DVF-RRT: Randomized Path Planning on Predictive Vector Fields

Abstract

Autonomous surface vehicle (ASV) navigation in marine environments is challenging due to the disturbances caused by water currents and their spatiotemporal variations. Existing methods take into account only spatial variations of vector fields that are measured through vehicle sensors, but neglect temporal variations of vector fields. Effective path planning for ASVs also requires critical reasoning about the prediction of spatiotemporally varying water currents in marine environments. Therefore, this paper presents a method that integrates the prediction of water vector fields with a randomized path planner. We model the water flow of an area of interest as an unknown vector field and then train a Long-Short Term Memory (LSTM) neural network to learn such an unknown vector field accurately and effectively from real ocean current data. This allows the generation of a randomized path that moves along the vector field in a continuous space. To generate a randomized path on the predicted vector field, we present a Deep Vector Field - Rapidly-exploring Random Tree (DVF-RRT) algorithm for reaching a goal configuration starting from an initial configuration that leverages the strength of the RRT algorithm. The algorithm is validated through simulated randomized paths on predictive vector fields and benchmarking with regard to an existing VF-RRT method.