Legged Lab

Legged Lab is an external Isaac Lab extension for legged-robot reinforcement learning and motion imitation. The current version contains flat-ground velocity tracking for Unitree G1 and Go2, plus DeepMimic, Adversarial Motion Priors (AMP), and online policy distillation for the 29-DoF Unitree G1.

The AMP implementation depends on the project-specific RSL-RL fork.

Demo

Adversarial Motion Priors on Unitree G1:

rl-video-step-0.mp4

Version compatibility

Component	Current repository target
Isaac Lab	2.3.2
Isaac Sim	5.1.0 (the version paired with Isaac Lab 2.3.2)
Python	3.11 for the current development environment; the package declares Python >= 3.10
RSL-RL	zitongbai/rsl_rl, branch feature/amp; training checks for package version >= 3.0.1
Docker base image	nvcr.io/nvidia/isaac-lab:2.3.1 (the default in docker/.env.base)

Note

The native setup targets Isaac Lab 2.3.2. The checked-in Docker workflow still uses the 2.3.1 image by default; override ISAACLAB_IMAGE in docker/.env.base if you maintain a newer compatible image.

Features

• Flat-ground velocity tracking with PPO for Unitree G1 and Go2.
• DeepMimic reference-motion tracking for the 29-DoF Unitree G1.
• AMP training with asymmetric actor-critic observations, symmetry augmentation, and self-contact handling.
• Online, DAgger-style distillation of a trained G1 AMP teacher into a smaller MLP student.
• Batch and single-motion conversion from GMR pickle data to the Legged Lab motion format.
• JIT and ONNX policy export during playback.
• Native and Docker-based development workflows.

Repository layout

legged_lab/
├── source/legged_lab/legged_lab/
│   ├── assets/                 # Robot asset configurations and USD resources
│   ├── data/MotionData/        # Git LFS motion datasets
│   ├── envs/                   # Custom AMP and animation environments
│   ├── managers/ and sensors/  # Runtime components
│   └── tasks/locomotion/       # Velocity, DeepMimic, AMP, and animation configs
├── scripts/rsl_rl/             # Training and playback entry points
├── scripts/tools/retarget/     # GMR-to-Legged-Lab conversion tools
└── docker/                     # Container image and lifecycle scripts

Installation

Native Isaac Lab environment

Prerequisites:

• Linux with an NVIDIA GPU and a driver compatible with Isaac Sim 5.1.0.
• Isaac Lab 2.3.2 installed and working.
• Git LFS.
• The feature/amp branch of the project RSL-RL fork.

Clone Legged Lab outside the Isaac Lab source tree and download its tracked assets:

git clone https://github.com/zitongbai/legged_lab.git
cd legged_lab
git lfs install
git lfs pull

Install Legged Lab with the Python interpreter used by Isaac Lab:

python -m pip install -e source/legged_lab

Clone and install the required RSL-RL fork in a separate directory:

git clone --branch feature/amp https://github.com/zitongbai/rsl_rl.git
cd rsl_rl
python -m pip install -e .

Run all commands below from the Legged Lab repository root with that same Python environment active.

Docker environment

The Docker scripts expect the two repositories to be siblings by default:

lab_dev/
├── legged_lab/
└── rsl_rl/

For another layout, set RSL_RL_PATH in docker/.env.base. The path is resolved relative to docker/.

Build and start the container:

bash docker/build.sh
bash docker/run.sh
bash docker/enter.sh

The startup script mounts both repositories, installs them in editable mode, enables all GPUs, and stores Isaac Sim caches and user data under ~/docker/isaac-sim. It also replaces .vscode/settings.json with the container settings.

Stop and remove the container with:

bash docker/stop.sh

After changing the Dockerfile, stop the existing container, rebuild the image, and run it again.

Available tasks

Task IDs are defined by the Gym registrations in source/legged_lab/legged_lab/tasks/locomotion.

Purpose	Training task	Playback task
G1 flat velocity	LeggedLab-Isaac-Velocity-Flat-Unitree-G1-v0	LeggedLab-Isaac-Velocity-Flat-Unitree-G1-Play-v0
Go2 flat velocity	LeggedLab-Isaac-Velocity-Flat-Unitree-Go2-v0	LeggedLab-Isaac-Velocity-Flat-Unitree-Go2-Play-v0
G1 DeepMimic	LeggedLab-Isaac--Deepmimic-G1-v0	LeggedLab-Isaac--Deepmimic-G1-Play-v0
G1 AMP	LeggedLab-Isaac-AMP-G1-v0	LeggedLab-Isaac-AMP-G1-Play-v0
G1 AMP distillation	LeggedLab-Isaac-AMP-G1-Distill-v0	Use the same task with play.py

The double hyphen in the registered DeepMimic IDs is intentional. There is also a debugging environment named LeggedLab-Isaac--Deepmimic-G1-Debug-v0.

To query the registrations directly from the installed extension:

python scripts/list_envs.py

Training and playback

The following examples use headless simulation. Remove --headless to open the simulator UI.

Velocity tracking

# Unitree G1
python scripts/rsl_rl/train.py \
    --task LeggedLab-Isaac-Velocity-Flat-Unitree-G1-v0 \
    --headless

# Unitree Go2
python scripts/rsl_rl/train.py \
    --task LeggedLab-Isaac-Velocity-Flat-Unitree-Go2-v0 \
    --headless

DeepMimic

python scripts/rsl_rl/train.py \
    --task LeggedLab-Isaac--Deepmimic-G1-v0 \
    --headless \
    --max_iterations 10000

AMP

python scripts/rsl_rl/train.py \
    --task LeggedLab-Isaac-AMP-G1-v0 \
    --headless

To select a non-default GPU, set both the simulator and agent devices:

python scripts/rsl_rl/train.py \
    --task LeggedLab-Isaac-AMP-G1-v0 \
    --headless \
    --device cuda:1 \
    agent.device=cuda:1

Training outputs are written to logs/rsl_rl/<experiment_name>/<timestamp>_<run_name>/.

Online AMP policy distillation

Distillation requires a trained AMP checkpoint. --load_run is the teacher run directory under logs/rsl_rl/g1_amp, and --checkpoint may be a checkpoint filename such as model_2999.pt.

python scripts/rsl_rl/train.py \
    --task LeggedLab-Isaac-AMP-G1-Distill-v0 \
    --headless \
    --load_run <teacher-run-directory> \
    --checkpoint <teacher-checkpoint>

The current student network uses hidden layers [256, 128, 64]; the frozen teacher uses [512, 256, 128]. The default distillation run lasts 3,000 iterations and is stored under the g1_amp experiment.

Play, record, and export

Use the corresponding -Play-v0 task where one exists:

python scripts/rsl_rl/play.py \
    --task LeggedLab-Isaac-AMP-G1-Play-v0 \
    --headless \
    --num_envs 64 \
    --video \
    --checkpoint logs/rsl_rl/g1_amp/<run-directory>/model_<iteration>.pt

Playback writes a video to <checkpoint-directory>/videos/play/ and exports the policy to <checkpoint-directory>/exported/policy.pt and policy.onnx. If --checkpoint is omitted, the script resolves a checkpoint from --load_run and the task's experiment directory.

Use python scripts/rsl_rl/train.py -h and python scripts/rsl_rl/play.py -h for all command-line options.

Motion data

Ready-to-use G1 motion files are stored with Git LFS in:

• source/legged_lab/legged_lab/data/MotionData/g1_29dof/amp/walk_and_run
• source/legged_lab/legged_lab/data/MotionData/g1_29dof/deepmimic

To add a GMR-retargeted dataset:

1. Retarget the source motion to the G1 model with GMR.

2. Put the resulting .pkl files in an input directory, for example temp/gmr_data.

3. Convert the complete dataset:

   python scripts/tools/retarget/dataset_retarget.py \
       --robot g1 \
       --input_dir temp/gmr_data \
       --output_dir temp/lab_data \
       --config_file scripts/tools/retarget/config/g1_29dof.yaml \
       --loop clamp \
       --headless

4. Move the converted files into the appropriate MotionData directory and update motion_data.motion_dataset.motion_data_dir and motion_data_weights in the task configuration.

The batch converter processes every .pkl file in the input directory and does not clip frame ranges. See scripts/tools/retarget/single_retarget.py and scripts/tools/retarget/gmr_to_lab.py for single-motion conversion and the serialized data format.

To inspect a configured G1 animation before training:

python scripts/play_anim.py --robot g1_29dof

Development

Code uses 4-space indentation, a 120-character line limit, Black formatting, and isort's Black profile. Run the configured checks before committing:

pre-commit run --all-files

Generated artifacts normally belong in logs/, outputs/, or temp/ and should not be committed. Motion datasets and robot assets that are intentionally versioned should remain managed by Git LFS.

Roadmap

• Flat-ground velocity tracking for G1 and Go2
• DeepMimic and AMP for G1
• AMP symmetry augmentation, asymmetric actor-critic, and self-contact handling
• Online AMP policy distillation
• Rough-terrain AMP locomotion
• Sim-to-sim validation in MuJoCo
• Image observations
• Additional humanoid robots such as Unitree H1

Acknowledgements

• Isaac Lab — simulation and environment framework.
• RSL-RL — reinforcement-learning foundation.
• AMP_for_hardware — AMP implementation reference.
• GMR — human-to-robot motion retargeting.
• MimicKit — imitation-learning reference.

License

This project is licensed under the Apache License 2.0.

AMP