Task Refinement for Autonomous Robots using Complementary Corrective Human Feedback

摘要

A robot can perform a given task through a policy that maps its sensed state to appropriate actions. We assume that a hand-coded controller can achieve such a mapping only for the basic cases of the task. Refining the controller becomes harder and gets more tedious and error prone as the complexity of the task increases. In this paper, we present a new learning from demonstration approach to improve the robot's performance through the use of corrective human feedback as a complement to an existing hand-coded algorithm. The human teacher observes the robot as it performs the task using the hand-coded algorithm and takes over the control to correct the behavior when the robot selects a wrong action to be executed. Corrections are captured as new state-action pairs and the default controller output is replaced by the demonstrated corrections during autonomous execution when the current state of the robot is decided to be similar to a previously corrected state in the correction database. The proposed approach is applied to a complex ball dribbling task performed against stationary defender robots in a robot soccer scenario, where physical Aldebaran Nao humanoid robots are used. The results of our experiments show an improvement in the robot's performance when the default hand-coded controller is augmented with corrective human demonstration.