LLM directly edits π0.5 VLA weights, no fine-tuning, lifting robot task success from 30% to 90%

Rewkang · x · 2026-11-18

Abdul Rahim Mirani had Claude Opus 5.5 directly edit the weights of the public π0.5 DROID policy — no training, no fine-tuning. After watching failures in a simulated colour-sorting task (Isaac Sim, Franka arm, Robotiq gripper), the model diagnosed the cause and turned up four attention heads 3×, changing just 0.05% of weights. Successes rose from 9/30 to 27/30 (30%→90%), and the cups-in-basket task — never used to design the edit — improved from 20/30 to 27/30 (67%→90%). All before/after numbers come from 30 unseen object placements per task. The human set the question and reviewed results; Opus 5.5 watched failures, designed and ran every experiment.

Original post →

More from coding & agent

coding & agent channel →