English | 中文

A Reinforcement Learning (RL) framework for legged robots with manipulator arms

Legged robots with manipulation arms present unique challenges — coordinating locomotion and dexterous manipulation simultaneously. LeggedManip Lab addresses this by providing a unified RL training framework built on Isaac Lab, supporting whole-body locomotion-manipulation policy training and deployment across 7 robot platforms, including Flat and Whole-Body Control (WBC) training modes.

Isaac Lab	Mujoco	Physical

Platform	Legged Robot	Arm	Screenshot
AGO-Z1	Unitree Aliengo	Unitree Z1
B1-Z1	Unitree B1	Unitree Z1
B2-Z1	Unitree B2	Unitree Z1
GO1-ARX5	Unitree Go1	ARX-X5
GO1-WX250S	Unitree Go1	WX250S
GO2-ARX5	Unitree Go2	ARX-X5
GO2-PIPER	Unitree Go2	Agilex Piper
...	...	...	...

Each platform supports the following 2 training modes:

• Flat — Locomotion + manipulation on flat terrain
• WBC — Whole-body control with end-effector pose tracking in mixed frame

• Install Isaac Lab by following the official installation guide. We recommend using the conda installation as it simplifies calling Python scripts from the terminal.

Recommended one-click install:

wget -O install_isaaclab.sh https://docs.robotsfan.com/install_isaaclab.sh && bash install_isaaclab.sh

After installation, please verify the installed RSL-RL version:

python -m pip show rsl-rl-lib

The version should be 5.0.1 or newer. If it shows 3.1.2, you are likely using the Isaac Lab v2.3.2 release tag, which is not compatible with this repository.

• Clone this repository separately from the Isaac Lab installation (i.e. outside the IsaacLab directory):

cd LeggedManip_Lab
python -m pip install -e source/LeggedManip_Lab

• Install the additional Python dependencies for MuJoCo deployment and utilities:

pip install -r requirements.txt

python scripts/list_envs.py

Train a policy with RSL-RL PPO:

# Flat terrain training
python scripts/rsl_rl/train.py --task GO2-PIPER-Flat --num_envs 4096 --headless --max_iterations 5000

# Whole-body control training
python scripts/rsl_rl/train.py --task GO2-PIPER-WBC --num_envs 4096 --headless --max_iterations 5000

python scripts/rsl_rl/play.py --task GO2-PIPER-Flat --headless

python scripts/zero_agent.py --task GO2-PIPER-Flat    # Zero-action agent
python scripts/random_agent.py --task GO2-PIPER-Flat  # Random-action agent

Step 1 — Export the policy. The play script automatically exports policy.pt (TorchScript) and policy.onnx upon loading your trained checkpoint:

python scripts/rsl_rl/play.py --task GO2-PIPER-Flat

The exported files are saved to logs/rsl_rl/{experiment_name}/{run_name}/exported/.

Step 2 — Copy the policy to the corresponding robot's policy folder:

cp logs/rsl_rl/{experiment_name}/{run_name}/exported/policy.pt mujoco/deploy/policy/{robot}/

Step 3 — Run in MuJoCo using the platform deployment script:

python mujoco/deploy/deploy_mujoco/{robot}/{robot}.py {config_name}.yaml

Example: Running GO2-PIPER

# 1. Export
python scripts/rsl_rl/play.py --task GO2-PIPER-Flat

# 2. Copy
cp logs/rsl_rl/go2_piper_flat/2026-01-01_00-00-00/exported/policy.pt mujoco/deploy/policy/go2_piper/

# 3. Deploy
python mujoco/deploy/deploy_mujoco/go2_piper/go2_piper.py config.yaml

• Environment Details, Project Structure & Key Files

In some VsCode versions, the indexing of part of the extensions is missing. In this case, add the path to your extension in .vscode/settings.json under the key "python.analysis.extraPaths".

Note: Replace <path-to-isaac-lab> with your own IsaacLab path.

{
    "python.languageServer": "Pylance",
    "python.analysis.extraPaths": [
        "${workspaceFolder}/source/LeggedManip_Lab",
        "/<path-to-isaac-lab>/source/isaaclab",
        "/<path-to-isaac-lab>/source/isaaclab_assets",
        "/<path-to-isaac-lab>/source/isaaclab_mimic",
        "/<path-to-isaac-lab>/source/isaaclab_rl",
        "/<path-to-isaac-lab>/source/isaaclab_tasks"
    ]
}

Please cite the following if you use this code or parts of it:

@software{junjiezhu2026LeggedManip_Lab,
  author = {Junjie Zhu},
  title  = {LeggedManip_Lab: A Reinforcement Learning Framework for Legged Manipulation},
  url    = {https://github.com/zzzJie-Robot/LeggedManip_Lab},
  year   = {2026}
}

The project uses some code from the following open-source code repositories:

• fan-ziqi/robot_lab
• unitreerobotics/unitree_mujoco
• google-deepmind/mujoco