Behavioral Cloning for Lateral Motion Control of Autonomous Vehicles using Deep Learning

Current trend of the automotive industry combined with research by the major tech companies has proved that self-driving vehicles are the future. With successful demonstration of neural network based autonomous driving, NVIDIA has introduced a new paradigm for autonomous driving software. The biggest challenge for self-driving cars is autonomous lateral control. An end-to-end model seems very promising in providing a complete software stack for autonomous driving. Although this system is not ready to be provided as a feature in the market today, it is one of the many steps in the right direction to make self-driving cars a reality. The work described in this paper focusses on how an end-to-end model is implemented. The subtleties of training a successful end-to-end model are highlighted with the aim of providing an insight on deep learning and software required for neural network training. Detailed analyses of data acquisition and training systems are provided and installation procedures for all required tools and software discussed. TORCS is used for developing and testing the end-to-end model. Approximately ten hours of driving data was collected from two different tracks. Using four hours of data from a track, we trained a deep neural network to steer a car inside simulation. Even with such a small training set, the end-to-end model developed demonstrated capabilities to maintain lanes and complete laps in different tracks. For a multilane track, like the one used for training, the model demonstrated an autonomy of 96.62% For single lane unknown tracks, the model steered the vehicle successfully for 89.02% of the time. The results indicate that end-to-end learning and behavioral cloning can be used to drive autonomously in new and unknown scenarios.