New preprint: Σ-LeWM: Learning World Models via S-Information for Robot Control. We combine forward, backward, and inverse dynamics to learn latent world models for more effective and efficient robot planning. [Project] · [Code]