ATLAS: Aligned Transport of Latent Structure for Reliable World Model Planning
ATLAS targets the latent geometry that planners actually use, not just the overall latent distribution.
The paper argues that regularizing only the latent marginal can weaken state-to-state novelty structure before planning. ATLAS preserves normalized pairwise relationships from an encoder representation while using Wasserstein embedding matching to calibrate the planning latent. In LeWM, it improves mean goal-reaching success on PushT, TwoRoom, and OGBench-Cube, with the largest reported gain on higher-novelty TwoRoom episodes. Diagnostics show stronger novelty structure, better marginal calibration, and lower multi-step prediction error. HF Daily Papers' note
The paper argues that regularizing only the latent marginal can weaken state-to-state novelty structure before planning. ATLAS preserves normalized pairwise relationships from an encoder representation while using Wasserstein embedding matching to calibrate the planning latent. In LeWM, it improves mean goal-reaching success on PushT, TwoRoom, and OGBench-Cube, with the largest reported gain on higher-novelty TwoRoom episodes. Diagnostics show stronger novelty structure, better marginal calibration, and lower multi-step prediction error. HF Daily Papers' note
score 4