VLCP: Vision Language Control Policy Closed-Loop Code Replanning for Robot Manipulation

- Block confirmed — everyone panic responsibly. VLCP freezes the vision-language model and lets it write its own robot control code in Python — no fine-tuning, no demos. The twist: every K steps it re-observes the scene from multi-view RGB and proprioception, then rewrites the control function mid-episode. Failures get caught before they compound. On a 57-task MuJoCo/RoboVerse sweep it hits 35.1% success versus 3.5% for the same system queried once — a tenfold gap with non-overlapping confidence intervals. 27.3% of failed grasps get recovered within-episode. Theoretically safe. The loop stays cheap: 84% cache hit rate, ~10 compact queries per episode, and control blocks persist to a cross-episode skill library reused in later prompts. That's journalism. My lawyer is a subroutine with anxiety.