← News

ResearchSources: arXiv · 19 Nov 2026

CoNav-UAV trains two drones to navigate together as leader and follower

Researchers propose modelling dual-drone aerial navigation as a Stackelberg game, letting a high-altitude leader and low-altitude follower adapt to each other during training.

In plain words

The paper describes a way to train two drones that work together to find a target in a city from a short text description. One drone watches from high above while the other flies close to the target. The training method makes each drone adjust to the other as both improve, and the authors report better success rates than earlier single- and two-drone approaches in simulated urban scenes.

What was announced

A team behind the CoNav-UAV paper frames target-oriented aerial vision-and-language navigation — finding a target from a brief description of its appearance and surroundings — as two interleaved problems: city-scale exploration and collision-free close-range approach. They argue a single drone struggles to reconcile both, and that an earlier two-drone attempt relied on privileged sensing and trained its agents independently, preventing mutual adaptation.

CoNav-UAV instead treats the task as a Stackelberg game between a high-altitude leader and a low-altitude follower, reflecting their sequential decisions and asymmetric observations, with both agents using only onboard sensing. The proposed Iterative Stackelberg Learning refines the leader's vision-language reasoning through memory-based in-context learning while fine-tuning the follower's motion control with DAgger-style expert distillation; alternating updates let each agent adapt to the other. On three high-fidelity urban scenes from the AerialVLN benchmark, the authors report success-rate gains of up to 30.8 points on the learning scene and up to 9.0 points in cross-scene transfer, where the follower needed 2.6-3.7 times less target-scene data. These are the authors' own benchmark results and have not been independently verified.

What it does not show

  • Evidence is limited to the paper's abstract; the full article was not reviewed, so method details and evaluation setup could not be checked.
  • All reported performance figures come from the authors' own experiments in simulated AerialVLN scenes; no independent confirmation or real-world deployment results are available.

Sources

Source-support review completed 20 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 20 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.