MetaFine
User Guide · Policies · StarVLA

StarVLA

StarVLA is a compact-but-capable VLA reference baseline; in MetaFine it's wired through the same --task-graph evaluate.py contract as every other supported backbone.

What it is

StarVLA — a compact VLA reference baseline. Vendored under core/policies/starvla. The StarVLA training path is verified.

Data

StarVLA trains on a LeRobot dataset. Produce one with the record → merge → replay → convert_to_lerobot flow (see Quickstart).

Train (verified)

Per core/policies/starvla/train/README.md:

  1. conda activate starVLA
  2. Place the LeRobot dataset under core/policies/starvla/starVLA/playground/Dataset/
  3. Register it in core/policies/starvla/starVLA/dataloader/gr00t_lerobot/mixtures.py
  4. StarVLA expects LeRobot 2.1 — convert a 3.0 dataset with lerobot_v30_to_v21 and add a modality.json to meta/
  5. Launch training:
$ bash run_libero_train.sh

Source of truth: core/policies/starvla/train/README.md (and eval/README.md for evaluation).