Developed a new Critic network architecture for the WE-Autopilot/Red-Team repository, targeting the f1tenth_gym reinforcement learning environment. The work introduced convolutional neural network-based state processing combined with fully connected layers to enhance the evaluation of state-action pairs, aiming to improve the stability and accuracy of policy updates. Leveraging PyTorch and deep learning techniques, the implementation incorporated an Adam optimizer with configurable learning rates and robust checkpointing to ensure reproducible and stable training. This approach addressed the need for more reliable value estimation in autonomous driving simulations, enabling faster experimentation and supporting ongoing research in reinforcement learning using Python.