rollout_probe_pi05_fullft

pi0.5 full fine-tune (ECE cluster, full-sotac data: 163 episodes, 4 objects) on the drifted rig, 2026-09-22. Same rig and day as rollout_probe_pi05 (pi05_ft) and rollout_probe_pi05_lora.

Red foam ball, 5 eps per scene in order 6,5,4,3,2,7: 1-6 5/5, 1-5 5/5, 1-4 5/5, 1-3 5/5, 1-2 5/5, 1-7 5/5. Total 30/30.

Notes:

• Motion as smooth as pi05_ft. Confirms the lora failure is the r=16 adapter, not the data recipe.
• Same systematic offset as ft: the gripper clips the right edge of the bowl at release. The ball lands inside clean. Offset smaller than ft's.
• 1-3 grasp retries: ep 3 failed twice then succeeded, ep 5 failed once then succeeded with an unstable grasp. Recovery behavior, no interventions.