Federated learning based decision making for autonomous driving in extreme scenarios

Abstract

Autonomous driving systems must make precise and reliable decisions to ensure safety and prevent collisions. In this paper, we address the challenge of effectively training autonomous vehicles to handle rarely encountered extreme scenarios in a virtual environment. To achieve this, we propose a hybrid approach that integrates reinforcement learning (RL) for autonomous driving within a federated learning (FL) framework. This approach enables individual vehicles to collaboratively develop a global model capable of handling diverse extreme tasks at intersections. Additionally, it allows local vehicles to train models in conjunction with roadside units (RSUs) without compromising sensitive data, such as perception information. Simulation results demonstrate that the proposed FL framework not only boosts the convergence speed during the training phase by up to 114.26%, but also improves autonomous driving performance, as evidenced by higher reward values, lower collision rates, and reduced travel time, compared to benchmark RL schemes.