This project aims to develop a vision-based system for Cross-view Action and Affordance Anticipation
that can transfer knowledge about object interactions between viewpoints. Imagine that a robot first
observes a human interacting with objects (exocentric view of a scene), where it learns the relationships between
objects and their affordances. Then, the robot is placed in a new environment observing it from its own, egocentric
perspective and encountering a different set of objects. Some objects may afford similar actions, while others may
afford entirely different ones. The challenge lies in enabling the robot to apply what it learned in the exocentric
view to this new egocentric scenario, even when the objects and context have changed. To address this challenging
problem, we propose using Graph Neural Networks, to model and align object relationships between exocentric and
egocentric views. The proposed approach incorporates temporal reasoning, enabling the robot to predict how object
interactions might unfold over time. This equips the system not only with the capability to recognize objects but
also to anticipate a sequence of actions necessary for a given task, even in unfamiliar environments. By transferring
learned affordances across views, the agent improves its adaptability, allowing it to anticipate and execute actions
in dynamic, real-world settings.