Answers are grounded in this paper and cite their sources.
Vision and language navigation (VLN) for unmanned aerial vehicles (UAVs) requires an agent to follow natural language instructions while perceiving scenes whose appearance changes with flight altitude and heading. Existing methods treat these variations as nuisances to be learned, spending model capacity on scale and orientation invariance that the imaging geometry already determines. We propose SOAR, a telemetry-conditioned framework that uses onboard altitude, heading, and camera intrinsics to normalize the viewpoint at the input. A parameter free metric tokenizer rescales and rotates each frame so that every token covers a fixed ground footprint in a north-up frame. Under near-nadir, locally planar conditions, this reduces the perspective to bird’s-eye-view transformation to a reshape, a single convolution, and a tilt-correcting homography. Map registration then becomes pure translation, and language grounding and waypoint prediction operate directly in metric units. SOAR trains on a single 24 GB GPU. On the validation–unseen splits of AerialVLN-S and CityNav, SOAR reached success rates of 8.0% and 8.5%, with navigation errors of 120.0 m and 121.0 m, respectively. We report these results next to benchmark values from the literature as reference points for context. Because the experimental configurations differed, these comparisons do not support claims of controlled superiority over previously published methods.
No takes yet. Share an insight, caveat, or question.
Khujamatov et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: