SmolVLA Fine-Tuning & Closed-Loop Evaluation
100K-update VLA adaptation and an audited 400-episode LIBERO benchmark
I completed an end-to-end SmolVLA reproduction on one RTX 4090, from pinned pretrained-policy initialization and runtime-contract validation to a frozen 100K-update checkpoint and 400 closed-loop LIBERO rollouts.
Impact
- 71.5% endpoint success (286/400) across all four LIBERO suites
- 100,000 optimizer updates / 6.4M sample presentations on one RTX 4090
- 40 tasks × 10 fixed-init episodes, with native success predicates
- Strict checkpoint reload: 0 missing, unexpected, or shape-mismatched keys
- Complete evidence chain: machine-readable results, logs, frozen configs, and 400 rollout videos
The strongest suite reached 82% success; Long remained the clearest failure boundary at 46%. The result is deliberately scoped to a single-seed simulated benchmark—not SOTA, a multi-seed significance claim, or real-robot deployment.
Representative rollouts
Spatial task 0 — successful episode
Long-horizon task 1 — successful episode