Research
Cycle WAM: Coupling Action and Video through Cycle Consistency
Cycle WAM introduces a unified training framework that treats video and action as paired modalities linked by causality, using bidirectional conversion and a differentiable cycle-reconstruction objective to enforce consistency between forward and inverse dynamics. It achieves state-of-the-art results on LIBERO-Pro and improves DreamZero on DROID, with sharper visual detail in forward-dynamics predictions compared to UWM.
World Action ModelsVideo ModelingRobot Learning