Human-Robot Instruction for Perceptual Teamwork

The utility of mobile robots would be greatly enhanced if they could be taught how to be competent, helpful members of military teams in the field. We propose a method for human instruction of robots, where the instructor teaches the robot how to recognize common situations and activities surrounding it, and how it should behave as part of those activities. Our approach is based on combining probabilistic models of group activities, prior domain and general knowledge, and human instruction. The instructor can guide the robot learning process by providing initial activity specifications, and corrections of the robot’s interpretation of its environment through noisy observations (tracks of nearby objects). In Phase 1, we will develop initial methods for activity model specification and learning with human instruction. We will use the system to investigate the tradeoff between prior knowledge, video training examples and human instruction. Greater levels of instruction can compensate for less training data, but would require more effort from the instructor.