Robots learn to improvise when their old playbook doesn't fit the new room

Instead of retraining a frozen robot from scratch for a new task, a coding model watches it stumble, diagnoses why, and rewrites the control harness around it — like a coach adjusting plays mid-season without benching the star player.

This means you can deploy a single robot policy across wildly different environments and tasks without touching its weights, jumping success rates from 10% to 40%+ on unseen scenes.