F1TENTH: An Over-taking Algorithm Using Machine Learning

In this paper we report on the development of a novel over-taking algorithm of cars in a simulated environment. The algorithm uses machine learning techniques, specifically recurrent neural networks (RNNs) and dense neural networks. We take LiDAR data, current speed, and current steering angle as input and produce control information in the form of output speed and steering angle. We obtain the training data by monitoring a human driver over-taking a car, controlled by a model predictive control (MPC) algorithm. After having trained several models (using Keras and Tensorflow), two unseen racetracks are used for evaluating the models. We set up experiments on these two racetracks in the simulator, to test whether the models can overtake in different and unseen cases. The best model (simple RNN) can pass 84 out of 90 cases on both racetracks. We identify faster training and lower risk of overfitting as key advantages for RNNs compared to other NNs we explored.