This paper presents a deep reinforcement learning (DRL) framework for autonomous driving. The framework integrates recurrent neural networks (RNNs) and attention models to handle partially observable scenarios and reduce computational complexity for embedded systems. The framework was tested in the open-source 3D car racing simulator TORCS, demonstrating successful learning of autonomous maneuvering in complex road conditions. Autonomous driving involves three main tasks: recognition, prediction, and planning. Recognition involves identifying objects in the environment, prediction involves forecasting future states, and planning involves generating a sequence of actions to navigate safely. DRL combines reinforcement learning (RL) with deep learning (DL) to achieve human-level control in autonomous driving. The framework uses deep Q-networks (DQN) and deep recurrent Q-networks (DRQN) for action learning, and integrates attention models to focus on relevant information. The attention models help reduce computational complexity by filtering out non-relevant data, making the framework suitable for real-time embedded systems. The framework was tested in the TORCS simulator, where it successfully learned to maintain lane position in complex driving scenarios. The results show that the framework can handle partial observability and reduce computational requirements, making it suitable for real-world autonomous driving applications. The paper also discusses the use of apprenticeship learning with demonstrated expert behavior, where the agent learns from expert actions to understand the reward function. This approach is particularly useful in scenarios where the reward function is not explicitly known. The proposed framework integrates RNNs and attention models to handle partially observable scenarios, and has been tested in the TORCS simulator. The results show that the framework can successfully learn autonomous driving tasks, demonstrating the potential of DRL in autonomous driving. Future work includes deploying the framework in a simulated environment with labeled ground truth for real-world applications.This paper presents a deep reinforcement learning (DRL) framework for autonomous driving. The framework integrates recurrent neural networks (RNNs) and attention models to handle partially observable scenarios and reduce computational complexity for embedded systems. The framework was tested in the open-source 3D car racing simulator TORCS, demonstrating successful learning of autonomous maneuvering in complex road conditions. Autonomous driving involves three main tasks: recognition, prediction, and planning. Recognition involves identifying objects in the environment, prediction involves forecasting future states, and planning involves generating a sequence of actions to navigate safely. DRL combines reinforcement learning (RL) with deep learning (DL) to achieve human-level control in autonomous driving. The framework uses deep Q-networks (DQN) and deep recurrent Q-networks (DRQN) for action learning, and integrates attention models to focus on relevant information. The attention models help reduce computational complexity by filtering out non-relevant data, making the framework suitable for real-time embedded systems. The framework was tested in the TORCS simulator, where it successfully learned to maintain lane position in complex driving scenarios. The results show that the framework can handle partial observability and reduce computational requirements, making it suitable for real-world autonomous driving applications. The paper also discusses the use of apprenticeship learning with demonstrated expert behavior, where the agent learns from expert actions to understand the reward function. This approach is particularly useful in scenarios where the reward function is not explicitly known. The proposed framework integrates RNNs and attention models to handle partially observable scenarios, and has been tested in the TORCS simulator. The results show that the framework can successfully learn autonomous driving tasks, demonstrating the potential of DRL in autonomous driving. Future work includes deploying the framework in a simulated environment with labeled ground truth for real-world applications.