What it is
An 8B VLM for language-guided robot navigation that uses only monocular RGB, predicts image-space waypoints, and trains efficiently from large-scale simulation. Instead of depth, multi-camera, or robot-specific coordinates, it navigates by pointing to the next waypoint in image space from monocular RGB, with displacement fallback, plus prefix-tree training and RL for recovery.
Key results
- R2R-CE: 77.4% SR, +10.5 pts over best monocular baseline
- R2R-CE: beats best depth/multi-camera system by 5.3 SR points
- RxR-CE: 75.1% SR, 68.7% SPL; RL adds +3.9 SR over SFT
Shows that image-space waypoint prediction plus efficient sim-first training can beat privileged-sensor navigation systems.