International Journal of Science, Engineering and TechnologyAuthors: Balaji M, Karpagavanayagam, Kishore Kumar A
Abstract: This project explores the development of an autonomous driving agent using Deep Reinforcement Learning (DRL), tailored for environments that simulate the complexity and unpredictability of Indian roads. Leveraging the Soft Actor-Critic (SAC) algorithm within a customized OpenAI Gym-compatible framework, the system is trained using the virtual simulation platform of Need for Speed: Most Wanted (2005). Real-time telemetry data is extracted through direct memory access, and control inputs are delivered via a virtual gamepad interface. The agent learns to optimize throttle, braking, and steering to minimize collisions, improve lap time, and maintain adherence to track paths. The training pipeline is structured in progressive phases, starting with basic behavioral learning and advancing to complex, dynamic scenarios that emulate heterogeneous traffic, poor lane discipline, and unexpected road conditions found in real-world Indian settings. Through rigorous simulation, the model demonstrates high adaptability and performance, offering a cost-effective, scalable, and safe alternative to real- world AV testing. This approach holds strong potential for future deployment in Indian traffic systems where conventional AV models struggle.
DOI: http://doi.org/
