Produced by vla-agent-loop: rollouts collected on the physical SO-101 rig; the episodes a VLM judged successful were merged with the base teleop dataset and the base policy was finetuned on HPC. No reward or progress model is involved.

NOT evaluated: deployed at the end of the run, but no collect measured it. Measured series from the same run (50-episode collects, VLM judge on side2+top): base 22.0% (11/50), then after each SFT round 4.0%, 8.0%, 5.0%, 10.0%. The four trained rounds pooled are 14/210 = 6.7% against the base's 22.0% (Fisher exact p=0.0025) -- the self-improvement loop DEGRADED the policy. Cause not identified. Leading hypothesis is overfitting: the base was already 5000 steps on the same 50 teleop episodes, and each round added 5000 more. A hypothesis that the successful rollouts' 25 s timeout tails poisoned training was checked against the recordings and NOT supported -- failures look like the base's (approach, then stall before the grasp).

Note that the rig's on-policy success rate is measured over very few episodes; treat any single evaluation here as an estimate with a wide confidence interval.