smolvla-libero-spatial-full

smolvla-libero-spatial-full

SmolVLA fine-tuned on LIBERO-Spatial (10 tasks, 432 demonstrations) by Pavan Yadav Annappa. Full action-expert fine-tune: vision encoder and VLM frozen, action expert + state/action projections trained (no LoRA), batch 32, 20000 steps, bf16, lr 0.0001 OneCycle, seed 0. Trained on one NVIDIA L4 with the pipeline in https://github.com/pavanyadava007/vla-lab (streaming shards, exact resume).

Result

LIBERO-Spatial success, 3 seeds x 50 episodes/task, re-plan every 10 steps: 77.6% +- 2.5 (pooled 95% CI 75.4-79.6, n=1500). Baseline LoRA run (001): 41.3% +- 2.2 open-loop, 55.5% with re-planning every 10 steps. All numbers with CIs: https://huggingface.co/spaces/pavanyadava07/vla-lab

Files

model.pt: {"model": state_dict, "meta": {...}} (450 M parameters; VLM weights bf16, trained parameters fp32). meta.extra carries the model config and the LIBERO normalisation stats, so the loader needs nothing else.

Use

git clone https://github.com/pavanyadava007/vla-lab && cd vla-lab   # env recipe in the README
python - <<'PY'
from huggingface_hub import hf_hub_download
from scripts.policy_wrappers import load_policy
p = load_policy(hf_hub_download("pavanyadava07/smolvla-libero-spatial-full", "model.pt"), n_action_steps=10)
# p.reset(); action = p.act(raw_libero_obs, "pick up the black bowl on the stove and place it on the plate")
PY
python scripts/eval_parallel.py --config configs/smolvla_libero_spatial_full.yaml --ckpt model.pt --out results/008.json --procs 5 --n-action-steps 10

Observation conventions (LIBERO render rotated 180 deg, 8-d state = eef pos + axis-angle + gripper) are handled in scripts/policy_wrappers.py.