Analyzing differences between teachers when learning object affordances via guided exploration

Abstract

Our work focuses on robots deployed in human environments. These robots, which will need specialized object manipulation skills, should leverage end-users to efficiently learn the affordances of objects in their environment. This approach is promising because prior work has shown that people naturally focus on showing salient aspects of objects when providing demonstrations. In our work, we use a guided exploration approach that combines self- and supervised learning. We present experimental results for a robot learning three affordances on four objects using 1219 interactions. We compare three conditions: (1) learning through self-exploration, (2) learning from supervised examples provided by 10 naïve users, and (3) a combined approach of self-exploration biased by user input. Previous analysis of this data focused on aggregate performance of these different strategies across all teachers, and showed that a combined approach is the most efficient and successful. In this article, we provide additional details on these specific strategies, as well as an analysis of the variance seen across teachers in this experiment. We provide a characterization of failure cases and insights for future work in learning from naïve end-users.