DiffWAM: A Fast and Efficient Navigation World Action Model

Can the motion implicit in future visual prediction be recovered directly from the predictive representations of a frozen video model, rather than by generating future videos and reconstructing geometry?

Using pretrained video foundation models for embodied UAV navigation usually requires costly future-video synthesis and geometric reconstruction, which is too expensive and slow for real-time continuous flight.

Earlier methods converted video-model priors into UAV motion by generating future video rollouts and performing geometric reconstruction, then inferring motion from those outputs. Their shortcoming is the high computational cost and latency of decoding future frames and reconstructing multi-frame geometry during deployment.

DiffWAM is a geometry-conditioned navigation world-action model that maps multi-level predictive features from a frozen video model directly to continuous camera trajectories. A Grid-Motion module preserves spatial-temporal motion associations, and a Latent2Pose module grounds those features with first-frame geometry to produce metrically meaningful 3D motion. During deployment it does not decode future video or reconstruct multi-frame geometry; those are used only for offline supervision. FastDreamer overlaps predictive and geometric computation with ongoing flight and performs timestamp-aware asynchronous trajectory handoff for continuous UAV execution.

The abstract reports a trajectory RMSE of 0.3492 m. It also begins to state an endpoint success metric but the abstract is truncated, so that value is not fully reported. No baselines, trial counts, or simulation/real-world split are given in the abstract.

The abstract does not list explicit limitations. Evident limitations include a single reported trajectory error, an incomplete endpoint success metric, no baseline comparison, no stated number of trials or hardware platform, and no code or compute details. As a preprint, the results have not yet been peer-reviewed.