DevDAgger Data Efficient Visual Imitation Learning Through Variational Autoencoders and Gaussian Process