Performance Analysis of RL Based End-to-End Learning Technology for Autonomous Driving

As autonomous driving technology is being researched as a core technology for the future automotive industry, there has been rapid advancement in related sensors, components, platforms, and artificial intelligence technologies. The core technology of autonomous driving, based on such hardware and platforms, lies in AI-driven software, with various studies being conducted in this field. In particular, key technologies for advancing autonomous driving include the mechanism of interaction among modules: the perception module that recognizes the ego vehicle's situation from multiple sensor inputs, the decision module that plans driving strategies based on the recognized situation, and the control module that operates physical components such as acceleration and steering based on the driving plan. This also includes the orchestration technology for feedback control in response to changing environments. This paper focuses on E2E (End-to-End Learning), an integrated control module encompassing perception, decision, and control modules, specifically investigating the performance of reinforcement learning in the control section. The research places particular emphasis on evaluating end-to-end learning performance using the CARLA online simulator (Car Learning to Act). It examines the neural network structure of TCP (Trajectory-guided Control Prediction) and trains and validates it under identical conditions. The results show an increase in performance when comparing a simple reinforcement learning model to the original: Driving Score improved by 4.81%, Route Completion by 1.91%, and Infraction Penalty remained at 0.0%.