Spatially Small-scale Approach-avoidance Behaviors Allow Learning-free Machine Inference of Object Preferences in Human Minds

Estimating human preference is an essential capability of a social robot. Such a machine Theory of Mind for others’ preferences is studied predominantly under the framework of inverse reinforcement learning, which however requires millions of approach-avoidance trajectories in a simulated grid world as training samples to achieve state-of-the-art results. Here, the present study explores the utility of taking spatially small-scale approach-avoidance behaviors in real life as predictive cues for efficiently estimating human preferences. Using a toy-playing scenario as an example (ΣN = 58 young and older adults), we found that a person’s subjectively reported and objectively inferred levels of attention and happiness could well predict the person’s subjective levels of toy preference. Importantly, attention and happiness can be rapidly estimated from gaze directions and emotional expressions by computer vision techniques in a learning-free manner. Compared with preference learning from spatially larger-scale movement behaviors, robot estimation of human preferences from these smaller-scale behavioral cues can be more efficient and generalizable to unlearned situations in real life.