Multi-Modal Multi-Task Deep Learning for Autonomous Driving

Several deep learning approaches have been applied to the autonomous driving task, many employing end-toend deep neural networks. Autonomous driving is complex, utilizing multiple behavioral modalities ranging from lane changing to turning and stopping. However, most existing approaches do not factor in the different behavioral modalities of the driving task into the training strategy. This paper describes a technique for using Multi-Modal Multi-Task Learning that considers multiple behavioral modalities as distinct modes of operation for an end-to-end autonomous deep neural network utilizing the insertion of modal information as secondary input data. Using labeled data from hours of driving our fleet of 1/10th scale model cars, we trained multiple neural networks to imitate the steering angle and driving speed of human control of a car. We show that in each case, our models trained with MTL can match or outperform multiple networks trained on individual tasks, while using a fraction of the parameters and having more distinct modes of operation than a network trained without MTL on the same multi-modal data. These results should encourage Multi-Modal MTL-style training with the insertion of Modal Information for tasks with related behaviors.