Speeding up deep neural network-based planning of local car maneuvers via efficient B-spline path construction

Abstract-- This paper demonstrates how an efficient representation of the planned path using B-splines, and a construction procedure that takes advantage of the neural network's inductive bias, speed up both the inference and training of a DNN-based motion planner. We build upon our recent work on learning local car maneuvers from past experience using a DNN architecture, introducing a novel B-spline path construction method, making it possible to generate local maneuvers in almost constant time of about 11 ms, respecting a number of constraints imposed by the environment map and the kinematics of a car-like vehicle. I. INTRODUCTION Although autonomous vehicles are researched intensively, Learning through the interaction seems to carry out the most research on motion planning for these vehicles focuses important information to improve the performance of the mostly on managing traffic scenarios and rules [1], [2], paying trained system [7], while it does not impose any upperbounds less attention to the local maneuvers that are necessary on it, unlike supervised learning, which performance to park a car in a crowded city center, to enter a shopping is bounded by the quality of the reference trajectories or mall's garage, or to avoid a collision with another car that human demonstrations. Human drivers perform Although our previously introduced DNN [5] keeps the such local maneuvers intuitively, leveraging the experience path computation time below 50 ms, some emergency maneuvers from similar situations they have encountered in the past. Therefore, we contribute in this seconds) avoiding collisions in dangerous situations, and paper a novel path parametrization and procedure of its satisfying the constraints of a car-like vehicle. A car is nonholonomic, construction, which enables our method to compute yet has a limited steering angle and some physical better paths in an even shorter time in comparison to dimensions, while the planned path should allow control [5] Although these requirements planning function, breaks up with the Markov Decision call for a solution that is rather a reactive behavior than Process formalism used in [5], instead plans the whole a classic planning algorithm, reactive methods [3] rarely maneuver at once.