Paper Title
Deep Reinforcement Learning For Autonomous Vehicle Navigation: A Comprehensive Study Using Proximal Policy Optimization and VariationalAutoencoders in Carla Simulation Environment

Abstract
One of the biggest challenges faced in contemporary artificial intelligence studies has been the advancement of autonomous driving systems. The article captures a detailed study of Deep Reinforcement Learning (DRL) on piloting autonomous cars on a new brew of Proximal Policy Optimization (PPO) and VariationalAutoencoders (VAE) on the CARLA simulator. The analysis overrides the limitations posed by the more traditional modular philosophy and the imitation learning methods by including an end-to-end learning system that construes the sensory input in the environment as direct translations to the vehicle control actions. Its methodology uses an advanced state representation of a synthetically enhanced state-space that considers images captured by the front camera, fragmented by semantics and coded by a trained VAE alongside external vehicle inputs such as speed, past steering angles, distance to journey waypoints, and facing alignments. Its action space is continuous with the PPO agent producing steering and throttle commands as it drives along a predetermined route in two townships of different complexity levels in CARLA. A well-engineered dense reward function motivates ideal driving because it includes speed-keeping (target: 20 km/h), lane keeping (maximum offset: 3 m), and orientation matching (maximum angular offset: 20◦). Experimentation shows that autonomous navigation was achieved successfully at 750-meter and 780-meter pre-planned routes at Town 7 and Town 2, respectively, with convergence reached after 900–1200 training episodes. Its encoded state representation, based on VAE using a 95-dimensional latent space of 160×80 pixel semantically segmented images, achieved substantial speed-up of learning relative to raw image inputs. The asynchronous training paradigm (with early termination condition and recovery contingency mechanisms) increased the efficiency of the training method by 34% over its synchronous counterparts. Average lane deviation was less than 1.2 meters, and convergence success rates were above 87%, confirming that the proposed architecture is a viable solution to complex urban navigation situations without dynamic traffic participants. Keywords - Deep Reinforcement Learning, Autonomous Driving, Proximal Policy Optimization, VariationalAutoencoder, CARLA Simulation, Computer Vision, Policy Gradient Methods, End-to-End Learning