UC Irvine

Global shortest distance motion planning through a field of obstacles using gradient basedoptimization techniques is a nonlinear and non-convex problem which is very challenging to optimize. Exact motion planning techniques such as visibility graphs scale poorly with the number of obstacles, and sample based methods find optimal paths asymptotically with samples, making them not ideal for online planning applications. Inspired by the the fact that visibility graphs can find shortest path solutions in fields of circular obstacles, and that these solutions can be characterized by the direction of the path around the obstacles, we look to train an Artificial Neural Network to predict the initialization needed for locally converging to the global minimum in our shortest-distance optimization. We found this method works well for obstacle courses with low numbers of obstacles, achieving 90% test accuracy for unseen obstacle courses, but has issues scaling due to an exponential increase in the amount of data needed for generalization across the entire feature space. This technique is meant for online planning applications that require many path queries, in a configurable environment, as it leverages the O(1) time complexity of a classifier, combined with the interior point algorithm for finding optimal paths rapidly.