Export the Policy with LEAPP#

This tutorial explains how to export the fine-tuned policy into a deployable LEAPP bundle.

Exporting the policy with LEAPP lets you deploy the policy from pure C++ without any Python involved, even though the policy was originally created in Python.

Deploying policies through pure C++ has the benefit of faster and more deterministic inference while using fewer resources than deploying through Python.

Run this tutorial on an x86_64 workstation. LEAPP export for GR00T is not supported on Jetson AGX Thor.

Prerequisites#

If you have not already done so as part of Fine-Tune GR00T N, clone the repository and set it up.

git clone --branch n17-export --recurse-submodules https://github.com/nvidia-isaac/gr00t-leapp-export.git
cd gr00t-leapp-export
uv sync --python 3.10

Run the Export#

Important

For G1 fine-tunes, always pass --joint_config export/data/g1_joints.json. Without it, the export auto-generates placeholder element names (left_leg_0, left_leg_1, …) that the deploy stack’s joint-state mapping cannot resolve, producing wrong (or silently incorrect) action predictions on hardware.

cd gr00t-leapp-export
uv run python export/export_with_leapp.py \
    --model_path <checkpoint-dir> \
    --embodiment_tag new_embodiment \
    --dataset_path <lerobot-dataset-dir> \
    --joint_config export/data/g1_joints.json \
    --output_name <export-name>
Argument reference

  • --model_path — path to the checkpoint-NNNN/ directory produced by Fine-Tune GR00T N.

  • --embodiment_tag — always new_embodiment for G1 fine-tunes.

  • --dataset_path — any LeRobot dataset with the same modality keys the model was trained on; used only as a sample input during tracing.

  • --joint_config — path to g1_joints.json shipped in the export repository. Provides the real ROS joint names the deploy stack needs; without it, placeholder names are used (refer to warning above).

  • --output_name — directory name for the exported bundle.

Tracing takes 5–15 minutes on a workstation GPU. The script ends with Export completed: <export-name>.

Output#

A directory at <export-name>/ containing the deployable bundle:

  • <export-name>.yaml — LEAPP graph spec; entry point for deploy.

  • <export-name>.png — visualization of the export graph.

  • backbone.onnx + backbone.onnx.data — VLM backbone.

  • action_head.onnx + action_head.onnx.data — diffusion action head.

  • preprocess_state.onnx — state normalization.

  • preprocess_video.onnx — image preprocessing.

  • decode_action.onnx — action decoding.

  • log.txt — export log.

The .yaml is what the deploy stack consumes. The .onnx and .onnx.data files must remain co-located with it.

Next#

Continue with Deploy the Fine-Tuned Policy to run the exported policy on the Unitree G1.