AirNav: leaderboard
AirNav (Cai et al., 2026) is a UAV vision-and-language navigation benchmark on real urban aerial data: a model reads a natural-language route instruction written in one of ten user personas and flies a simulated drone with discrete moves from the start point to a described target. Success rate in percent on the test-unseen split, the share of episodes that stop within the success radius of the target; higher is better. General multimodal models run zero-shot with the instruction, drone state, action history and a sample of past views.
Metric: Success rate (%; test-unseen split, share of episodes that stop within the success radius of the target; zero-shot MLLMs given the instruction, UAV state, action history and current plus sampled past views). Source: arxiv.org. Saturation forecast: Around October 2028. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3 VL 235B A22B | 5.21 | #477 |
| 2 | GPT-4o | 4.56 | #741 |
| 3 | Qwen 2.5 VL 32B Instruct | 2.99 | #1055 |
| 4 | Qwen 2.5 VL 7B Instruct | 1.7 | #1842 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=airnav · How It Works · Data refreshed daily, snapshot 2026-11-12.