Latent Evolution Operator Network (LEON) for Robot World Action Models

LEON replaces Transformer-based latent transitions with context-modulated operator-based propagation grounded in Koopman generator theory. This architecture improves how robot policies predict scene state evolution by modeling temporal dynamics through learned observable spaces rather than simple token interaction.