Recognizing affect in human touch of a robot

A pet cat or dog's ability to respond to our emotional state opens an interaction channel with high visceral impact, which social robots may also be able to access. Touch is a key but understudied element; here, we explore its emotional content in the context of a furry robot pet. We asked participants to imagine feeling nine emotions located in a 2-D arousal-valence affect space, then to express them by touching a lap-sized robot prototype equipped with pressure sensors and accelerometer. We found overall correct classification (Random Forests) within the 2-D grid of 36% (all participants combined) and 48% (average of participants classified individually); chance 11%. Rates rose to 56% in the high arousal zone. To better understand classifier performance, we defined and analyzed new metrics that better indicate closeness of the gestural expressions. We also present a method to combine direct affect recognition with affect inferred from gesture recognition. This analysis provides a unique first insight into the nature and quality of affective touch, with implications as a design tool and for incorporating unintrusive affect sensing into deployed interactions.