Sim-to-MuJoCo Transfer#

AGILE includes a generic, task-agnostic framework (agile/sim2mujoco/) for running trained policies in MuJoCo simulation. The framework automatically handles observation/action mapping by parsing the exported I/O descriptor YAML file – no code changes needed for different tasks.

Quick Start#

  1. Export policy and I/O descriptors from your trained checkpoint:

python scripts/eval.py --task Velocity-G1-History-v0 --checkpoint /path/to/checkpoint.pt
python scripts/export_IODescriptors.py --task Velocity-G1-History-v0 --output_dir /path/to/output

  1. Get robot MJCF from Unitree’s official repository or bring your own:

git clone https://github.com/unitreerobotics/unitree_mujoco.git
# G1 robot: unitree_mujoco/unitree_robots/g1/scene_29dof.xml

  1. Run evaluation in MuJoCo:

python scripts/sim2mujoco_eval.py \
    --checkpoint /path/to/exported/policy.pt \
    --config /path/to/exported/config.yaml \
    --mjcf unitree_mujoco/unitree_robots/g1/scene_29dof.xml \
    --duration 10.0

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:

python scripts/sim2mujoco_eval.py \
    --checkpoint /path/to/policy.pt \
    --config /path/to/config.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 (joint positions, velocities, accelerations, torques, commands, root state) to parquet files. Output is compatible with the existing agile.algorithms.evaluation.plotting utilities:

python scripts/sim2mujoco_eval.py \
    --checkpoint /path/to/policy.pt \
    --config /path/to/config.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