A dual recurrent neural network model of human-like motion for artificial agents and its evaluation in a VR mirror game turing test

TL;DR

This paper explores whether artificial agents with human-like motion can be mistaken for humans in a social interaction test.

Contribution

A novel dual recurrent neural network model is proposed and evaluated in a VR Turing test for human-like social interaction.

Findings

Participants correctly identified the interaction partner in 74% of cases.

Subjects were less likely to detect artificial agents when focused on the joint task.

Sensorimotor patterns significantly influence the subjective experience of social interaction.

Abstract

Action-oriented approaches to cognition which emphasize the constitutive role of sensorimotor patterns for perception are gaining importance for the study of cognitive processes in the human brain as well as for endowing artificial agents with cognitive capabilities. It is still debated whether motor-based action-effect contingencies can be extended to social contexts. Here, we investigate the hypothesis that social sensorimotor contingencies (socSMCs) substantially contribute to successful social interaction, and that endowing an artificial agent with socSMCs could make it an interaction partner evaluated like a human. We studied a variant of a Turing test in which human participants had to decide whether they interacted with an artificial agent or another human. To disguise the true nature of the partner, movements were mapped to standardized avatars who interacted in a virtual environment. Depending on individual traits of the participants and the duration of the interaction, in about 74% of instances participants correctly identified the interaction partner. Subjects were less likely to detect an artificial agent the more they focused on the joint task rather than on the partner. Our results suggest that the subjective experience of physical social interaction to a significant extent accrues from basic sensorimotor patterns.

Figures