Maximum-Reward Motion in a Stochastic Environment: The Nonequilibrium Statistical Mechanics Perspective

Abstract

We consider the problem of computing the maximum-reward motion in a reward field in an online setting. We assume that the robot has a limited perception range, and it discovers the reward field on the fly. We analyze the performance of a simple, practical lattice-based algorithm with respect to the perception range. Our main result is that, with very little perception range, the robot can collect as much reward as if it could see the whole reward field, under certain assumptions. Along the way, we establish novel connections between this class of problems and certain fundamental problems of nonequilibrium statistical mechanics . We demonstrate our results in simulation examples.