On self-organizing map and rapidly-exploring random graph in multi-goal planning

Abstract

This paper reports on ongoing work towards an extension of the self-organizing maps for the traveling salesman problem to more challenging problems of multi-goal trajectory planning for complex robots with a high-dimensional configuration space. The main challenge of this problem is that the distance function needed to find a sequence of the visits to the goals is not known a priori and it is not easy to compute. To address this challenge, we propose to utilize the unsupervised learning in a trade-off between the exploration of the distance function and exploitation of its current model. The proposed approach is based on steering the sampling process in a randomized sampling-based motion planning technique to create a suitable motion planning roadmap, which represents the required distance function. The presented results shows the proposed approach quickly provides an admissible solution, which may be further improved by additional samples of the configuration space.