GR00T post-training (VLA fine-tuning)

COMPASS distillation datasets can fine-tune VLA models like NVIDIA Isaac-GR00T to bolt on navigation capabilities. Three steps: convert, fine-tune, evaluate.

Step 1 — Convert HDF5 → GR00T LeRobot format

You need an HDF5 distillation dataset first; see Recording distillation data. Then convert with the bundled script:

python scripts/hdf5_to_lerobot_episodic.py \
    --hdf5-dir <path/to/hdf5/directory> \
    --output-path <path/to/lerobot/format>

The script is pure-Python (no Isaac Lab) — runs anywhere with the standard COMPASS Python environment. It walks the HDF5 dataset, repacks per-episode into LeRobot’s parquet format, and emits the chunk + metadata layout that GR00T’s training pipeline expects.

Step 2 — Post-train GR00T

Follow the post-training instructions in the Isaac-GR00T getting-started guide.

A ready-to-use navigation data configuration lives on this branch: liuw/nav_fine_tune.

Step 3 — Evaluate the post-trained GR00T model in COMPASS

Launch the GR00T inference server (see the Isaac-GR00T repo) on port 8888, with the same data configuration you used during training. Then evaluate from COMPASS:

python run.py \
    -c configs/eval_config.gin \
    -o <output_dir> \
    -b <path/to/x_mobility_ckpt> \
    --enable_cameras \
    --gr00t-policy

--gr00t-policy tells run.py to dispatch action queries to the inference server instead of loading a local checkpoint. Eval parameters (scene, embodiment, episode count) live in configs/eval_config.gin.