Cycle WAM: Coupling Action and Video through Cycle Consistency

This research introduces Cycle WAM, a unified training framework that treats video and action as paired modalities linked by causal relationships. By employing bidirectional conversion and a differentiable cycle-reconstruction objective, the model reinforces cross-modal correspondence to recover both forward and inverse dynamics. The approach leverages pretrained video priors to achieve state-of-the-art performance on the LIBERO-Pro benchmark and shows substantial improvements over DreamZero on the DROID dataset, while demonstrating sharper visual detail in forward-dynamics predictions compared to UWM.