Behavior Cloning Imitation Learning For Vision Transformer To Predict Steering Angle And Speed Of Self-Driving Vehicles

Abstract

Controlling a self-driving vehicle on highways is a complex and essential task that involves numerous challenges. Rule-based expert systems and traditional self-driving systems offer limited solutions when dealing with complex real-world scenarios. This has led to a shift towards using learning-based planning, supported by big data, particularly through behavior cloning for End-to-End Architectures. This study proposes the use of transformers that rely on self-attention mechanisms in computer vision, which have proven successful in natural language processing tasks and effective in patch-embedding colored images. The proposed model is trained on a comprehensive dataset that includes driving behaviors, environmental conditions, and vehicle dynamics, with the goal of predicting the steering angle and speed of self-driving vehicles on highways. The results show that the model achieves a significantly lower Mean Absolute Error (MAE) compared to traditional deep learning techniques, by using vision transformers. The paper introduces two models: the first uses a single Multi-Layer Perceptron Head with two output neurons for both the steering angle and vehicle speed, while the second uses two separate prediction heads with one output neuron for each. The mean absolute error for the first model was 0.175 and for the second model 0.37, while the lowest mean absolute error in studies using traditional self-driving systems was 0.491.