Event
Goal2Pixel: Reframing VLN-CE as Pixel Grounding
1 reports · 1 sources
Muyi Bao and other authors propose Goal2Pixel, which reframes Vision-and-Language Navigation in Continuous Environments (VLN-CE) as "navigable pixel grounding": the model predicts a visible navigable pixel on the image plane, then back-projects it into a 3D waypoint for moving forward; three auxiliary regions—left, right, and down—respectively encode turning and stop actions. The authors claim that, among methods trained without external data, Goal2Pixel achieves competitive state-of-the-art levels on R2R-CE and RxR-CE; on R2R-CE Val-Unseen it achieves 54.1% success rate (SR) and 52.5% SPL, calling the vision-language model only 7.75 times per episode. These are results reported in the paper.
Report timeline
- Goal2Pixel: A Vision-Language Navigation Method That Grounds Goals to Pixels
Goal2Pixel proposes a pure pixel paradigm that reformulates vision-and-language navigation in continuous environments (VLN-CE) as navigable pixel grounding: the model predicts navigable pixels on the image plane, which are then back-projected into 3D waypoints for forward movement, while turning and stopping are encoded by auxiliary left, right, and bottom regions.
arXiv cs.RO ·