VEHICLE DYNAMICS MODELING AND MOTION PLANNING WITH PREDICTABLE TIMING AND VELOCITY

Abstract

Intelligent transportation systems and autonomous vehicles can improve how we drive, save energy and avoid traffic accidents. However, we still need to make more effort on the process of interdisciplinary research to turn these systems into reality. My work, aimed to contribute to this process, consists of two parts. In the first part, I address a longitudinal motion planning problem in which a vehicle aims to arrive at a given position on a road segmented based on various driving conditions at a given time and velocity. I show that it is possible to fully describe the set of all reachable arrival configurations using a table of closed-form equations, under a simplified model of vehicles with linear acceleration. Then I devise a sampling-based algorithm to solve this motion planning problem, using the table to check whether a feasible plan exist. The simulation results showed that the proposed sampling-based algorithmwith heuristics has a higher probability of success than the simple random sampling approach. After that, I will discuss how to use the feasible set on real vehicles. In this part, the task of planning on a segment is executed by using the bisection method, which utilizes the dynamics model of a vehicle. This model, however, is not always available and empirically expensive to obtain. Therefore, in the second part, I focus on the problem of learning a vehicle dynamics model. I introduce an instance-based learning method to learn a performance model automatically, and compare it with the artificial neural network and matrix factorization methods. Furthermore, an exploration strategy called plan-based exploration, based on planning using the reference model is given to speed up the learning process. Our experimental results demonstrated that the instance-based learning method coupled with the plan-based exploration strategy has the fastest learning rate.