Deep Reinforcement Learning Approaches in CARLA’s Autonomous Driving Simulation

The convergence of sophisticated algorithms and evolving simulation platforms holds promise for advancements in autonomous driving. Employing the CARLA simulator, our research undertakes a comprehensive exploration of three cardinal Reinforcement Learning (RL) algorithms: A3C, DDQN, and DDPG. Our findings unearth distinct behavioral patterns and proficiencies. The DDPG, undeniably audacious, pushes for higher speeds, yet occasionally strays from the optimal path, especially in challenging curves. In contrast, A3C’s wavering trajectory in both training and unfamiliar environments signifies a more cautious approach. The real revelation, however, is the DDQN which, in a nuanced ballet of speed and precision, consistently holds a central trajectory, signaling a harmonious blend of speed and safety. Notably, the embrace of Variational Autoencoder (VAE), inspired by prior works demonstrating its efficacy in expediting training, serves as an auxiliary catalyst in our experiments. This synthesis offers invaluable insights for future endeavors aiming to refine driving strategies in ever-complicated urban terrains.