AffordanceWAM: Affordance-Aware Joint World-Action Modeling for Robot Manipulation

This paper introduces AffordanceWAM, an affordance-aware generative World Action Model designed for generalizable robot manipulation. The method addresses the challenge of combining costly action-labeled robot videos with diverse egocentric human videos that lack robot actions and differ in embodiment. AffordanceWAM jointly models scene evolution prediction, feasible interaction identification, and action determination, targeting the multibody dynamics and contact mechanics domain of robotic manipulation.