lerobot-legged-zoo

Legged robot models (MJCF + assets) plus MJLab training examples in one place. Run from the repo root with uv.

Status by robot:

• LeRobot Humanoid family: in development (common CAD source): https://cad.onshape.com/documents/fb645318a27646d1d8840be6/w/d1cae8805fb652b4d1614997/e/621ef473e7cde4418aaec2ed
• Leggy: WIP concept; no hardware yet.
• Open Duck: reference repo https://github.com/apirrone/Open_Duck_Playground

Installation

This project uses uv. A CUDA-capable GPU is required for training/play.

uv sync

Run Training

# Train with default settings
uv run train Mjlab-Velocity-Flat-Open-Duck-v2

# Adjust number of parallel environments
uv run train Mjlab-Velocity-Flat-Open-Duck-v2 --env.scene.num-envs 2048

# See all options
uv run train Mjlab-Velocity-Flat-Open-Duck-v2 --help

Run Play/Evaluation

# Play with a wandb checkpoint
uv run play Mjlab-Velocity-Flat-Open-Duck-v2 --wandb-run-path <your-wandb-path>

# Play with a local checkpoint
uv run play Mjlab-Velocity-Flat-Open-Duck-v2 --checkpoint-file logs/rsl_rl/open_duck_v2/<run>/model_*.pt

# Use random actions (sanity check)
uv run play Mjlab-Velocity-Flat-Open-Duck-v2 --agent random

Available Tasks

Task Name	Description
Mjlab-Velocity-Flat-Leggy	Train Leggy to walk on flat ground
Mjlab-Velocity-Rough-Leggy	Train Leggy to walk on rough terrain
Mjlab-Velocity-Flat-Open-Duck-v2	Train Open Duck v2 to walk on flat ground
Mjlab-Velocity-Rough-Open-Duck-v2	Train Open Duck v2 to walk on rough terrain
Mjlab-Velocity-Flat-Unitree-G1-23dof	Train Unitree G1 (23-DoF) on flat ground
Mjlab-Velocity-Rough-Unitree-G1-23dof	Train Unitree G1 (23-DoF) on rough terrain
Mjlab-Velocity-Flat-Unitree-G1-29dof	Train Unitree G1 (29-DoF) on flat ground
Mjlab-Velocity-Rough-Unitree-G1-29dof	Train Unitree G1 (29-DoF) on rough terrain
Mjlab-Velocity-Flat-LeRobot-Humanoid	Train LeRobot Humanoid to walk on flat ground
Mjlab-Velocity-Rough-LeRobot-Humanoid	Train LeRobot Humanoid to walk on rough terrain
Mjlab-Velocity-Flat-LeRobot-Humanoid-full	Train LeRobot Humanoid (full) on flat ground
Mjlab-Velocity-Rough-LeRobot-Humanoid-full	Train LeRobot Humanoid (full) on rough terrain
Mjlab-Velocity-Flat-LeRobot-Humanoid-no-arms	Train LeRobot Humanoid (no arms) on flat ground
Mjlab-Velocity-Rough-LeRobot-Humanoid-no-arms	Train LeRobot Humanoid (no arms) on rough terrain

Robots (Details)

Each robot has a model under models/ and a matching training example under training_exemples/.

Leggy

• Model: models/leggy/robot.xml
• Training: training_exemples/leggy/env_cfgs.py
• Constants: training_exemples/leggy/leggy_constants.py
• Tasks: Mjlab-Velocity-Flat-Leggy, Mjlab-Velocity-Rough-Leggy
• Small biped with parallel legs, designed for compact locomotion experiments.

Open Duck v2 (Backlash)

• Model: models/open_duck_v2/open_duck_v2_backlash.xml
• Training: training_exemples/open_duck_v2/env_cfgs.py
• Constants: training_exemples/open_duck_v2/open_duck_v2_constants.py
• Tasks: Mjlab-Velocity-Flat-Open-Duck-v2, Mjlab-Velocity-Rough-Open-Duck-v2
• Updated Open Duck with backlash joints and improved geometry.

LeRobot Humanoid Full (20-DOF)

• Model: models/lerobot_humanoide/mjcf/robot.xml
• Training: training_exemples/lerobot_humanoid_full/env_cfgs.py
• Constants: training_exemples/lerobot_humanoid_full/lerobot_humanoid_full_constants.py
• Tasks: Mjlab-Velocity-Flat-LeRobot-Humanoid-full, Mjlab-Velocity-Rough-LeRobot-Humanoid-full
• Full humanoid with upper‑body joints for whole‑body control research.

LeRobot Humanoid (12-DOF)

• Model: models/bipedal_plateform/mjcf/robot.xml
• Training: training_exemples/lerobot_humanoid/env_cfgs.py
• Constants: training_exemples/lerobot_humanoid/lerobot_humanoid_constants.py
• Tasks: Mjlab-Velocity-Flat-LeRobot-Humanoid, Mjlab-Velocity-Rough-LeRobot-Humanoid
• 12‑DOF lower‑body humanoid for fast velocity‑tracking experiments.

LeRobot Humanoid No-Arms

• Model: models/bipedal_plateform_no_arms/mjcf/robot.xml
• Training: training_exemples/lerobot_humanoid_no_arms/env_cfgs.py
• Constants: training_exemples/lerobot_humanoid_no_arms/lerobot_humanoid_no_arms_constants.py
• Tasks: Mjlab-Velocity-Flat-LeRobot-Humanoid-no-arms, Mjlab-Velocity-Rough-LeRobot-Humanoid-no-arms
• Lower‑body‑only humanoid variant without arms.

Modify Rewards / Cost Function

Edit training_exemples/<robot>/env_cfgs.py and tune cfg.rewards:

# Example: change weights
cfg.rewards["upright"].weight = 1.0
cfg.rewards["body_ang_vel"].weight = -0.05

# Example: add a new cost term
from mjlab.managers.reward_manager import RewardTermCfg
from mjlab.tasks.velocity import mdp

cfg.rewards["self_collisions"] = RewardTermCfg(
  func=mdp.self_collision_cost,
  weight=-1.0,
  params={"sensor_name": "self_collision"},
)

After edits, rerun your task with uv run train ....