Deploy Policies

LEAPP and Isaac ROS Deploy

To deploy policies sim-to-sim and sim-to-real this project uses LEAPP and Isaac ROS Deploy. This allows you to deploy your policy as fast and efficient C++ code without rewriting any code.

LEAPP packages a policy into a self-contained artifact, a LEAPP bundle, which can then be executed by a LEAPP runtime.

Isaac ROS Deploy provides LEAPP runtimes for real robots and simulation-in-the-loop deployment. The runtimes execute policies through C++, providing deterministic, low-latency deployment without requiring policy code changes.

For a complete Unitree G1 runtime walkthrough, see the Isaac ROS AGILE G1 deployment tutorial.

Export a Policy with LEAPP

Use scripts/export_policy_leapp.py to export a supported AGILE training checkpoint. Choose the task and checkpoint that match the policy you want to deploy:

uv run scripts/export_policy_leapp.py \
    --task <task-name> \
    --checkpoint /path/to/checkpoint.pt \
    --export_save_path /path/to/leapp-bundles \
    --disable_graph_visualization

The command creates a directory named for the task. Keep that entire directory together: its YAML configuration references the ONNX model and, when present, the safetensors state artifact through relative paths.

Example: Export the G1 Velocity Policy

The repository ships a Velocity-G1-History-v0 checkpoint that can be exported for the Isaac ROS Controller Manager:

export AGILE_LEAPP_EXPORT_DIR="$HOME/agile-leapp/velocity-g1"
uv run scripts/export_policy_leapp.py \
    --task Velocity-G1-History-v0 \
    --checkpoint agile/data/policy/velocity_g1/unitree_g1_velocity_history_state_dict.pt \
    --export_save_path "${AGILE_LEAPP_EXPORT_DIR}" \
    --disable_graph_visualization

Use ${AGILE_LEAPP_EXPORT_DIR}/Velocity-G1-History-v0/Velocity-G1-History-v0.yaml as the LEAPP configuration path. Copy the complete Velocity-G1-History-v0 directory when moving the bundle to the runtime machine.