Publication

Autonomous driving is a major technological challenge, offering potential improvements in safety, efficiency, and comfort for daily transportation. However, the full-scale implementation of fully automated vehicles (AVs) remains distant, primarily due to safety concerns. Confined areas like ports, mines, and quarries present opportunities for early deployment of these vehicles, as they provide controlled environments with reduced safety risks from external factors. Effective coordination of fully automated vehicles in such settings is crucial, as it can increase productivity and possibly reduce the number of operating vehicles needed.

Optimization-based control methods are useful for planning AV operations, considering key operational constraints. However, these methods can be slow for real-time applications due to the complexity of solving the optimization problems involved, especially for coordinating multiple vehicles. This thesis introduces a method using optimization-based heuristics to simplify and approximate these problems.

The method involves a two-stage optimization approach for AV coordination in confined sites. Specifically, the combinatorial part of the coordination problem that is related to the occupancy orders of the conflict zones is formulated as a Mixed Integer Quadratic Program (MIQP). In the second stage, the optimal control commands for each vehicle are found under a fixed crossing order by solving a Nonlinear Program (NLP). To additionally improve the computational demand of the approach we propose a decomposition strategy based on a graph theory, where the centralized NLP is decomposed into multiple, parallelly solvable NLPs. Utilizing the Lagrange dual variables we propose a method that can further decompose the NLP and can be used to find a trade-off between improved computation time and optimality.

Finally, we adapted the optimization-based method to be able to handle the scenarios when human-driven vehicles (HDVs) are present in the confined site. Specifically, the heuristic predicts HDV behavior using a model that accounts for various human reactions. In cases where an HDV follows another vehicle, a car-following model is used, where the HDV's movement depends on the lead vehicle. This allows for partial control over the HDV movement, especially if the lead vehicle is an AV. In particular, the lead AV could slow down or speed up the HDV such that a desired occupancy order is achieved, resulting in a more efficient motion of the AV fleet. The MIQP is adapted to include HDV motion estimates and to determine if the HDV can use a car-following model. The NLP is modified to capture HDV movements and establish safety constraints between AVs and HDVs. Through closed-loop receding horizon control, we demonstrate how the occupancy order for the zones can be dynamically adapted to current conditions and HDV motion predictions.