Learning Probabilistic Decision Making by a Service Robot with Generalization of User Demonstrations and Interactive Refinement.

When learning abstract probabilistic decision making models for multi-modal service robots from human demonstrations, alternative courses of events may be missed by human teachers during demonstrations. We present an active model space exploration approach with generalization of observed action effect knowledge leading to interactive requests of new demonstrations to verify generalizations. At first, the robot observes several user demonstrations of interacting humans, including dialog, object poses and human body movement. Discretization and analysis then lead to a symbolic-causal model of a demonstrated task in the form of a preliminary Partially observable Markov decision process. Based on the transition model generated from demonstrations, new hypotheses of unobserved action effects, generalized transitions, can be derived along with a generalization confidence estimate. To validate generalized transitions which have a strong impact on a decision policy, a request generator proposes further demonstrations to human teachers, used in turn to implicitly verify hypotheses. The system has been evaluated on a multi-modal service robot with realistic tasks, including furniture manipulation and execution-time interacting humans.