Latent World Models with Monotone Planning Costs for Image-Goal Navigation
Overview
This research presents a latent world model designed for image-goal navigation that utilizes a frozen DINO-family encoder. The authors propose two training objectives: an autoregressive rollout loss to align multi-step planning rollouts and a Monotone Cost Ranking (MCR) loss to ensure action sequences with higher perturbations receive higher planning costs, preventing misleading sampling-based planners like Cross-Entropy Method. By avoiding InfoNCE-based action-contrastive training which distorts latent geometry, the method achieves state-of-the-art performance on the GNM navigation dataset, reducing orientation error by 2.7 times compared to a DINO WM baseline with the same encoder. The model is also demonstrated in zero-shot deployment on a physical robot, successfully following goal-directed paths in both indoor and outdoor environments.
Generated from attributed reports · Updated 37 minutes ago
Event evidence and corrections
0 attributed source owners. Ownership does not establish independent confirmation. Quantities are reported separately and are never added together.
No current evidence-backed claims. Missing information remains not reported.
Report timeline
Follow attributed reports and material updates.
- arXiv Robotics — research abstractsSignalLatent World Models with Monotone Planning Costs for Image-Goal Navigation
This research presents a latent world model designed for image-goal navigation that utilizes a frozen DINO-family encoder. The authors propose two training objectives: an autoregressive rollout loss to align multi-step planning rollouts and a Monotone Cost Ranking (MCR) loss to ensure action sequences with higher perturbations receive higher planning costs, preventing misleading sampling-based planners like Cross-Entropy Method. By avoiding InfoNCE-based action-contrastive training which distorts latent geometry, the method achieves state-of-the-art performance on the GNM navigation dataset, reducing orientation error by 2.7 times compared to a DINO WM baseline with the same encoder. The model is also demonstrated in zero-shot deployment on a physical robot, successfully following goal-directed paths in both indoor and outdoor environments.
Event coverage history
There is not enough continuous observation data to show a trend.
Timezone · UTC
Article dates follow your selected timezone. Briefing editions use Hong Kong time (UTC+8).