Lets a frozen model learn robot manipulation from one demonstration by rehearsing in a simulated world and storing feedback in context rather than weights. It reports 83.4% mean success across 50 RoboTwin 2.0 tasks.