Reusing a robot's behavioral mechanisms to model and manipulate human mental states

Abstract

In a task domain characterized by physical actions and where information has value, competing teams gain advantage by spying on and deceiving an opposing team while cooperating teammates can help the team by secretly communicating new information. For a robot to thrive in this environment it must be able to perform actions in a manner to deceive opposing agents as well as to be able to secretly communicate with friendly agents. It must further be able to extract information from observing the actions of other agents. The goal of this research is to expand on current human robot interaction by creating a robot that can operate in the above scenario. To enable these behaviors, an architecture is created which provides the robot with mechanisms to work with hidden human mental states. The robot attempts to infer these hidden states from observable factors and use them to better understand and predict behavior. It also takes steps to alter them in order to change the future behavior of the other agent. It utilizes the knowledge that the human is performing analogous inferences about the robot's own internal states to predict the effect of its actions on the human's knowledge and perceptions of the robot. The research focuses on the implicit communication that is made possible by two embodied agents interacting in a shared space through nonverbal interaction. While the processes used by a robot differ significantly from the cognitive mechanisms employed by humans, each face the similar challenge of completing the loop from sensing to acting. This architecture employs a self-as-simulator strategy, reusing the robot's behavioral mechanisms to model aspects of the human's mental states. This reuse allows the robot to model human actions and the mental states behind them using the grammar of its own representations and actions.