H-WM: Hierarchical World Model for TAMP

Key Claims

• Two-level world model — high-level logical (symbolic) predictor + low-level visual predictor operating in tandem
• Mitigates error accumulation — the symbolic layer provides a stabilizing prior that limits how far visual rollouts can drift
• Targets TAMP (Task and Motion Planning) — multi-step manipulation problems that have historically resisted end-to-end learning

Why This Matters

Matches LeCun's hierarchical planning thesis — he argues that a true world model must operate at multiple levels of abstraction (airport trip: "go to airport" at high level, "move left foot" at low level). H-WM is one operationalization of that argument for robotics: the symbolic layer handles "pick up cup, move to sink, release" while the visual layer handles pixel-level outcomes.

Notes

First-pass stub. Key robotics-application anchor for the world-models concept page.

Source: H-WM