Task Recognition and Generalization in Long-Term Robot Teaching

Several research efforts have addressed the challenge of having non-technical users teach or demonstrate a task to a robot, focusing on the teaching method and the task acquisition. In this paper, we consider an autonomous robot that persists over time interacting with users and the problem of teaching an additional task to the robot. We believe that the assumption that a user would know all the tasks previously taught to the robot does not hold. We hence investigate the problem of recognizing when a user is teaching a task similar to one it already knows. We present a graph-based task representation, and contribute algorithms to measure task similarity, to perform task generalization, and to generate task proposals, as a user teaches new task actions to an agent. Tasks are accumulated in a library, and the agent can learn and represent common patterns among tasks. We experimentally demonstrate our methods in a robot manipulator and a mobile service robot with large task libraries. We show that our methods lead to a significant reduction in necessary user-agent interaction and average task representation size.