End-to-End Deep Learning for Autonomous Steering Angle Prediction: A Dataset-Driven Approach

Trained on a dataset of 8,036 samples from several camera viewpoints, this work proposes an end-to-end deep learning method for guiding angle prediction in autonomous driving. With an eye toward forward-driving situations, the dataset comprises steering, throttle, and braking settings. Using Mean Squared Error (MSE), model performance was assessed, a low validation MSE of 0.0107 indicated reliable steering predictions. With kurtosis and skewness assessments stressing rare but important driving circumstances, dataset analysis revealed a balanced distribution of steering values and skewed throttle values. The model’s minimal overfitting and consistent training progress highlight its possible practical application. Future developments will try to increase resilience and lower sensitivity to outliers, hence supporting safer and more flexible autonomous driving systems.