Sim-to-MuJoCo Transfer#

AGILE can run LEAPP-exported policies in MuJoCo for cross-simulator validation. The Sim2MuJoCo runner uses LEAPP’s Python InferenceManager to reload the exported policy graph, maps LEAPP semantic inputs to MuJoCo state, and applies LEAPP joint target/gain outputs to the MuJoCo robot.

For an overview of LEAPP bundles and policy export, see Deploy Policies. This page uses the exported bundle with AGILE’s Python MuJoCo validation runner.

Quick Start#

  1. Export the policy through LEAPP:

uv run scripts/export_policy_leapp.py \
    --task Velocity-Height-G1-History-v0 \
    --checkpoint /path/to/model.pt

  1. Download the public robot assets or bring your own MJCF:

uv run agile-download-assets
# G1 robot: external_assets/unitree_mujoco/unitree_robots/g1/scene_29dof.xml

  1. Run the LEAPP bundle in MuJoCo:

uv run scripts/sim2mujoco_eval.py \
    --leapp-yaml logs/rsl_rl/<experiment>/<run>/Velocity-Height-G1-History-v0/Velocity-Height-G1-History-v0.yaml \
    --mjcf external_assets/unitree_mujoco/unitree_robots/g1/scene_29dof.xml \
    --duration 10.0

--leapp-yaml points to the LEAPP YAML file in the exported bundle. The bundle is self-contained: the policy joints’ PD gains ride in the exported graph and the control frequency is recorded under pipeline.configs.frequency. The runner derives everything else from the bundle and the MJCF: the control decimation from the MJCF physics timestep, the reset pose from zeros, joint armature/limits from the MJCF, and default gains for any joint the policy does not control. No companion file is required.

Tip

If the robot is unstable in MuJoCo, try --pd-scale 0.3 to reduce PD gains.

Interactive Control#

The Sim2MuJoCo module supports keyboard teleoperation. Remove --no-viewer to enable the interactive viewer:

  • Arrow keys (or I/J/K/L) for movement

  • U/O for turning

  • Page Up/Down (or 9/0) for height control

  • SPACE to stop

Deterministic Evaluation#

For reproducible evaluations, use YAML-driven command schedules. These reuse the same eval config format as the Isaac Lab evaluation pipeline:

uv run scripts/sim2mujoco_eval.py \
    --leapp-yaml logs/rsl_rl/<experiment>/<run>/Velocity-Height-G1-History-v0/Velocity-Height-G1-History-v0.yaml \
    --mjcf /path/to/scene.xml \
    --eval-config agile/sim2mujoco/configs/x_velocity_sweep.yaml \
    --save-data --no-viewer

Pre-built sweep configs in agile/sim2mujoco/configs/:

Config

Description

x_velocity_sweep.yaml

Forward/backward velocity sweep

y_velocity_sweep.yaml

Lateral velocity sweep

yaw_rate_sweep.yaml

Turning rate sweep

height_sweep.yaml

Base height sweep (velocity+height tasks)

Data Logging#

Use --save-data to record per-step data to parquet files:

uv run scripts/sim2mujoco_eval.py \
    --leapp-yaml logs/rsl_rl/<experiment>/<run>/Velocity-Height-G1-History-v0/Velocity-Height-G1-History-v0.yaml \
    --mjcf /path/to/scene.xml \
    --save-data --output-dir logs/sim2mujoco/my_eval

Output structure:

logs/sim2mujoco/<task>/<eval>_<timestamp>/
  trajectories/
    episode_000.parquet
  metadata.json