Trajectory Following using Nonlinear Model Predictive Control and 3D Point-Cloud-based Localization for Autonomous Driving

In autonomous driving, the trajectory follower is one of the critical controllers which should be capable of handling different driving scenarios. Most of the existing controllers are limited to a particular driving scenario and for a specific vehicle model. In this work, the trajectory follower is formulated as a nonlinear model predictive control problem and solved using the multiple-shooting trajectory optimization method, Gauss-Newton Multiple Shooting. This solver has already been used for other control applications and provides the flexibility to use different nonlinear models. The controller is tested using a retrofitted autonomous driving platform, along with the 3D point-cloud-based mapping and localization algorithms. The nonlinear model being used is a classical kinematic bicycle model. Due to the high nonlinearity between the vehicle inputs, throttle and brake, and the acceleration, the longitudinal speed control uses an additional piece-wise linear mapping. The results from the initial tests, while following a predefined trajectory on a Go-Kart test-track, are evaluated and presented here.