Learning safe navigation in uncertain environments

We present a strategy for the navigation of mobile robots in open sparse uncertain environments. Our work assumes that the robot has no access to external informations about its absolute position, and that the environment consists of a set of objects whose separation is large compared to the range of the perception sensors. The approach is based on: (i) maximizing the probability of successful transition between any pair of known objects - by requiring if necessary that the robot uses other objects as intermediate landmarks - and (ii) guaranteeing - by controlling the extension of dead-reckoning motions and resorting if necessary to active search for perceptual information - that a reliable path to the homing region always exists. The paper presents experimental results that clearly demonstrate the increase in autonomy resulting from the proposed navigation strategy. (C) 2000 Elsevier Science B.V. All rights reserved.