SIMULATION-BASED BEHAVIOURAL CLONING FOR AUTONOMOUS VEHICLES USING AIRSIM

Abstract

Autonomous vehicles represent a transformative advancement in transportation technology, requiring robust testing and validation methodologies before real-world deployment. This research investigates behavioural analysis of autonomous driving systems using simulation-based deep learning techniques, specifically leveraging Unreal Engine 5 integrated with AirSim for high-fidelity synthetic environment creation. The study addresses the critical challenge of training AI models to navigate diverse road conditions, including degraded road surfaces such as potholes, cracks, and worn lane markings. We propose developing a comprehensive simulation environment featuring controlled test tracks representing both ideal and artificially degraded pavement conditions. A convolutional neural network (CNN) model, based on the established NVIDIA DAVE-2 architecture, will be trained using behavioural cloning methodology on data collected exclusively from this custom synthetic environment. The research aims to quantify the impact of road surface degradation on autonomous vehicle control performance, specifically analysing steering accuracy and control smoothness. By training and testing the model in a controlled simulation environment, this work provides empirical evidence for developing robust autonomous driving systems capable of handling visually challenging real-world scenarios. The findings will contribute to enhanced data augmentation strategies and normalization techniques, ultimately improving the safety and reliability of self-driving vehicles operating under diverse environmental conditions.