Disney researchers have managed to make a robot learn strategies from unstructured motion data. First, they extracted a latent spatial encoding by training a variational autoencoder and then used short motion windows from unstructured data as input.

Next, they used the embedding of the time-dependent latent code to train a conditional policy in a second step, enabling a mapping from kinematic inputs to dynamic outputs. By keeping these two steps separate, the team was able to benefit from self-supervised methods to obtain better latent codes and explicit imitation rewards to avoid mode collapse. The efficiency and robustness was demonstrated in simulation with unseen, user-specified movements and on a bipedal robot where dynamic movements were transferred to the real world.