arxivcs.ROcs.CV2026-07-14
FlowWAM: Optical Flow as a Unified Action Representation for World Action Models
Yixiang Chen, Peiyan Li, Yuan Xu, Qisen Ma, Jiabing Yang, Kai Wang, et al.
World Action Models (WAMs) are able to leverage pretrained video generators for both world modeling and action prediction. However, directly leveraging such video generators for control raises a new challenge: how to represent actions in a suitable form that aligns with pretraine…