rollout_probe_pi05

Full 6-scene evaluation of pi05-sotac-ft (pi0.5 full fine-tune, sotac ep 0-70, red foam ball) recorded 2026-09-22 on a rig with KNOWN camera-geometry drift relative to the Sep 5 training-era setup. Started as a drift probe, extended to the full protocol; the name kept the probe prefix.

Context: on this same drifted rig, the ACT baseline (act-sotac-runA-ep0-70) scored 1/5 at scene 1-6 versus its 4/5 on Sep 5. This dataset shows pi0.5 on the identical geometry. Scores here are NOT comparable to datasets recorded on the Sep 5 rig; they are comparable to rollout_probe_A (same day, same rig).

Protocol: bowl at point 1, ball at points 6,5,4,3,2,7, five episodes per scene, 60 s episode / 30 s reset, episodic strategy, RTC inference, cameras top+wrist, tactile recorded in the 318-dim state (policy is tactile-blind; a processor step trims state to 6 joints at inference).

Episode map and operator tally

episodes	scene	success
0-4	1-6	5/5
5-9	1-5	5/5
10-14	1-4	4/5
15-19	1-3	5/5, ep 15 intervened
20-24	1-2	5/5
25-29	1-7	5/5, ep 26 intervened

Operator count 29/30 (97%). Strict count excluding the two intervened episodes as successes: 27/30 (90%).

Notes

• Intervened episodes: ep 15 (ball dropped mid-transport, replaced by hand, policy then completed), ep 26 (operator intervention).
• Crash-resume seams after ep 15 and ep 16 (one empty-episode writer crash from an intermittent stall; no episode data lost). Two unused USB cameras were unplugged after ep 15; the stall did not recur.
• Consistent behavioral residual of the drift: the ball is released at the right edge of the bowl in most successes, inside but off-center.
• A partial mid-session snapshot exists as rollout_probe_pi05_20260922_142711; it is a crash artifact and should be ignored.