Research questionHow can robot perception encode action-relevant scene dynamics to improve manipulation generalization?Robot policies may recognize what is present without representing how objects and relevant regions change during manipulation. This can make visual control brittle when tasks or environments differ from those seen during learning.