Integrating a Neuro-Symbolic Planner for Safe Driving

Abstract

Ensuring the safety and reliability of autonomous driving systems remains an ongoing challenge. The complexity of real-world traffic scenarios highlights the importance of autonomous driving systems being able to reason about their surrounding environments and make safe decisions. This thesis proposes a modular neuro-symbolic framework that decouples trajectory prediction from motion planning. Rather than relying solely on an end-to-end deep-learning-based planner, the proposed approach uses a fine-tuned large language model to predict the future trajectories of surrounding actors. These predictions are incorporated as dynamic obstacles into a symbolic A* planner, which generates a goal-oriented trajectory for the ego vehicle subject to safety constraints. The proposed framework is evaluated on data generated from driving scenarios developed in Scenic and executed in the CARLA simulator, with its performance compared against an end-to-end deep-learning-based planner using the same set of 144 scenarios. The results demonstrate that this modular neuro-symbolic approach achieves a higher rate of safe, goal-reaching trajectories, with a success rate of 59.0% compared with 27.1% for the baseline planner. However, the neuro-symbolic planner also exhibits a higher collision rate of 32.6%, compared with 10.4% for the baseline planner. This difference is partly due to the baseline frequently producing safe but non-goal-reaching trajectories, with 57 scenarios falling into this category because the ego vehicle remains at least 100 meters away from the other actors throughout the scenario. These findings highlight a trade-off between collision avoidance and goal achievement, indicating the need for further evaluation to draw clearer conclusions regarding the safety of the approach.