_resi_moe --headless

3️⃣ Visualize:
After training, you can visualize the learned policy using the following command:

python legged_gym/scripts/play.py --task g1_16dof_loco/g1_16dof_resi_moe --load_run ${policy_path}

🕹️ Viewer Controls
You can manually control the robot behaviors during visualization.

Key	Function
W, A, S, D	Move forward, left, backward, right
Z, X, C	Switch gait command — Z: walk/run, X: high-knees, C: squat
[ , ]	Switch between robots
Space	Pause / Unpause simulation

4️⃣ Validate in Mujoco:
We provide four types of terrain environments for Mujoco validation:

Roughness	Pit	Stairs	Gap

The specific terrain to evaluate can be selected by configuring the YAML file (g1_16dof_resi_moe.yaml).

Run the following command for Mujoco validation:

python deploy/deploy_mujoco/deploy_mujoco_with_resi.py g1_16dof_resi_moe.yaml

📄 License

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.

Acknowledgements

We thank @lsx162534 and @ChenN-Scott for their help with the code implementation.

📝 Citation

If you find our work useful, please consider citing:

@misc{wang2025moremixtureresidualexperts,
      title={MoRE: Mixture of Residual Experts for Humanoid Lifelike Gaits Learning on Complex Terrains}, 
      author={Dewei Wang and Xinmiao Wang and Xinzhe Liu and Jiyuan Shi and Yingnan Zhao and Chenjia Bai and Xuelong Li},
      year={2025},
      eprint={2506.08840},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2506.08840}, 
}