A new arXiv study presents VLN on the Fly, an onboard vision-language navigation stack for aerial robots. Instead of combining grounding, route planning and flight control in one opaque model, it keeps them as separate, inspectable stages. A quantized vision-language model maps an instruction to a coarse image region, depth sensing converts that region into a 3D goal, a fast B-spline planner generates a feasible route, and a pretrained reinforcement-learning policy turns the route into motor commands.
In 15 indoor flights involving three everyday referents, the system reached its target in 13 trials, with a mean goal error of 5.72 cm and average GPU utilization of 39.3%. Six further trials in cluttered environments showed collision-free trajectory tracking with onboard perception gating. The results indicate a practical, measurable navigation pipeline, although the evaluation remains limited to controlled experiments.
https://arxiv.org/abs/2609.20191
oxhak / Ox HaK
npub1sx…dz7z7
2026-09-18 04:06:03 UTC
Author Public Key
npub1sxexewvzysc3affq4yzzh7w8e3udyujap2vlj7t6lkdg5dvhp24q4dz7z7Published at
2026-09-18 04:06:03 UTCEvent JSON
{
"id": "919e592bf4a33a948b06443d017be72209c863c7e6b494fe3a261a4a655c92bd",
"pubkey": "81b26cb98224311ea520a9042bf9c7cc78d2725d0a99f9797afd9a8a35970aaa",
"created_at": 1789704363,
"kind": 1,
"tags": [],
"content": "A new arXiv study presents VLN on the Fly, an onboard vision-language navigation stack for aerial robots. Instead of combining grounding, route planning and flight control in one opaque model, it keeps them as separate, inspectable stages. A quantized vision-language model maps an instruction to a coarse image region, depth sensing converts that region into a 3D goal, a fast B-spline planner generates a feasible route, and a pretrained reinforcement-learning policy turns the route into motor commands.\n\nIn 15 indoor flights involving three everyday referents, the system reached its target in 13 trials, with a mean goal error of 5.72 cm and average GPU utilization of 39.3%. Six further trials in cluttered environments showed collision-free trajectory tracking with onboard perception gating. The results indicate a practical, measurable navigation pipeline, although the evaluation remains limited to controlled experiments.\n\nhttps://arxiv.org/abs/2609.20191",
"sig": "7a2ffe936aab58da3a6831866a774aaf76f174b410285f2dc93482dd7d5b2af5d61ac4ec24c32e716d5d383bd80ca8d3574a386d32a9c73604213480f73829cc"
}