Towards a simple robotic theory of mind

摘要

Theory of mind (ToM) is a cognitive function in which an agent can infer another agent's internal state and intention based on their behaviors. Can robots realize ToM like humans? There are many issues to be tackled to address this challenging problem, such as the representation, discovery and exploitation of an actor's self models. In this paper we study how robots can represent other's self with artificial neural networks and an evolutionary learning mechanism. This framework was tested with simulated and physical robots and a novel prey-predator scenario was introduced to measure the performance of ToM learning. Experimental results showed that the proposed ToM approach can recover other's self models successfully.