Coaching: An Approach to Efficiently and Intuitively Create Humanoid Robot Behaviors

Abstract

The advances in humanoid robots in recent years have given researchers new opportunities to study and create algorithms for generating humanoid behaviors. Not surprisingly, most approaches for creating or modifying behaviors for complex humanoids require specialized knowledge and a large amount of work. Our aim is to provide an alternative, intuitive way to program humanoid behavior. To do this, we examine human-to-human skill transfer, specifically coaching, and adapt it to the humanoid setting. We enable a real-time scenario where a person, acting as a coach, interactively directs humanoid behavior to a desired outcome. This tightly coupled interaction between a person and a humanoid allows efficient, directed learning of new behaviors, where behavior characteristics can be modified on-demand. Communication is realized through demonstration and a coaching vocabulary, and changes are effected by transformation functions acting in the behavior domain