Confidence-Based Demonstration Selection for Interactive Robot Learning

Effective learning by demonstration techniques enable complex robot behaviors to be taught from a small number of demonstrations. Demonstrations are obtained based on a selection algorithm, which governs which states are labeled by the human teacher. In this work, we examine selection algorithms used by the robot to request demonstration examples. Previous approaches typically rely on a fixed confidence threshold. In this work, we highlight the drawbacks of using a single threshold, and contribute an algorithm for automatically setting multiple confidence thresholds designed to target domain states with the greatest uncertainty. Our evaluation compares the proposed multi-threshold selection method to confidence-based selection using a single fixed threshold, and manual data selection by the teacher. Our results indicate that the multi-threshold approach significantly reduces the number of demonstrations required to learn the task.