Safety-Driven Path Selection Using Reinforcement Learning in Autonomous Driving
· Paper · Reinforcement Learning / Autonomous Driving / P2P / K-Shortest Path
Published at RSAE 2025.
Autonomous-driving route selection depends on timely road data, but reports can be incorrect, stale, or deliberately misleading. This work combines peer-to-peer reporting with reinforcement learning to estimate the credibility of road-segment reports before using them for routing.
We first identify accident-related parameters from several traffic datasets and estimate their relative importance with machine learning. Vehicles report conditions as they pass each road segment. An RL model evaluates reports inside a fixed time window and updates the trust score of each reporting vehicle, gradually reducing the influence of malicious or faulty sources.
The validated risk estimates are then combined with K-shortest paths. The vehicle selects among candidate routes in real time, balancing lower-risk road segments against travel efficiency.
Experiments show better safety and travel-time efficiency, particularly when the overall reporting error rate is low enough for the RL model to distinguish real events from false reports.
Keywords: autonomous driving, road conditions, data quality, reinforcement learning, P2P network.