Mind the Unknown
Risk- and Occlusion-Aware Motion Planning for Autonomous Vehicles
Time: Fri 2025-05-09 09.00
Location: Kollegiesalen, Brinellvägen 8, Stockholm
Video link: https://kth-se.zoom.us/j/62547376681
Language: English
Doctoral student: Truls Nyberg , Robotik, perception och lärande, RPL
Opponent: Dr Andrea Censi, Institute for Dynamic Systems and Control, ETH Zurich
Supervisor: Jana Tumova, ACCESS Linnaeus Centre, Centrum för autonoma system, CAS, Robotik, perception och lärande, RPL; Patric Jensfelt, Robotik, perception och lärande, RPL
QC 20250415
Abstract
Autonomous vehicles (AVs) must navigate uncertain environments while ensuring safety, particularly in scenarios involving risk and occlusions. This thesis develops structured approaches to risk- and occlusion-aware motion planning, integrating theoretical advancements with real-world validation.
To address risk in motion planning, we introduce a framework that quantifies both the probability and severity of safety violations, enabling AVs to reason about risk while maintaining operational efficiency. Complementing this, we investigate pedestrian-aware motion planning in urban environments, incorporating a harm-based risk model to balance safety and progress in interactions with vulnerable road users.
Occlusions pose a major challenge by limiting direct visibility of critical road users. We develop a method for tracking and reasoning about hidden obstacles using reachability analysis and formal logics. By incorporating prior observations, our approach systematically refines possible states of occluded agents, reducing unnecessary conservatism. For high-speed driving, we refine velocity bounds on occluded traffic participants, preventing worst-case assumptions that could lead to excessive braking. Additionally, we explore vehicle-to-everything (V2X) communication to enhance situational awareness, enabling AVs to infer and share information about occluded regions in real time.
Finally, we propose an occlusion-aware planning framework that integrates tree-based motion planning with reachability-based occlusion tracking. This enables AVs to proactively reason about future observations—or their absence—ensuring robust decision-making under limited sensing. By reducing overly conservative constraints while maintaining safety guarantees, our approach addresses key issues in occlusion-aware motion planning.
Together, these contributions advance the ability of AVs to operate safely and efficiently in demanding environments, supporting scalable real-world deployment.