I don’t see a problem. It appears to use a shared latent to learn inverse dynamics.