Donut as I do: Learning from Failed Demonstrations

The canonical Robot Learning from Demonstration scenario has a robot observing human demonstrations of a task or behavior in a few situations, and then developing a generalized controller. Current work further refines the learned system, often to perform the task better than the human could. However, the underlying assumption is that the demonstrations are successful, and are appropriate to reproduce. We, instead, consider the possibility that the human has failed in their attempt, and their demonstration is an example of what not to do. Thus, instead of maximizing the similarity of generated behaviors to those of the demonstrators, we examine two methods that deliberately avoid repeating the human’s mistakes.