Hierarchical World Models for Visual Planning

H-JEPA is a novel hierarchical world model designed to improve visual planning and long-horizon control without relying on reward functions or pixel reconstruction. By organizing action-conditioned Joint-Embedding Predictive Architectures into multiple temporal layers, the system processes information at varying scales where higher levels handle abstract states and lower levels manage rapid details. This top-down hierarchy allows an agent to decompose complex tasks into simpler subgoals and discard unpredictable fast-varying features while preserving slow-moving traits. Across various simulated environments, hierarchical planning significantly outperforms single-level models while requiring fewer compute resources. Furthermore, the framework successfully adapts to real-robot manipulation tasks when combined with inverse-dynamics supervision