Latent world models offer a way to train end-to-end driving systems by predicting future states, providing self-supervision without explicit labels. But according to a new arXiv preprint, these models typically inherit dense features from pretrained visual encoders as their world-state interface. The paper argues that this creates a mismatch between what the visual encoder represents and what the driving task actually requires.

To address this, the authors propose MORE (Motion-Oriented Registers), which learns a set of motion-oriented registers to serve as the world-state interface. Rather than relying on dense visual features, these registers are designed to capture information relevant to motion and planning, potentially making the world model more efficient and better suited for downstream driving decisions.

The abstract is truncated in the source, so details on experimental results or comparisons are not available. The paper is presented as a new submission to arXiv, and the proposed approach is described as a way to close the gap between visual feature representations and the motion-centric needs of driving world models.