OnlineLearningandControlofComplexDynamical SystemsfromSensory Input-Supplementary Material

The decoder is a symmetric copy of the encoder. As our images are64 64, the last convolutional block yields afeature map with 8channels and1 1spatial dimension, which is reshaped into an8 1 vector. The latent code we consider is thus directly the output of the convolutional encoder. Toobtain videos of actuated systems, we have generated a set of reference trajectories and velocities to be followedbythesystems. Our model does not exhibit such limitations.