On Thursday 1 October 2026, 14:00-15:00, MF 3.122, Huaze will defend his Master thesis project titled
Shielding for Autonomous Highway Driving with STARK
Abstract: Reinforcement-learning-based autonomous driving agents can achieve
effective driving behaviours, but their decisions are difficult to verify and may
lead to unsafe situations. This thesis studies runtime safety shielding for the ABZ
2025 autonomous highway driving case study, focusing on a three-lane highway
setting in which pre-trained AI agents control an ego vehicle.
This thesis uses STARK to construct runtime shields for these AI agents. Before
each AI-proposed action is applied, the shield evaluates it against predicted future
traffic evolutions and rejects it if the prediction indicates an unsafe outcome.
The safety requirements are expressed using DisTL formulae, which compare
predicted behaviours with desired safety-related behaviours.
Two shield designs are implemented. The first shield uses a heuristic prediction
model based on simple assumptions about surrounding vehicles. The second
shield uses randomized IDM–MOBIL-based prediction to model possible future
reactions of nearby vehicles. Experimental results show that both shields substantially
reduce the collision rate of the adversarial AI agent. The recoverability-based
shield achieves a better balance between safety and operational performance
than the heuristic shield in our experiment setting. These results suggest
that STARK can provide a useful framework for constructing runtime shields that
combine predicted future behaviours with formal temporal logic evaluation.