aiqbench analysis
This research addresses the computational constraints of edge deployment for aerial robotics by maintaining a modular architecture that separates vision-language grounding, planning, and control. By utilizing quantized vision-language models and reinforcement learning policies, the system achieves high success rates in cluttered indoor environments while retaining safety checks lost in end-to-end approaches. This work advances practical onboard autonomy for drones without sacrificing the interpretability required for safe flight operations.
VLN on the Fly: An Onboard Vision-Language Navigation Stack for Aerial Robots
arXiv:2609.20191v1 Announce Type: cross Abstract: Running vision-language navigation fully onboard an aerial robot is hard, since grounding, planning, and control must share limited compute and a single-stage error is difficult to isolate in flight. End-to-end aerial policies fuse these stages into one network, giving up the observability and safety checks a modular stack keeps available. We propose VLN on the Fly, an onboard stack that keeps grounding, planning, and control as separate, inspectable stages. A quantized VLM grounds an instruction to a coarse image cell, depth lifts it to a 3D goal, a fast B-spline planner returns a feasible trajectory, and a pretrained reinforcement learning policy tracks it to motor commands across quadrotors. Across 15 onboard flights over three everyday referents in a controlled indoor volume, the stack reaches the target in 13 of 15 trials with 5.72 cm mean goal error and 39.3% average GPU utilization. In 6 additional cluttered-environment trials, the stack tracks collision-free trajectories under onboard perception gating.
Abstract:Running vision-language navigation fully onboard an aerial robot is hard, since grounding, planning, and control must share limited compute and a single-stage error is difficult to isolate in flight. End-to-end aerial policies fuse these stages into one network, giving up the observability and safety checks a modular stack keeps available. We propose VLN on the Fly, an onboard stack that keeps grounding, planning, and control as separate, inspectable stages. A quantized VLM grounds an instruction to a coarse image cell, depth lifts it to a 3D goal, a fast B-spline planner returns a feasible trajectory, and a pretrained reinforcement learning policy tracks it to motor commands across quadrotors. Across 15 onboard flights over three everyday referents in a controlled indoor volume, the stack reaches the target in 13 of 15 trials with 5.72 cm mean goal error and 39.3% average GPU utilization. In 6 additional cluttered-environment trials, the stack tracks collision-free trajectories under onboard perception gating.
Robotics (cs.RO); Artificial Intelligence (cs.AI); Computer Vision and Pattern Recognition (cs.CV); Machine Learning (cs.LG); Systems and Control (eess.SY)
arXiv:2609.20191 [cs.RO]
(or arXiv:2609.20191v1 [cs.RO] for this version)
https://doi.org/10.48550/arXiv.2609.20191 arXiv-issued DOI via DataCite (pending registration)
From: Ricardo Vilela De Godoy Dr. [view email] [v1] Fri, 7 Aug 2026 20:34:33 UTC (7,770 KB)
Full text under the publisher's open licence · 206 words
Related coverage
LightVLN: Efficient Aerial Vision-and-Language Navigation with Compact Memory and History-Guided Local Aggregation
arXiv · Oct 6, 2026
TADreamer: Zero-Shot Language-Guided 3D Navigation for Terrestrial-Aerial Bimodal Robots via Video Imagination
arXiv · Sep 18, 2026
Dyna-2: General Purpose Physical Agents with Hour-Long Autonomy — Jason Ma, Dyna Robotics
Latent Space · Oct 13, 2026
VoLN: Vision-Only Long-Horizon Navigation---Paradigm, Benchmark, and Method
arXiv · Jul 24, 2026