PhD defense Aloysio Galvão Lopes: Trajectory Planning in the Presence of Obstacles, Prediction and Real-Time Scheduling
LIX (bâtiment Turing, 1 Rue Honoré d’Estienne d’Orves, F-91120 Palaiseau, France) amphi Sophie Germain
and in videoconferencing
Jury
- Matthias Althoff, Professor, TUM School of Computation Information and Technology (Reviewer)
- Liliana Cucu-Grosjean, Research Director, INRIA (Reviewer)
- Jesse Read, Professor, École polytechnique (Examiner)
- Frank Singhoff, Professor, Université de Bretagne Occidentale (Examiner)
- Éric Goubault, Professor, École polytechnique (Thesis Supervisor)
- Laurent Pautet, Professor, Télécom Paris (Co-supervisor)
- Sylvie Putot, Professor, École polytechnique (Guest)
Abstract
Despite the growing interest, fully driverless operation of autonomous vehicles (AVs) remains limited. In this context, our research develops methods that enable AVs to operate safely while efficiently managing their hardware. We focus on two core sources of uncertainty: (i) the behavior of surrounding vehicles and (ii) the variable execution times of tasks within the autonomous system. The former relates to prediction, which is essential for safety, while the latter concerns real-time system operation, crucial for efficient resource allocation and timely reactions.
To address the prediction problem, we leverage recent advances in conformal prediction, a theory for uncertainty quantification that provides prediction zones with formal probabilistic guarantees. In this context, we introduce ConForME, a method for computing tight prediction zones around multi-step predictions from neural networks. ConForME advances the state of the art by reducing prediction zone sizes considerably.
Building on this, we propose PROSPERS, a framework that combines ConForME with mixed-criticality scheduling. By switching between criticality modes based on probabilistic trajectory predictions, PROSPERS ensures safe mode transitions. This approach not only improves resource allocation, but also bridges a key gap by addressing both uncertainties simultaneously. Additionally, we develop a Rapidly-Exploring Random Tree (RRT)-based planning algorithm that computes dynamically feasible short-term plans and provide a proof of its probabilistic completeness.
Alongside PROSPERS, we present a modular ROS 2 implementation, which is compatible with CommonRoad scenarios and we use it to simulate safe overtaking maneuvers. In conclusion, this thesis provides a comprehensive perspective on probabilistic safety and efficient resource allocation for AVs, paving the way for more reliable and adaptable autonomous driving systems in complex real-world environments.