legged_rl_lab

🧰️ Setup

Use pip to install isaaclab pip install isaaclab
Create conda environment

conda create -n legged_rl_lab python=3.11
conda activate legged_rl_lab
pip install -U torch==2.7.0 torchvision==0.22.0 --index-url https://download.pytorch.org/whl/cu128
pip install --upgrade pip

Install isaacsim 5.1 and isaaclab 2.3

pip install isaaclab[isaacsim,all]==2.3.0 --extra-index-url https://pypi.nvidia.com

Verify the installization

isaacsim

Install the project

python -m pip install -e source/legged_rl_lab
python -m pip install -e source/legged_rl_lab/legged_rl_lab/rsl_rl

List the tasks available in the project

python scripts/list_envs.py

🚀 Train

🐕️ Go2

Walk (Flat)

Walk (Flat)

#Train
python scripts/rsl_rl/train.py \
  --task=LeggedRLLab-Isaac-Velocity-Flat-Unitree-Go1-v0 \
  --num_envs 4096 \
  --headless \
  --resume \
  --load_run /path/to/log/folder \
  --checkpoint model_xx.pt

#Play
python scripts/rsl_rl/play.py \
    --task=LeggedRLLab-Isaac-Velocity-Flat-Unitree-Go1-v0 \
    --num_envs 16

Walk (Rough)

Walk(rough)

#Train
python scripts/rsl_rl/train.py \
  --task=LeggedRLLab-Isaac-Velocity-Rough-Unitree-Go1-v0 \
  --num_envs 4096 \
  --headless

CUDA_VISIBLE_DEVICES=0,1,2,3 python -m torch.distributed.run \
    --nproc_per_node=4 \
    --master_port=54321 \
    scripts/rsl_rl/train.py \
    --task LeggedRLLab-Isaac-Velocity-Rough-Unitree-Go2-v0 \
    --num_envs 4096 \
    --headless

#Play
python scripts/rsl_rl/play.py \
    --task=LeggedRLLab-Isaac-Velocity-Rough-Unitree-Go1-v0 \
    --num_envs 16

Handstand

Footstand

python scripts/rsl_rl/train.py \
  --task=LeggedRLLab-Isaac-Velocity-Footstand-Unitree-Go2-v0 \
  --num_envs 4096 \
  --headless

python scripts/rsl_rl/play.py \
    --task=LeggedRLLab-Isaac-Velocity-Handstand-Unitree-Go2-v0 \
    --num_envs 16

🤖️ G1

Walk (Flat)

#Train
python scripts/rsl_rl/train.py \
  --task=LeggedRLLab-Isaac-Velocity-Flat-Unitree-G1-v0 \
  --num_envs 4096 \
  --headless

#Play
python scripts/rsl_rl/play.py \
    --task=LeggedRLLab-Isaac-Velocity-Flat-Unitree-G1-v0 \
    --num_envs 16

Walk (Rough)

#Train
python scripts/rsl_rl/train.py \
  --task=LeggedRLLab-Isaac-Velocity-Rough-Unitree-G1-v0 \
  --num_envs 4096 \
  --headless

#Play
python scripts/rsl_rl/play.py \
    --task=LeggedRLLab-Isaac-Velocity-Rough-Unitree-G1-v0 \
    --num_envs 16

Cross-Embodied G1+Go2 (Mixed)

#Train (multi-GPU)
python -m torch.distributed.run \
  --nproc_per_node=4 \
  scripts/rsl_rl/train_cross_embodied_shared.py \
  --num_envs 4096 \
  --headless

#Train (single-GPU)
python scripts/rsl_rl/train_cross_embodied_shared.py \
  --num_envs 4096 \
  --headless

#Play
python scripts/rsl_rl/play_cross_embodied_shared.py \
  --num_envs 32

Procedural Quadruped

# Flat – Train
python scripts/rsl_rl/train.py \
    --task LeggedRLLab-Isaac-CrossEmboided-Flat-Procedural-Quadruped-v0 \
    --num_envs 4096 \
    --headless

# Flat – Play
python scripts/rsl_rl/play.py \
    --task LeggedRLLab-Isaac-CrossEmboided-Flat-Procedural-Quadruped-Play-v0 \
    --num_envs 32

# Rough – Train
python scripts/rsl_rl/train.py \
    --task LeggedRLLab-Isaac-CrossEmboided-Rough-Procedural-Quadruped-v0 \
    --num_envs 4096 \
    --headless

# Rough – Play
python scripts/rsl_rl/play.py \
    --task LeggedRLLab-Isaac-CrossEmboided-Rough-Procedural-Quadruped-Play-v0 \
    --num_envs 32

Procedural Humanoid

# Flat – Train
python scripts/rsl_rl/train.py \
    --task LeggedRLLab-Isaac-CrossEmboided-Flat-Procedural-Humanoid-v0 \
    --num_envs 4096 \
    --headless

# Flat – Play
python scripts/rsl_rl/play.py \
    --task LeggedRLLab-Isaac-CrossEmboided-Flat-Procedural-Humanoid-Play-v0 \
    --num_envs 32

# Rough – Train
python scripts/rsl_rl/train.py \
    --task LeggedRLLab-Isaac-CrossEmboided-Rough-Procedural-Humanoid-v0 \
    --num_envs 4096 \
    --headless

# Rough – Play
python scripts/rsl_rl/play.py \
    --task LeggedRLLab-Isaac-CrossEmboided-Rough-Procedural-Humanoid-Play-v0 \
    --num_envs 32

Procedural Mixed (Humanoid + Quadruped)

Trains a single policy across procedurally generated bipeds and quadrupeds simultaneously. Three pluggable obs-encoder back-ends are available:

Encoder	Flat Train Task	Rough Train Task
Mask (default)	`…-Flat-Procedural-Mixed-v0`	`…-Rough-Procedural-Mixed-v0`
Transformer	`…-Flat-Procedural-Mixed-Transformer-v0`	`…-Rough-Procedural-Mixed-Transformer-v0`
GCN	`…-Flat-Procedural-Mixed-GCN-v0`	`…-Rough-Procedural-Mixed-GCN-v0`

Architecture note: encoder lives in mdp/cross_procedural_mdp.py; all three procedural env types (ProceduralHumanoidRobotEnv, ProceduralQuadrupedRobotEnv, ProceduralMixedRobotEnv) inherit from CrossProceduralEnv which provides the unified morphology-params interface.

# ── Flat – Mask (default) ────────────────────────────────────────────────
# Train
python scripts/rsl_rl/train.py \
    --task LeggedRLLab-Isaac-CrossEmboided-Flat-Procedural-Mixed-v0 \
    --num_envs 4096 \
    --headless

# Play
python scripts/rsl_rl/play.py \
    --task LeggedRLLab-Isaac-CrossEmboided-Flat-Procedural-Mixed-Play-v0 \
    --num_envs 32

# ── Flat – Transformer ───────────────────────────────────────────────────
# Train
python scripts/rsl_rl/train.py \
    --task LeggedRLLab-Isaac-CrossEmboided-Flat-Procedural-Mixed-Transformer-v0 \
    --num_envs 4096 \
    --headless

# Play
python scripts/rsl_rl/play.py \
    --task LeggedRLLab-Isaac-CrossEmboided-Flat-Procedural-Mixed-Play-v0 \
    --num_envs 32

# ── Flat – GCN ───────────────────────────────────────────────────────────
# Train
python scripts/rsl_rl/train.py \
    --task LeggedRLLab-Isaac-CrossEmboided-Flat-Procedural-Mixed-GCN-v0 \
    --num_envs 4096 \
    --headless

# Play
python scripts/rsl_rl/play.py \
    --task LeggedRLLab-Isaac-CrossEmboided-Flat-Procedural-Mixed-GCN-Play-v0 \
    --num_envs 32

# ── Rough – Mask (default) ───────────────────────────────────────────────
# Train
python scripts/rsl_rl/train.py \
    --task LeggedRLLab-Isaac-CrossEmboided-Rough-Procedural-Mixed-v0 \
    --num_envs 4096 \
    --headless

# Play
python scripts/rsl_rl/play.py \
    --task LeggedRLLab-Isaac-CrossEmboided-Rough-Procedural-Mixed-Play-v0 \
    --num_envs 32

# ── Rough – Transformer ──────────────────────────────────────────────────
# Train
python scripts/rsl_rl/train.py \
    --task LeggedRLLab-Isaac-CrossEmboided-Rough-Procedural-Mixed-Transformer-v0 \
    --num_envs 4096 \
    --headless

# Play
python scripts/rsl_rl/play.py \
    --task LeggedRLLab-Isaac-CrossEmboided-Rough-Procedural-Mixed-Play-v0 \
    --num_envs 32

# ── Rough – GCN ──────────────────────────────────────────────────────────
# Train
python scripts/rsl_rl/train.py \
    --task LeggedRLLab-Isaac-CrossEmboided-Rough-Procedural-Mixed-GCN-v0 \
    --num_envs 4096 \
    --headless
    
# Play
python scripts/rsl_rl/play.py \
    --task LeggedRLLab-Isaac-CrossEmboided-Rough-Procedural-Mixed-GCN-Play-v0 \
    --num_envs 32

🧗 Parkour

Depth

TS-Depth Teacher

# G1 — phase 1 teacher / privileged-latent training
python scripts/rsl_rl/train.py \
  --task LeggedRLLab-Isaac-Parkour-Depth-Unitree-G1-v0 \
  --num_envs 4096 \
  --headless

# G1 — teacher play
python scripts/rsl_rl/play.py \
  --task LeggedRLLab-Isaac-Parkour-Depth-Unitree-G1-Play-v0 \
  --num_envs 50

# Go2 — phase 1 teacher / privileged-latent training
python scripts/rsl_rl/train.py \
  --task LeggedRLLab-Isaac-Parkour-Depth-Unitree-Go2-v0 \
  --num_envs 4096 \
  --headless

# Go2 — teacher play
python scripts/rsl_rl/play.py \
  --task LeggedRLLab-Isaac-Parkour-Depth-Unitree-Go2-Play-v0 \
  --num_envs 50

TS-Depth Student

Use the phase-1 checkpoint as the teacher source. With --resume, the distill config automatically uses that checkpoint as algorithm.teacher_checkpoint_path.

# G1 — phase 2 student distillation
python scripts/rsl_rl/train.py \
  --task LeggedRLLab-Isaac-Parkour-Depth-Unitree-G1-Distill-v0 \
  --num_envs 4096 \
  --headless \
  --resume \
  --load_run <teacher_run_folder> \
  --checkpoint model_xxx.pt

# Go2 — phase 2 student distillation
python scripts/rsl_rl/train.py \
  --task LeggedRLLab-Isaac-Parkour-Depth-Unitree-Go2-Distill-v0 \
  --num_envs 4096 \
  --headless \
  --resume \
  --load_run <run_folder> \
  --checkpoint model_xxx.pt

# Export the student depth policy from the distillation run
python scripts/rsl_rl/export_ts_depth_policy.py \
  --checkpoint logs/rsl_rl/go2_parkour_depth_distill/<student_run_folder>/model_xxx.pt \
  --onnx

Attention

AME Attention

This task uses an AME-style CNN + multi-head attention encoder over a 33x21x3 local terrain map, matching the AME paper-style resolution. Training uses the full-size AME parkour terrain preset: 10x20 terrain curriculum, 8m x 8m tiles, 50m border, 0.05m horizontal scale, up/down stairs, AME stakes, concentric gaps, stone bridge, and rails. Play keeps a smaller curated obstacle layout for inspection. It is a normal PPO task, separate from the TS-Depth teacher/student runner.

# G1 — train AME attention parkour
python scripts/rsl_rl/train.py \
  --task LeggedRLLab-Isaac-Parkour-Attention-Unitree-G1-v0 \
  --num_envs 256 \
  --headless

# G1 — play
python scripts/rsl_rl/play.py \
  --task LeggedRLLab-Isaac-Parkour-Attention-Unitree-G1-Play-v0 \
  --num_envs 50

# Go2 — train AME attention parkour
python scripts/rsl_rl/train.py \
  --task LeggedRLLab-Isaac-Parkour-Attention-Unitree-Go2-v0 \
  --num_envs 256 \
  --headless

# Go2 — play
python scripts/rsl_rl/play.py \
  --task LeggedRLLab-Isaac-Parkour-Attention-Unitree-Go2-Play-v0 \
  --num_envs 50

🏃 Mimic

Datasets

Place motion datasets under the project motion directory:

source/legged_rl_lab/legged_rl_lab/data/motion/
├── LAFAN1_Retargeting_Dataset/   # Motion capture retargeted CSV (30 FPS)
│   └── g1/                       # 40 CSV clips (walk, run, dance, jump, fight, fall …)
└── AMASS_Retargeted_for_G1/      # Large-scale motion capture NPZ (25 sub-libraries, 17,714 files)
    └── g1/
        ├── CMU/
        ├── KIT/
        └── ...

LAFAN1 retargeted data: LAFAN1_Retargeting_Dataset
AMASS retargeted data: AMASS_Retargeted_for_G1

LAFAN1 is stored as retargeted .csv, so convert it to .npz before AMP or tracking training.

# Convert one CSV.
python scripts/csv_to_npz.py \
  --input_file source/legged_rl_lab/legged_rl_lab/data/motion/LAFAN1_Retargeting_Dataset/g1/walk1_subject1.csv \
  --input_fps 30 \
  --output_fps 30 \
  --headless

# Or batch-convert every CSV under the G1 LAFAN1 folder.
# Remove --overwrite if you want to keep existing NPZ files.
python scripts/csv_to_npz.py \
  --input_dir source/legged_rl_lab/legged_rl_lab/data/motion/LAFAN1_Retargeting_Dataset/g1 \
  --input_fps 30 \
  --output_fps 30 \
  --headless \
  --overwrite

For AMP locomotion training, put the converted LAFAN1 run*.npz and walk*.npz files into source/legged_rl_lab/legged_rl_lab/data/motion/LAFAN1_Retargeting_Dataset/g1_amp_run_walk_fall_getup/.

# Optional: replay one converted NPZ in Isaac Sim to verify the body state.
python scripts/replay_npz.py \
  --file source/legged_rl_lab/legged_rl_lab/data/motion/LAFAN1_Retargeting_Dataset/g1_amp_run_walk_fall_getup/walk1_subject1.npz

AMP

# Train — G1 humanoid, flat terrain, AMP + RSI
# Default expert motion: LAFAN1_Retargeting_Dataset/g1/walk1_subject1.npz
python scripts/amp/train.py \
    --task LeggedRLLab-Isaac-AMP-Flat-Unitree-G1-v0 \
    --num_envs 4096 \
    --headless

# Train on one specific motion file
python scripts/amp/train.py \
    --task LeggedRLLab-Isaac-AMP-Flat-Unitree-G1-v0 \
    --num_envs 4096 \
    --headless \
    --motion_file source/legged_rl_lab/legged_rl_lab/data/motion/LAFAN1_Retargeting_Dataset/g1/walk1_subject1.npz

# Train on a directory of motions
# Use the folder containing converted LAFAN1 run/walk NPZ files.
python scripts/amp/train.py \
  --task LeggedRLLab-Isaac-AMP-Flat-Unitree-G1-v0 \
  --motion_file source/legged_rl_lab/legged_rl_lab/data/motion/LAFAN1_Retargeting_Dataset/g1_amp_run_walk_fall_getup \
  --num_envs 4096 \
  --headless \
  --max_iterations 20000

# Resume from a checkpoint
python scripts/amp/train.py \
    --task LeggedRLLab-Isaac-AMP-Flat-Unitree-G1-v0 \
    --num_envs 4096 \
    --headless \
    --resume

# Play / visualise
python scripts/amp/play.py \
    --task LeggedRLLab-Isaac-AMP-Flat-Unitree-G1-Play-v0 \
    --num_envs 50 \
    --motion_file source/legged_rl_lab/legged_rl_lab/data/motion/LAFAN1_Retargeting_Dataset/g1_amp_run_walk_fall_getup

skrl AMP (alternative AMP implementation with 3-way discriminator loss):

# Train — G1 humanoid, flat terrain, skrl AMP
python scripts/skrl/train.py \
    --task LeggedRLLab-Isaac-AMP-Flat-Unitree-G1-skrl-v0 \
    --algorithm AMP \
    --num_envs 4096 \
    --headless \
    --max_iterations 20000

# Play — auto-loads the latest checkpoint under
# logs/skrl/unitree_g1_amp_flat_skrl/<run>/checkpoints/
python scripts/skrl/play.py \
    --task LeggedRLLab-Isaac-AMP-Flat-Unitree-G1-skrl-v0 \
    --algorithm AMP \
    --num_envs 50

# Play — load a specific checkpoint by absolute path
python scripts/skrl/play.py \
    --task LeggedRLLab-Isaac-AMP-Flat-Unitree-G1-skrl-v0 \
    --algorithm AMP \
    --num_envs 50 \
    --checkpoint logs/skrl/unitree_g1_amp_flat_skrl/<run_folder>/checkpoints/agent_24000.pt

# Play — load a checkpoint copied to the local ckpt/ directory
python scripts/skrl/play.py \
    --task LeggedRLLab-Isaac-AMP-Flat-Unitree-G1-skrl-v0 \
    --algorithm AMP \
    --num_envs 50 \
    --ckpt agent_24000.pt

Motion Tracking

Task ID	Description
`Tracking-Flat-G1-v0`	Standard, with state estimation
`Tracking-Flat-G1-Wo-State-Estimation-v0`	No state estimation (closer to real deployment)
`Tracking-Flat-G1-Low-Freq-v0`	Half-frequency control

# Step 3 — Train
python scripts/rsl_rl/train.py \
  --task Tracking-Flat-G1-v0 \
  --motion_file </path/to/npz/file> \
  --num_envs 4096 --headless

# Resume
python scripts/rsl_rl/train.py \
  --task Tracking-Flat-G1-v0 \
  --motion_file <path/to/npz/file> \
  --resume --load_run <run_folder> --checkpoint model_xxx.pt \
  --num_envs 4096 --headless

# Step 4 — Play
python scripts/rsl_rl/play.py \
  --task Tracking-Flat-G1-v0 \
  --motion_file /path/to/motion.npz \
  --num_envs 16

python scripts/rsl_rl/play.py \
  --task Tracking-Flat-G1-v0 \
  --motion_file source/legged_rl_lab/legged_rl_lab/data/motion/LAFAN1_Retargeting_Dataset/g1_jump/jumps1_subject1.npz \
  --num_envs 16 \
  --checkpoint logs/rsl_rl/g1_flat/2026-04-02_02-32-52/model_11000.pt

Sim2Sim

Terrain Generator: use the terrain generator script, see terrain_tool for details.

python3 deploy/utils/terrain_tool/terrain_generator.py

Go1 Walk

See deploy/go1_deploy/README.md for details.

pip install mujoco
python deploy/go1_deploy/sim2sim_walk.py --model go1_flat.pt

Go2 Walk / Handstand

See deploy/go2_deploy/README.md for details.

pip install mujoco
# Walk
python deploy/go2_deploy/sim2sim_walk.py --model go2_rough.pt
# Handstand
python deploy/go2_deploy/sim2sim_handstand.py --model go2_handstand.pt

G1 Walk

See deploy/g1_deploy/README.md for details.

pip install mujoco
python deploy/g1_deploy/sim2sim_walk.py --model g1_flat_1.onnx --config g1_walk.yaml

Sim2Real

Go1 Walk

See deploy/go1_deploy/README.md for details.

# Dependency: unitree_legged_sdk (see README)
python deploy/go1_deploy/sim2real_walk.py --mode real --model policy.pt

Go2 Walk

See deploy/go2_deploy/README.md for details.

python deploy/go2_deploy/sim2real_walk.py --mode real --model policy.pt

G1 Walk

See deploy/g1_deploy/README.md for details.

# Dependency: cyclonedds + unitree_sdk2_python (see README)
python deploy/g1_deploy/sim2real_walk.py

Troubleshooting

Pylance Missing Indexing of Extensions

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".

{
    "python.analysis.extraPaths": [
        "<path-to-ext-repo>/source/legged_rl_lab"
    ]
}

Restart Terminal

pkill -f "python.*train.py"

Name		Name	Last commit message	Last commit date
Latest commit History 133 Commits
.claude		.claude
.vscode		.vscode
AME_Locomotion		AME_Locomotion
GMR		GMR
deploy		deploy
media		media
metamorphosis		metamorphosis
scripts		scripts
source/legged_rl_lab		source/legged_rl_lab
.dockerignore		.dockerignore
.flake8		.flake8
.gitattributes		.gitattributes
.gitignore		.gitignore
.pre-commit-config.yaml		.pre-commit-config.yaml
README.md		README.md

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Repository files navigation

legged_rl_lab

🧰️ Setup

🚀 Train

🐕️ Go2

Walk (Flat)

Walk(rough)

Footstand

🤖️ G1

🧗 Parkour

TS-Depth Teacher

TS-Depth Student

AME Attention

🏃 Mimic

Datasets

Sim2Sim

Sim2Real

Troubleshooting

Pylance Missing Indexing of Extensions

Restart Terminal

Acknowledgements

About

Uh oh!

Releases

Contributors

Uh oh!

Languages

Folders and files

Latest commit

History

Repository files navigation

legged_rl_lab

🧰️ Setup

🚀 Train

🐕️ Go2

Walk (Flat)

Walk(rough)

Footstand

🤖️ G1

🧗 Parkour

TS-Depth Teacher

TS-Depth Student

AME Attention

🏃 Mimic

Datasets

Sim2Sim

Sim2Real

Troubleshooting

Pylance Missing Indexing of Extensions

Restart Terminal

Acknowledgements

About

Topics

Resources

Uh oh!

Stars

Watchers

Forks

Releases

Contributors

Uh oh!

Languages