Learning task sequences from scratch: applications to the control of tools and toys by a humanoid robot

The goal of this work is to build perceptual and motor control systems for a humanoid robot, starting from an infant's early ability for detecting repetitive or abruptly varying world events from human-robot interactions, and walking developmentally towards robust perception and learning. This work presents strategies for learning task sequences from human-robot interaction cues. Demonstration by human teachers facilitates robot learning to recognize new objects, such as tools or toys, and their functionality. Self-exploration of the world extends the robot's knowledge concerning object properties. Multi-modal percepts are then acquired and recognized by robotic manipulation of toys and tools.