MulRobBench: leaderboard
MulRobBench (2026) tests whether multimodal UAV agents make protocol-compliant, safe decisions from real aerial observations under glare, missing data and ambiguous operator language. Semantic protocol-decision score in percent, aggregated over constraint extraction, target disambiguation, safe action, reobservation and collaboration decisions; higher is better.
Metric: Semantic protocol-decision score (%; aggregate over constraint extraction, target disambiguation, safe action, abstention or reobservation and collaboration trigger; 3,024-sample strict evaluation set of real UAV observations with protocol constraints, degradations and operator language; controlled semantic scores in [0, 1] x100). Source: arxiv.org. Saturation forecast: Around 2030. 17 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3.5 9B | 49.78 | #803 |
| 2 | Qwen 3.5 4B | 48.88 | #1175 |
| 3 | Gemma 4 E4B | 46.29 | #2214 |
| 4 | Qwen 2.5 VL 7B | 44.99 | #1787 |
| 5 | Gemma 4 E2B | 43.49 | #3892 |
| 6 | Qwen 3.5 2B | 43.27 | #4810 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=mulrobbench · How It Works · Data refreshed daily, snapshot 2026-11-13.