Expanding Constrained Kinodynamic Path Planning Solutions through Recurrent Neural Networks

Abstract

Path planning for autonomous systems with the inclusion of environment and kinematic/dynamic constraints encompasses a broad range of methodologies, often providing trade-offs between computation speed and variety/types of constraints satisfied. Therefore, an approach that can incorporate full kinematics/dynamics and environment constraints alongside greater computation speeds is of great interest. This thesis explores a methodology for using a slower-speed, robust kinematic/dynamic path planner for generating state path solutions, from which a recurrent neural network is trained upon. This path planning recurrent neural network is then used to generate state paths that a path-tracking controller can follow, trending the desired optimal solution. Improvements are made to the use of a kinodynamic rapidly-exploring random tree and a whole-path reinforcement training scheme for use in the methodology. Applications to 3 scenarios, including obstacle avoidance with 2D dynamics, 10-agent synchronized rendezvous with 2D dynamics, and a fully actuated double pendulum, illustrate the desired performance of the methodology while also pointing out the need for stronger training and amounts of training data. Last, a bounded set propagation algorithm is improved to provide the initial steps for formally verifying state paths produced by the path planning recurrent neural network.