This paper presents a deep reinforcement learning (DRL) framework for autonomous driving, addressing the challenges of handling complex environments and interactions with other vehicles. The authors motivate their approach by highlighting the limitations of traditional supervised learning methods in autonomous driving, where strong interactions with the environment are common. They propose a framework that integrates Recurrent Neural Networks (RNNs) and attention models to handle partially observable scenarios and reduce computational complexity for embedded systems. The framework is tested in the open-source 3D car racing simulator TORCS, demonstrating successful learning of autonomous maneuvering in scenarios with complex road curvatures and interactions with other vehicles. The paper also provides a review of recent advances in DRL, including Q-learning, Deep Q Networks (DQN), Deep Deterministic Actor Critic (DDAC), and Deep Recurrent Q Networks (DRQN). The authors conclude by discussing future work, including deploying the framework in simulated environments with labeled ground truth to extend it to real driving scenarios.This paper presents a deep reinforcement learning (DRL) framework for autonomous driving, addressing the challenges of handling complex environments and interactions with other vehicles. The authors motivate their approach by highlighting the limitations of traditional supervised learning methods in autonomous driving, where strong interactions with the environment are common. They propose a framework that integrates Recurrent Neural Networks (RNNs) and attention models to handle partially observable scenarios and reduce computational complexity for embedded systems. The framework is tested in the open-source 3D car racing simulator TORCS, demonstrating successful learning of autonomous maneuvering in scenarios with complex road curvatures and interactions with other vehicles. The paper also provides a review of recent advances in DRL, including Q-learning, Deep Q Networks (DQN), Deep Deterministic Actor Critic (DDAC), and Deep Recurrent Q Networks (DRQN). The authors conclude by discussing future work, including deploying the framework in simulated environments with labeled ground truth to extend it to real driving scenarios.