Is the fundamental challenge for AI directed movement in 3D space the continuity of the environment? Vs LLM's which train on discrete and limited token sets.