World-action models generate visual futures to condition robot actions, but execution feedback can make parts of the prediction invalid. RTP maintains the visual future as a persistent condition and revises it using a learned revision bridge that resumes an intermediate denoising state saved during initial generation and adapts its continuation to current observations. An adaptive policy decides whether to keep the plan, revise it, or start fresh.
On RoboMME and RMBench tasks, RTP achieved success rates of 48.6% and 84.8% respectively. The method outperforms alternatives that either reuse the unchanged plan or fully regenerate from noise. This is interesting because it allows the robot to preserve useful task structure from the original prediction while correcting parts invalidated by feedback, improving both efficiency and closed-loop performance.