LEARNING SAMPLING DISTRIBUTIONS FOR RANDOMIZED MOTION PLANNING: ROLE OF HISTORY SIZE

摘要

Recently, we have proposed a novel motion planning algorithm based on random walks. One of its main features is that it can incorporate adaptive components. This means that the developer is not required to provide all the parameters which specifies the stochastic mechanism through which the free configuration space is explored. In fact, the algorithm adapts to the shape of the space it is currently moving in from the last generated H samples, where H is the history size which is a priori fixed. Then, according to this information, it suitably modifies the random distribution from which the next sample is drawn, in order to speed up the exploration. In this paper we investigate via numerical experiments how the choice of the history size influences the performance of the algorithm, as well as the effectiveness of the learning process itself.