[CVPR 2024] GeoWorld: Solving Long-Horizon Planning with Hyperbolic Geometry

TL;DR

Predicting the future is hard, but planning for it over long steps is even harder. GeoWorld introduces a shift from flat Euclidean latent spaces to curved Hyperbolic manifolds. By leveraging the natural hierarchical structure of decision trees, it prevents the "geometric drift" that plagues current SOTA world models like V-JEPA 2, delivering significant boosts in multi-step visual planning success rates.

Problem & Motivation: The Euclidean Bottleneck

In the quest for autonomous intelligence, "World Models" allow agents to imagine future outcomes. Unlike generative models that waste bits on pixel-perfect reconstruction, Predictive World Models (like JEPA) work in latent space.

However, there is a fundamental mismatch:

1. Geometric Neglect: The real world of possibilities branches exponentially. Euclidean space—the "flat" default for most AI—cannot efficiently represent this hierarchical tree without distorting distances.
2. Multi-step Shortcoming: As an agent plans 5, 10, or 20 steps ahead, small prediction errors in Euclidean space compound into massive "drifts," causing the plan to lose its physical and logical coherence.

The authors' insight? The future is a tree, and trees live in Hyperbolic space.

Methodology: H-JEPA and Geometric RL

GeoWorld re-imagines the world model through two core innovations:

1. Hyperbolic JEPA (H-JEPA)

Instead of mapping a video frame to a flat vector, H-JEPA uses the Poincaré Ball model. It projects visual features onto a hyperbolic manifold using an exponential map. In this space, the distance between two states isn't a straight line; it's a geodesic (a curve that accounts for the space's negative curvature).

2. Geometric Reinforcement Learning (GRL)

To refine the predictor for long-haul stability, the authors treat planning as energy minimization. They introduce a unique regularization based on the Triangle Inequality. By ensuring that $d(s_1,s_3)\le d(s_1,s_2)+d(s_2,s_3)$ in the hyperbolic sense, they force the model to learn "geodesic-consistent" trajectories that don't warp over long horizons.

Experiments: Superior Long-Horizon Stability

The model was tested on CrossTask and COIN—benchmarks involving complex human activities like "Replacing a Memory Chip."

Performance Gains

GeoWorld achieved a SOTA success rate, particularly as tasks became more complex. While Euclidean models (V-JEPA 2) saw their accuracy plummet when the planning horizon $T$ moved from 3 to 8, GeoWorld remained robust.

Visualizing the energy landscape: GeoWorld (Right) shows a much sharper, better-defined energy basin than V-JEPA 2 (Left).

Key Results:

• Success Rate: +3% in 3-step planning.
• Long Horizon (T=8): While standard models collapsed to <5% SR, GeoWorld's SFT+GRL approach maintained nearly 14%—nearly a 3x relative improvement in extreme scenarios.

Critical Analysis & Conclusion

GeoWorld proves that Inductive Bias matters. By explicitly embedding the hierarchical nature of time and choice into the geometry of the latent space, we get planners that don't just "guess" the next step but understand the global manifold of the task.

Limitations: Currently, the model relies on video-based goal observations. Future work will likely explore how to integrate language instructions (e.g., "Fix the sink") into this hyperbolic latent structure.

Takeaway: If your data has a hierarchy—be it a file system, a chemical structure, or the branching futures of a robot—stop treating it as flat. The future of planning is curved.