Navigation World Model
Compact future-view prediction for navigation, summarized by quality and latency metrics.
This project studies compact future-view prediction for embodied navigation. The teaser compares a reference rollout, a baseline, the final model, and a faster variant on the same navigation prompt.
On the current shared validation slice, the fast variant stays close to the final model in decoded-frame quality while reducing loaded-model latency from 2.56s/sample to 0.53s/sample. The comparison is still intentionally small and used here as a compact demo rather than a full benchmark.