A decision theoretic approach for task coordination in social robots

Abstract

We present a new method for designing and implementing socially interactive mobile robots built around a 3-layer hybrid control architecture. Our main contribution is at the deliberative level, where we introduce multiply sectioned Markov decision processes (MS-MDPs) as a mechanism for efficient task specification, policy generation and execution. Using MS-MDPs, we partition the task into a number of subtasks, each assigned to an MDP, such that each one can be specified and solved independently. Each MDP controls one aspect of the global task, and they all are executed concurrently, coordinated implicitly by common state variables. We validate our approach by presenting experiments performed using our robot HOMER, the human oriented messenger robot. HOMER is a stereo-vision guided mobile robot designed for performing a message delivery task, which allows for rich and complex robot-human interactions using a multi-modal interface. HOMER's deliberative layer includes 3 MDPs: the navigator, the dialogue manager and the gesture generator. Together they coordinate 10 behaviors for accomplishing the message delivery task.