Leveraging Machine Learning to Improve Adaptive Primitive-Based Motion Planning

Abstract

This paper introduces a new approach for adding replanning capabilities to the maneuver automaton. We call this approach “maneuver interruption.” Maneuver interruption enables replanning by identifying maneuver segments that are dynamically similar to the current vehicle state. As a result, the vehicle can exit a maneuver if new information emerges or the environment changes. We use machine learning to enhance the performance of maneuver interruption. Specifically, we examine how supervised learning can predict dynamic similarity and utilize the learned network to enable maneuver interruption. A variety of models are compared for their ability to quantify the feasibility of a maneuver-to-maneuver transition. The multilayer perceptron is found to be the most effective at this task and was therefore selected for generating maneuver-to-maneuver transitions for replanning. Additionally, we use Monte Carlo methods and pruning to reduce the transition library size by an order of magnitude with minimal loss in performance. We test learning-enhanced maneuver interruption on obstacle evasion tasks with a medium-fidelity ZOHD Drift flight dynamics model. On randomly generated obstacle fields, maneuver interruption is demonstrated to enable longer collision-free flights at a minor cost to control performance.