Application of a Deep Learning-Based Heuristic for Weighted A-Star Algorithm

The A-Star algorithm provides an excellent graphical framework for optimal path generation between a designated start and goal position pair. Its performance comes from the fact that, at any time, it can estimate the amount of the remaining distance and steer the next step selection in the right direction. One of the main issues with the generally used heuristics is that it can not really do scene understanding, for example, to consider the obstacles. In scenarios with surrounding objects, the runtime of A-Star is considerably higher due to the spanning of the search tree with high-node expansion being hindered around the obstructions. The introduction of the weighted extension brought a trade-off between optimality and speed, and to further influence its calculation resource needs, we created a deep learning-based solution, to encode the surroundings in the format of a birds-eye-view image and utilizing this to regress the remaining path length to a given goal position. We used this estimation as a replacement for the classical heuristics and integrated it into the planning solution to measure its effect on system performance.