Learning Gaze Behaviors for Balancing Participation in Group Human-Robot Interactions
summary
Learns gaze policies from prior Furhat group interaction logs using imitation learning, batch offline reinforcement learning, and a four-action gaze set: look speaker / avert / look listener / do nothing with the robot acting twice per second. RL uses a reward = reduction in participation unevenness per turn multiplied by turn length and estimates a Q-function from batch data. Offline user study (N=50, twelve groups per new condition) shows RL increases number of turns significantly, IL raises overall voice activity versus a prior speaker-listener heuristic, and neither policy breaks interaction appropriateness.
AI-written summary · about