Researchers at Beihang and Peking University described Revisable Temporal Planning, or RTP, in a paper submitted Sept. 28, 2026. It is a way for a robot's AI to fix its plan as it goes instead of starting over after every move. RTP works with world action models, AI models for robots that predict how a scene will change and which moves the robot can make to shape it. In the researchers' tests, revising the plan led to more successful tasks than replanning from scratch each time.
The model RTP builds on predicts a short video of what should happen next and uses it to steer the robot's next move. RTP keeps that predicted video as the guide for the robot's actions. As the model creates the video, it saves partly finished versions along the way. After the robot acts, a trained part the researchers call a revision bridge picks up one of those saved versions. It reworks the rest of the video to fit what the robot now sees, using a record of what has already happened. Before each move, the system picks whether to keep the plan, revise it or start a fresh one.
The researchers built RTP on a version of an existing model, LingBot-VA, adapted to the test tasks. In the researchers' tests on two sets of robot tasks, revising the plan succeeded more often than starting over each time, and more often than every other method they compared.
The results come only from the researchers' own tests on those two sets of tasks. No outside lab has tested the method yet, and whether the gains carry over to other robots and jobs is an open question.