Online Learning for Crowd-sensitive Path Planning

Abstract

In crowded environments, the shortest path for an autonomous robot navigator may not be the best choice - another plan that avoids crowded areas might be preferable. Such a crowd-sensitive path planner, however, requires knowledge about the crowd's global behavior. This paper formulates a Bayesian approach that relies only on an onboard range scanner to learn a global crowd model online. Two new algorithms, CUSUM-A* and Risk-A*, use local observations to continuously update the crowd model. CUSUM-A* tracks the spatio-temporal changes in the crowd; Risk-A* adjusts for changes in navigation cost due to human-robot interactions. Extensive evaluation in a challenging simulated environment demonstrates that both algorithms generate plans that significantly reduce their proximity to moving obstacles, and thereby protect people from actuator error and inspire their trust in the robot.