Continuing Robot Skill Learning after Demonstration with Human Feedback

Abstract

Though demonstration-based approaches have been successfully applied to learning a variety of robot behaviors, there do exist some limitations. The ability to continue learning after demonstration, based on execution experience with the learned policy, therefore has proven to be an asset to many demonstration-based learning systems. This paper discusses important considerations for interfaces that provide feedback to adapt and improve demonstrated behaviors. Feedback interfaces developed for two robots with very different motion capabilities - a wheeled mobile robot and high degree-of-freedom humanoid - are highlighted.