Planning-based Prediction for Pedestrians

Abstract

In this paper, we describe a novel uncertainty-based technique for predicting the future motions of a moving person. Our model assumes that people behave purposefully - efficiently acting to reach intended destinations. We employ the Markov decision process framework and the principle of maximum entropy to obtain a probabilistic, approximately optimal model of human behavior that admits efficient inference and learning algorithms. The method learns a cost function of features of the environment that explains previously observed behavior. This enables generalization to physical changes in the environment, and entirely different environments. Our approach enables robots to plan paths that balance time-to-goal and pedestrian disruption. We quantitatively show the improvement provided by our approach.