Collaborative CNN-Based Federated Learning for Steering Control in Diverse Driving Conditions

The rapid evolution of autonomous vehicular technologies demands advanced solutions for reliable navigation in diverse and often unfamiliar environments. This paper introduces a novel CNN-based federated learning approach for vehicular systems, designed to enhance their self-driving adaptability and decision-making capabilities in novel environments through collaborative training. Our method leverages the decentralized nature of FL to enable vehicles to learn collectively from shared experiences while maintaining the privacy of individual data. We utilize the CARLA simulation environment to generate a wide range of driving scenarios, including multiple vehicles operating under varied weather conditions. This diverse dataset serves as a testbed for our FL framework, allowing us to evaluate its effectiveness in accurately predicting steering wheel angles while simultaneously adapting to different environmental challenges. Our results demonstrate that vehicles trained using our FL framework exhibit enhanced performance in predictive analytics, showing greater resilience to environmental changes and improved decision-making in real-time scenarios with a much lower computational complexity.