Detecting and Correcting Model Anomalies in Subspaces of Robot Planning Domains

Abstract

Making decisions based on accurate models enables robots to exploit domain knowledge to act intelligently. However, in many realistic domains, it is impossible to have globally accurate models, as the world may exhibit modes of behav-ior during deployment that were unforeseeable during model building. This paper addresses the problem of adaptation in domains in which robots have access to a model of the world that is generally accurate, but which is inadequate in par-ticular sets of similar situations –i.e., subspaces of the task domain. Using optimization techniques to find parametric approximations to these subspaces, our framework general-izes from sparse observations to find and correct for statisti-cal incongruences between expected and observed behavior. We demonstrate this framework in a domain in which single deployment adaptation is essential: a team of soccer robots keeping the ball away from a previously unknown opponent. Empirical results show that the framework improves model accuracy and task performance over timescales comparable to a single soccer game.