Research map / Vision-language and multimodal / Multimodal LLMs

Is your VLM Sky-Ready? A Comprehensive Spatial Intelligence Benchmark for UAV Navigation

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: VLM capability analysis · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces SpatialSky-Bench, a 13-subtask spatial intelligence benchmark for UAV-view VLMs, a 1M-sample auto-generated dataset built from UAVScenes annotations, and Sky-VLM (Qwen2.5-VL-7B fine-tuned with SFT then GRPO), which reports large gains over baselines.

Problem

Existing spatial intelligence benchmarks and VLMs focus on ground or egocentric views and do not assess spatial perception from UAV perspectives, which matters for UAV navigation.

Approach

Generate a UAV-view spatial QA benchmark and training set automatically from mask, LiDAR, and pose annotations, then train a 7B VLM with SFT followed by GRPO using task-specific rewards (point L1, choice exact match, box IoU).

Key results

• Sky-VLM (7B) reports total average 53.30 on SpatialSky-Bench versus 23.07 for the best baseline (GPT-5).
• SFT-only Sky-VLM-SFT scores 48.29 total; RL stage raises environment perception from 52.53 to 60.33.
• Qwen2.5-VL-7B baseline scores 16.93 total average on the benchmark.

Surprising findings

• Performance reportedly decreases as SFT sample size grows beyond 300K (30.43 at 300K vs 48.29 at 1M), which conflicts with the paper's own claim of monotonic scaling.
• Closed-source models such as GPT-5 and Gemini-2.5-Pro score only around 21-23 average on the UAV benchmark.

Limitations

• Benchmark and training data are automatically generated with VLM-written captions and templates, possibly introducing noise.
• Open-ended scoring relies on GPT-4o as judge with BLEU, with limited validation shown.

Experimental setup

Datasets: UAVScenes

Benchmarks: SpatialSky-Bench

Metrics: mIoU point accuracy accuracy BLEU GPT-4o score 1-10

Baselines: GPT-4o GPT-5 Gemini-2.5-Pro Gemini-2.5-Flash Qwen-VL-Max InternVL3.5 Qwen3-VL-8B Qwen2.5-VL-72B SpatialVLM SpaceR VILASR

Builds on: Qwen2.5-VL GRPO UAVScenes

Closest papers by meaning

• Towards Realistic UAV Vision-Language Navigation: Platform, Benchmark, and Methodology (ICLR 2025 · similarity 0.91)
• SatNav: A Scalable Benchmark for Long-Horizon UAV Vision-Language Navigation from Satellite Imagery (NeurIPS 2026 · similarity 0.89)
• SpaceDrive: Infusing Spatial Awareness into VLM-based Autonomous Driving (CVPR 2026 · similarity 0.87)
• GTR-Bench: Evaluating Geo-Temporal Reasoning in Vision-Language Models (ICLR 2026 · similarity 0.87)
• Parse, Search, and Confirmation: Training-Free Aerial Vision-and-Dialog Navigation with Chain-of-Thought Reasoning and Structured Spatial Memory (CVPR 2026 · similarity 0.87)
• Spatial Reasoning with Vision-Language Models in Ego-Centric Multi-View Scenes (ICLR 2026 · similarity 0.87)