Towards learning path planning for solving complex robot tasks

Abstract

For solving complex robot tasks it is necessary to incorporate path planning methods that are able to operate within different high-dimensional configuration spaces containing an unknown number of obstacles. Based on Advanced A*-algorithm (AA*) using expansion matrices instead of a simple expansion logic we propose a further improvement of AA* enabling the capability to learn directly from sample planning tasks. This is done by inserting weights into the expansion matrix which are modified according to a special learning rule. For an examplary planning task we show that Adaptive AA* learns movement vectors which allow larger movements than the initial ones into well-defined directions of the configuration space. Compared to standard approaches planning times are clearly reduced.