DreamTrue researchers report fewer interaction defects in robot video predictions

Researchers report that DreamTrue, a model that predicts videos of robot actions, reduced human-assessed interaction defects from 48.12% to 6.25% on AgiBot. The preprint addresses predictions that follow commands inaccurately or favor successful outcomes because training data contains too few failed interactions. The system aligns rendered action trajectories with recorded videos, then modifies trajectories to generate a wider range of interactions. A reward model trained on human annotations guides further training toward physically plausible outcomes. The reported reduction concerns predicted videos, rather than a measured reduction in physical robot failures.