Docs

Dated notes on the work after the October pages: new data recorded so the destination comes from the command, two retrains of π0.5, and how each one did on the arm.

1. 9 Nov 2026

   Arm test, 9 Nov: π0.5 A at 16000 steps, against B

   A, trained from the pi05 base, on the same 31 moves as B: 21 placed correctly and it meets the 12-move gate narrowly. Also the probes and the camera check.

2. 8 Nov 2026

   Arm test, 8–9 Nov: π0.5 B at 16000 steps

   B retrained from base12k with the quantiles pinned: 11 of 31 rollouts placed correctly, and the destination now follows the command.

3. 7 Nov 2026

   Retraining π0.5 on a university cluster, 6–7 Nov

   Two full fine-tunes on the 220 contrastive demonstrations, one from the earlier checkpoint and one from the pi05 base, trained side by side on two 80 GB A100 GPUs in 12-hour jobs.

4. 4 Nov 2026

   Arm test, 4 Nov: π0.5 B at 4000 steps

   The first retrain on the 220 contrastive demonstrations lost its grasp. The cause was a normalisation mismatch, not the policy.