Deep End-to-end Imitation Learning for Missile Guidance With Infrared Images

In this paper, we propose an end-to-end missile guidance algorithm that uses raw infrared image pixels by imitating a conventional guidance law. To bypass the need to extract the state from raw image data for an expert policy, our method leverages privileged data in training episodes. We demonstrate that this approach not only prevents performance degradation due to estimation error but also successfully imitates a conventional guidance law without expensive sensors. Our method shows successful imitation results even in noisy and agile environments through an image augmentation strategy. We also analyze the possibility of predicting the failure of the guidance through ensemble networks and show that the variance between ensemble networks can help predict the risk of the neural-network guidance system.