Human-in-the-Loop Feedback to Improve Robot Teammate Learning

Author(s):

Julia L. Wright and Maggie Wigness

Abstract:

Participants oversaw a robot autonomously navigating to four waypoints in a specified order. The robot initially preferred to traverse over grass, and participants were instructed to train it to drive on the road. Each participant’s task was to observe the robot’s actions, intervene when its behavior was undesirable, and demonstrate the desired behavior by teleoperating the robot. Participants determined when to intervene and how long to demonstrate the desired behavior. After each demonstration, the robot updated its behavior model to better mimic the demonstrated actions and then continued on using the newly updated model. Each participant completed three trials under different transparency conditions. The within-subjects analysis showed that demonstrations effectively changed the robot’s behavior. When robot intent was transparent, participants reduced the number of demonstrations. This led to a decrease in participant workload and an increase in performance satisfaction. Feedback on their demonstrations further decreased both the number and distance of demonstrations, improving demonstration quality and further reducing participant workload, while also boosting confidence in the robot’s capabilities. The results provide strong evidence that nonexperts can successfully train agents in new behaviors after deployment, and that greater transparency of the agent improves outcomes for both the human user and the robot.