Maximum Mean Discrepancy Imitation Learning

摘要

Imitation learning is an efficient method for many robots to acquire complex skills. Some recent approaches to imitation learning provide strong theoretical performance guarantees. However, there remain crucial practical issues, especially during the training phase, where the training strategy may require execution of control policies that are possibly harmful to the robot or its environment. Moreover, these algorithms often require more demonstrations than necessary to achieve good performance in practice. This paper introduces a new approach called Maximum Mean Discrepancy Imitation Learning that uses fewer demonstrations and safer exploration policy than existing methods, while preserving strong theoretical guarantees on performance. We demonstrate empirical performance of this method for effective navigation control of a social robot in a populated environment, where safety and efficiency during learning are primary considerations.