The paper demonstrates that existing hierarchical planning methods using (JEPAs) often generate physically unrealisable sub-goals when trying to plan over long horizons. The authors introduce Metro-WM, which instead of generating latent sub-goals, retrieves them from a graph of previously observed frames from offline demonstrations or random-action trajectories. This approach guarantees that sub-goals correspond to real states.
In experiments, Metro-WM achieved up to 37.33 percentage points higher success rates than the next best hierarchical method, while being up to 10.9 times faster and requiring 13-56 times less offline compute. The system also found shorter paths than the original demonstrations and performed robustly even with very sparse data.