What Will You Do Next? A Cognitive Model for Understanding Others’ Intentions Based on Shared Representations

Abstract

Goal-directed action selection is the problem of what to do next in order to progress towards goal achievement. This problem is computationally more complex in case of joint action settings where two or more agents coordinate their actions in space and time to bring about a common goal: actions performed by one agent influence the action possibilities of the other agents, and ultimately the goal achievement. While humans apparently effortlessly engage in complex joint actions, a number of questions remain to be solved to achieve similar performances in artificial agents: How agents represent and understand actions being performed by others? How this understanding influences the choice of agent’s own future actions? How is the interaction process biased by prior information about the task? What is the role of more abstract cues such as others’ beliefs or intentions? In the last few years, researchers in computational neuroscience have begun investigating how controltheoretic models of individual motor control can be extended to explain various complex social phenomena, including action and intention understanding, imitation and joint action. The two cornerstones of control-theoretic models of motor control are the goal-directed nature of action and a widespread use of internal modeling. Indeed, when the control-theoretic view is applied to the realm of social interactions, it is assumed that inverse and forward internal models used in individual action planning and control are re-enacted in simulation in order to understand others’ actions and to infer their intentions. This motor simulation view of social cognition has been adopted to explain a number of advanced mindreading abilities such as action, intention, and belief recognition, often in contrast with more classical cognitive theories - derived from rationality principles and conceptual theories of others’ minds - that emphasize the dichotomy between action and perception. Here we embrace the idea that implementing mindreading abilities is a necessary step towards a more natural collaboration between humans and robots in joint tasks. To efficiently collaborate, agents need to continuously estimate their teammates’ proximal goals and distal intentions in order to choose what to do next.We present a probabilistic hierarchical architecture for joint action which takes inspiration from the idea of motor simulation above. The architecture models the casual relations between observables (e.g., observed movements) and their hidden causes (e.g., action goals, intentions and beliefs) at two deeply intertwined levels: at the lowest level the same circuitry used to execute my own actions is re-enacted in simulation to infer and predict (proximal) actions performed by my interaction partner, while the highest level encodes more abstract task representations which govern each agent’s observable behavior. Here we assume that the decision of what to do next can be taken by knowing 1) what the current task is and 2) what my teammate is currently doing. While these could be inferred via a costly (and inaccurate) process of inverting the generative model above, given the observed data, we will show how our organization facilitates such an inferential process by allowing agents to share a subset of hidden variables alleviating the need of complex inferential processes, such as explicit task allocation, or sophisticated communication strategies.