MichaelJH/ravla-libero-spatial-heldout

RA-VLA reproduction weights (held-out LIBERO-Spatial)

Private artifacts from a reproduction of RA-VLA: Retrieval-Augmented VLA for Test-Time Adaptation (ICML 2026, OpenReview ut6HebnnQe) built on NVIDIA GR00T N1.5. Code, logs, and the narrative are in the companion private GitHub repo MichaelAJH/RA-VLA_reproduce (see REPRO_LOG.md there).

These weights are fine-tunes of nvidia/GR00T-N1.5-3B and are distributed under the NVIDIA Open Model License. The RA-VLA retrieval code that produced them is the paper authors' supplementary material and has no redistribution license; do not make this repo public without their permission.

Contents

Path	What	Result
ravla/spatial_heldout/checkpoint_3/	RA-VLA action head, 3 epochs on Object+Goal+Long (Spatial never seen)	92/500 = 18.4% on LIBERO-Spatial (paper 32.0%)
retriever/spatial_heldout/model.safetensors, config.json, metadata.pt	BERT retriever trained on Object+Goal+Long	phase-alignment rho 0.97
retriever/spatial_heldout/memory_eval.pt	Test-time buffer: 3 Spatial demos/task, random.sample seed 0	used for the 18.4% run
retriever/spatial_heldout/memory_eval_seed1.pt	Same, seed 1	18.8% with the 5-epoch checkpoint

Not included: memory_train.pt (14 GB, training-only), the 5-epoch extension checkpoint, the oracle memories, and the 1-epoch author-protocol Goal checkpoint. Ask if you need them.

Serving

# in the GitHub repo, with its .venv active and PYTHONPATH set to the repo root
python scripts/rag/4_serve_ravla.py --model_path <hf>/ravla/spatial_heldout/checkpoint_3 \
    --retriever_path <hf>/retriever/spatial_heldout --memory_path <hf>/retriever/spatial_heldout/memory_eval.pt --port 8071
python examples/Libero/eval/run_libero_eval.py --task_suite_name libero_spatial --num_trials_per_task 50 --port 8071 --headless

Check scripts/rag/4_serve_ravla.py for the exact flag names; hyperparameters (tau 0.01, stride 4, margin 0.01, lambda 0.1, k=1, N=3, 4 denoising steps) are the code defaults.