Checkpoints from FlowVLA-RL: critic-free GRPO on SmolVLA's flow-matching action expert, made possible by a marginal-preserving ODE-to-SDE retrofit that gives every denoising step a closed-form log-probability.

folder	what	LIBERO-Object success (n=200, deterministic ODE)
`sft-100pct-baseline/`	SFT on the full LIBERO dataset (25k steps, bs 64)	58.5%
`grpo-seed11-update300/`	the above + 300 GRPO updates (~2000 sim episodes)	68.5%

68.5% is this seed's peak. Across the seed family the fixed-budget mean is 64.8% (+6.3 pp, 4/4 seeds improve); protocol, statistics and caveats are in the repo's REPORT.md.

Usage

Checkpoints are in lerobot v0.6.0 SmolVLA format (3 camera slots; LIBERO provides 2, the third is an empty camera).

bash

hf download MorpheusTzz/smolvla-grpo-libero-object --local-dir ckpts

# evaluate with the FlowVLA-RL harness
git clone https://github.com/BlackMirean/FlowVLA-RL
python scripts/run_baseline_eval.py --checkpoint ckpts/grpo-seed11-update300 \
  --suite libero_object --episodes 20 --batch-size 5 --seed 1000

One thing to know before evaluating: lerobot v0.6.0 computes _max_episode_steps for LIBERO but never applies it, so the caller must enforce the step caps (280/280/300 per suite) or success rates come out inflated. The harness above does this; details in the repo's eval protocol.

MIT license, covering the checkpoints and this card. Base model: lerobot/smolvla_base.