Towards a "theory of mind" in simulated robots

摘要

The psychology term Theory of Mind (ToM) refers to the ability of an agent to recognize that an observed actor acts according to intentions and plans. In humans and some primates, ToM is fundamental to effective cooperation and competition, and is a key component of high-level cognition. In this paper, we explore the use of evolutionary robotics methods to create a robotic ToM. We use a co-evolutionary setup to evolve controllers that retrospectively explain an observed actor's behavior, and new actions that elicit new and more revealing behaviors. Evolved controllers can then be used to predict, manipulate and exploit the observed actor's behavior for cooperation or competition. Experimental results are shown in a physically-realistic simulation environment, and demonstrate an significant performance improvement compared to a direct estimation baseline.