Accurate steering angle prediction is crucial for smooth and safe autonomous driving, especially in complex and dynamic conditions. Many traditional models struggle with balancing accuracy and addressing non-convex optimization challenges, limiting their real-time effectiveness. To address this, we propose a customized Convolutional Neural Network (CNN) designed specifically for real-time steering angle prediction. The architecture incorporates multiple convolutional layers along with tailored pooling, dropout, and dense layers, which efficiently extract features while minimizing overfitting. For optimization, we utilize the NADAM optimizer, which combines Nesterov momentum with Adam’s adaptive learning, improving gradient anticipation and speeding up convergence. NADAM excels at resolving non-convex optimization problems, making it particularly suited for this task. In tests using a comprehensive dataset with diverse driving conditions, our model consistently outperformed conventional methods, achieving a minimum loss of 0.061. The use of average pooling in the final layers further improved generalization and reduced overfitting, making the model highly scalable for real-time self-driving applications.