Best Dissertation Award - IEEE ITS Society
Holger Banzhaf was awarded with the Best Dissertation Award for his dissertation named Nonholonomic Motion Planning for Automated Vehicles in Dense Scenarios
Motion planning is one of the crucial components in the software stack of an automated vehicle. It is responsible for the computation of a safe and preferably optimal trajectory from a given start state to a desired goal. While a local solution to this problem is sufficient for highway driving, this thesis focuses on the computation of a global solution, which is typically required to handle unstructured environments or complex maneuvers. Relevant scenarios include dead ends, blocked lanes, or various parking problems that have proven difficult for automated vehicles to solve, particularly when space is tight. The contributions of this thesis can be grouped into three parts. The first part focuses on steering functions for car-like robots, which play a major role in both search-based and sampling-based motion planning. Within this context, the novel steering function hybrid curvature (HC) steer is introduced that computes smoother paths than the well-known Reeds-Shepp steering function  and shorter paths than continuous curvature (CC) steer . Especially in tight environments, HC steer proves to be a powerful tool for the computation of directly executable motion plans with continuous curvature between direction switches. In addition to that, the two novel steering functions continuous curvature rate (CCR) and hybrid curvature rate (HCR) steer are presented that extend the smoothness of both CC and HC steer from curvature to curvature rate continuity. This allows to increase the comfort for passengers as well as the tracking performance of the low level motion controller. The second part of this thesis focuses on motion planning under uncertainty aiming to improve the robustness of the motion plans by explicitly considering the localization and control errors of the system. For this purpose, the previously mentioned steering functions are extended to belief space in which every vehicle state is associated with its respective uncertainty. Furthermore, two novel algorithms for probabilistic collision checking are introduced in order to bound the collision probability of the computed vehicle motion. The third part addresses the problem of slow convergence in sampling-based motion planning if samples are only drawn from a uniform distribution. To overcome this problem, a data-driven approach is presented that utilizes a convolutional neural network to predict a distribution over future vehicle poses given the current environment and the boundary conditions of the planning problem. Samples from this distribution can then be used to bias the motion planner towards promising regions in the state space allowing to improve the planning performance in complex scenarios. Finally, the proposed methods from all three parts are integrated into the sampling-based motion planner RRT*  and its bidirectional extension BiRRT*  to demonstrate their benefits in a broad set of challenging environments. The motion planner is not only tested in simulation, but also integrated into a research vehicle proving its effectiveness in real-world applications.
From the research group Angewandte Technisch-Kognitive Systeme