Embodied vision - perceiving objects from actions

摘要

For the purposes of visual manipulation of objects by a robot, the latter has to learn, about object properties, as well as actions that the robot may apply on it. This paper presents strategies to acquire such competencies based on human-robot interactions. Perception is driven by manipulation from an actor, either human or robotic. Interaction with human teachers facilitates robot learning of new objects and their functionality, or the acquisition of new competencies as an actor. Self-exploration of the world extends the robot's knowledge concerning object properties, and consolidates the execution of learned tasks.