Learning to Manipulate Objects with a Human Teacher

Our research is motivated by the promise of robots that operate in human environments to assist people on a daily basis. It is not feasible to pre-program these robots will every skill needed, they will have to adapt and learn after they are deployed. In our view, these robots should be able to learn on their own–discovering new skills and mastering known skills through practice. Additionally, they must be able to learn from everyday people who are not familiar with robotic systems or Machine Learning algorithms. Importantly, these two “modes” of learning (social and non-social) are not mutually exclusive, but can each provide a different kind of learning opportunity for the robot. In this project we study the differences in social and non-social learning of object affordances. Does a human teacher make affordance learning faster or more efficient? What strategies to people employ to scaffold the robot’s learning process? There are several prior works aim to let a human teach skills to a robot by demonstration (Atkeson and Schaal 1997; Kuniyoshi, Inaba, and Inoue 1994; Nicolescu and Matarić 2003), or with high-level feedback to an exploratory learner (Kaplan et al. 2002; Saksida, Raymond, and Touretzky 1998). In our own prior work (Thomaz and Breazeal 2008), we began to explore the mutually beneficial relationship between self-exploration and guided learning. In the work presented here, we are working to give the teacher an even more interactive role in the teaching process, rather than simply providing verbal feedback and guidance.