Search papers, labs, and topics across Lattice.
The paper introduces GeoWorld, a geometric world model that addresses limitations of Euclidean latent spaces in energy-based predictive world models by mapping latent representations to hyperbolic manifolds using a Hyperbolic JEPA. This approach aims to better capture geometric and hierarchical relationships among states, leading to improved long-horizon prediction. Experiments on CrossTask and COIN datasets demonstrate a 3% and 2% improvement in success rate for 3-step and 4-step planning, respectively, compared to V-JEPA.
Escape Euclidean limitations: GeoWorld uses hyperbolic geometry to learn better world models, boosting multi-step planning success rates.
Energy-based predictive world models provide a powerful approach for multi-step visual planning by reasoning over latent energy landscapes rather than generating pixels. However, existing approaches face two major challenges: (i) their latent representations are typically learned in Euclidean space, neglecting the underlying geometric and hierarchical structure among states, and (ii) they struggle with long-horizon prediction, which leads to rapid degradation across extended rollouts. To address these challenges, we introduce GeoWorld, a geometric world model that preserves geometric structure and hierarchical relations through a Hyperbolic JEPA, which maps latent representations from Euclidean space onto hyperbolic manifolds. We further introduce Geometric Reinforcement Learning for energy-based optimization, enabling stable multi-step planning in hyperbolic latent space. Extensive experiments on CrossTask and COIN demonstrate around 3% SR improvement in 3-step planning and 2% SR improvement in 4-step planning compared to the state-of-the-art V-JEPA 2. Project website: https://steve-zeyu-zhang.github.io/GeoWorld.
