SpatialUAV Benchmark for Low-Altitude UAV Intelligence

• SpatialUAV is a benchmark that evaluates UAV spatial intelligence by testing multi-view 3D inference, cross-perspective reasoning, and dynamic motion understanding.
• It comprises 4,331 curated instances over 14 fine-grained task types, ranging from semantic discrimination to complex geometric and motion reasoning.
• Empirical findings reveal that current models fall short of human-level performance, highlighting weaknesses in cross-view association, structured grounding, and temporal viewpoint understanding.

Searching arXiv for papers directly related to “SpatialUAV” and nearby UAV spatial intelligence work. SpatialUAV is a real low-altitude UAV benchmark for evaluating spatial intelligence in aerial perception, collaboration, and motion. It is designed to test whether vision-LLMs can perform 3D spatial inference, multi-view collaboration, scene-dynamic reasoning, and heterogeneous answer generation, rather than only image-level recognition or single-view understanding. The benchmark comprises 4,331 curated instances across 14 fine-grained task types, organized under a unified visual-input–question–answer schema and evaluated with task-specific metrics for labels, correspondences, geometric values, bounding boxes, and free-form motion descriptions. Its central empirical finding is that current models remain far from human-level performance, with especially pronounced weaknesses in cross-view association, structured grounding, geometric reasoning, and temporal viewpoint understanding (Zhang et al., 26 Jun 2026).

1. Definition and research scope

SpatialUAV was introduced to address a specific gap in UAV evaluation: existing UAV benchmarks often emphasize image-level recognition, single-view understanding, or narrow answer formats, leaving 3D spatial inference, multi-view collaboration, scene dynamics, and diverse task formulations insufficiently evaluated. This gap is especially consequential for low-altitude UAVs, where perspective distortion, altitude-dependent scale changes, partial occlusion, aerial–ground viewpoint mismatch, and cross-view reasoning are routine operating conditions rather than corner cases (Zhang et al., 26 Jun 2026).

The benchmark is organized around five capability areas: semantic discrimination, spatial relation reasoning, aerial–aerial collaboration, aerial–ground collaboration, and motion understanding. This design makes SpatialUAV diagnostic rather than purely aggregate: it does not ask only whether a model can recognize scene content, but whether it can localize regions, compare distances, associate objects across views, infer camera transformation, recover occluded targets, translate positions between aerial and ground perspectives, and describe UAV motion over time (Zhang et al., 26 Jun 2026).

A plausible implication is that SpatialUAV formalizes “spatial intelligence” for UAV-centered multimodal systems as a composite of geometric inference, cross-view grounding, and temporally coherent motion understanding, rather than as a synonym for aerial visual recognition alone.

2. Dataset composition and task structure

SpatialUAV contains 4,331 curated instances and 14 fine-grained task types. These are distributed across four reasoning settings: 1,315 single-image samples, 1,231 aerial–aerial samples, 785 aerial–ground samples, and 1,000 video-motion samples. Each instance belongs to one of the five capability areas, but the benchmark preserves a single canonical representation for all tasks (Zhang et al., 26 Jun 2026).

The underlying formalism represents each example as

$(\mathcal{V}, \tau, q, y^*)$

where $\mathcal{V} = (I_1,\ldots,I_K)$ is the ordered visual input, $\tau \in \mathcal{T}$ is one of the 14 task types, $q$ is the question or instruction, and $y^* \in \mathcal{Y}_\tau$ is the canonical ground-truth answer. The visual input cardinality satisfies

$K \in \{1,2,5,16\},$

covering single images, paired views, candidate-view selection, and video frames (Zhang et al., 26 Jun 2026).

Capability area	Task types	Typical output forms
Semantic discrimination	Region Recognition; Anomaly Detection	Region label; option letter
Spatial relation	Direction Recognition; Distance Comparison	Direction label; clock direction; option letter
Aerial–aerial collaboration	Collaboration Recognition; Shared Association; Object Matching; Camera Transformation; Occlusion Removal	Option letter; region pair list; bounding box; angle value; angle-distance pair
Aerial–ground collaboration	Shared Association; Collaboration Recognition; View Translation; Path Planning	Region pair list; option letter; region label; direction label
Motion understanding	Global Motion	Free-form text

The 14 tasks include Region Recognition, Anomaly Detection, Direction Recognition, Distance Comparison, Collaboration Recognition, Shared Association, Object Matching, Camera Transformation, Occlusion Removal, View Translation, Path Planning, and Global Motion. Some task names recur across aerial–aerial and aerial–ground settings because the reasoning primitive is similar while the input geometry differs (Zhang et al., 26 Jun 2026).

This structure is significant because it prevents a benchmark from being dominated by one answer type or one visual regime. SpatialUAV explicitly mixes region selection, geometric estimation, correspondence recovery, and natural-language motion description within one evaluation framework.

3. Input configurations and answer formats

SpatialUAV supports seven input configurations and nine answer formats. The input configurations include single clean images, single detected or annotated images, detected images with hidden category labels and neutral region IDs, annotated pair images, clean paired images, annotated UAV-plus-ground path inputs, and ordered video frame sequences. Concrete mappings are task-dependent: Region Recognition uses detected images with region IDs, Collaboration Recognition uses clean paired images, Occlusion Removal uses annotated pair images, and Global Motion uses ordered video frames (Zhang et al., 26 Jun 2026).

The benchmark’s nine canonical answer formats are: option letter, region label, direction label, clock direction, region pair list, bounding box, angle value, angle-distance pair, and free-form text. The benchmark analysis identifies Angle-Distance Pair, Bounding Box, and Region Pair List as especially difficult structured-output regimes (Zhang et al., 26 Jun 2026).

This heterogeneity is not incidental. The benchmark was designed to avoid the common reduction of UAV reasoning to multiple-choice classification. Its output space includes discrete labels, structured correspondences, geometric quantities, and open-ended motion descriptions. The paper’s ablation results show that converting structured-output tasks into multiple-choice questions yields only small and inconsistent gains, which indicates that the underlying difficulty is not primarily answer formatting but the spatial reasoning itself (Zhang et al., 26 Jun 2026).

A related misconception is that poor performance arises mainly from low image resolution or small aerial targets. SpatialUAV explicitly tests higher-resolution inputs and reports only marginal and inconsistent improvement, suggesting that the dominant failure mode is weak spatial prior generalization to UAV geometry rather than simple visibility loss (Zhang et al., 26 Jun 2026).

4. Data construction and validation pipeline

The benchmark’s construction pipeline integrates detector-assisted regions, depth supervision, metadata-derived rules, extensive manual annotation, blind filtering, and multi-turn human validation. Data are curated from five resources: BEDI, AirCopBench, MAVREC, AirScape, and University-1652. The pipeline preserves source labels, paired-view correspondences, scene identifiers, temporal ordering, and camera or trajectory annotations when available (Zhang et al., 26 Jun 2026).

For region-level tasks, images are annotated with bounding boxes and neutral region IDs. The original category names are hidden from the model to reduce shortcut learning. For distance-related tasks, the benchmark uses Metric3D depth estimates to derive depth-aware answers, particularly for Distance Comparison. Metadata-derived rules support tasks such as Region Recognition, Collaboration Recognition, Object Matching, and Camera Transformation. Tasks judged unsuitable for reliable auto-derivation, including anomaly localization, shared cross-view associations, occlusion recovery, view translation, path-planning directions, and free-form motion descriptions, are manually labeled (Zhang et al., 26 Jun 2026).

A distinctive element of the pipeline is blind filtering for language-only solvability. DeepSeek-V4-Pro and Qwen3.6-27B are run in a setting where they receive only the text prompt; if either model answers correctly without vision, the sample is removed. This procedure is intended to suppress wording leakage, answer-pattern artifacts, and language priors that would otherwise make a task non-visual (Zhang et al., 26 Jun 2026).

Validation occurs in two rounds. First, exhaustive human cross-validation inspects the visual input, question, answer format, and label, with disagreements resolved manually. Second, targeted model-assisted validation runs Qwen3-VL-30B, Qwen3.6-35B, and InternVL3.5-38B on all samples; if all three disagree with the ground truth, the instance is reopened for human review (Zhang et al., 26 Jun 2026).

This pipeline is methodologically important because SpatialUAV does not present itself as a raw aggregation of pre-existing data. It is a synthesized and filtered benchmark intended to enforce grounded visual reasoning rather than text exploitation or annotation noise tolerance.

5. Evaluation protocol and scoring design

Because SpatialUAV mixes heterogeneous outputs, it uses task-specific metrics rather than a single global exact-match criterion. Each task reports a normalized per-sample score $s_i \in [0,1]$, aggregated as

$100 \times N^{-1}\sum_i s_i$

for $N$ test instances of that task (Zhang et al., 26 Jun 2026).

For option letters, clock directions, path-planning labels, and single-region outputs, the benchmark uses exact matching: $s_i = \mathbf{1}[\hat{y}=y].$ For multi-region outputs such as Region Recognition and Anomaly Detection, it uses conservative partial credit: $\mathcal{V} = (I_1,\ldots,I_K)$0 This penalizes over-prediction strongly (Zhang et al., 26 Jun 2026).

For Shared Association, the score is pair-level $\mathcal{V} = (I_1,\ldots,I_K)$1: $\mathcal{V} = (I_1,\ldots,I_K)$2 where $\mathcal{V} = (I_1,\ldots,I_K)$3 is the predicted correspondence set, $\mathcal{V} = (I_1,\ldots,I_K)$4 is the ground-truth set, and $\mathcal{V} = (I_1,\ldots,I_K)$5 is the number of matched pairs. Bounding-box evaluation combines overlap, center consistency, and size consistency: $\mathcal{V} = (I_1,\ldots,I_K)$6 with $\mathcal{V} = (I_1,\ldots,I_K)$7 (Zhang et al., 26 Jun 2026).

For geometric outputs in Camera Transformation, the benchmark computes

$\mathcal{V} = (I_1,\ldots,I_K)$8

with thresholds $\mathcal{V} = (I_1,\ldots,I_K)$9 and $\tau \in \mathcal{T}$0 meters. The score is exact-threshold based for angle-only or angle-distance outputs. For Global Motion, free-form descriptions are scored by an LLM-based semantic similarity function

$\tau \in \mathcal{T}$1

implemented with a GPT-5.4-mini judge (Zhang et al., 26 Jun 2026).

The evaluation design is notable for two reasons. First, it treats structured grounding and free-form temporal description as first-class outputs rather than collapsing them into proxy labels. Second, it encodes task semantics into the metric itself, which makes the benchmark more diagnostic but also more demanding.

6. Empirical findings and position within UAV spatial intelligence research

SpatialUAV evaluates 18 representative vision-LLMs spanning closed-source, spatial-specific, and open-source categories. Human performance on a 20% sampled subset is 89.0 average. The best reported model, GPT-5.4, reaches 56.7 average, while the best open-source model, Qwen3.6-27B, reaches 49.5 average. Models perform relatively better on semantic discrimination and simpler spatial relation tasks, but exhibit severe degradation on cross-view association, structured grounding, geometric reasoning, and temporal viewpoint understanding (Zhang et al., 26 Jun 2026).

Human scores remain high across tasks, including Region Recognition at 99.3, Anomaly Detection at 100, Direction Recognition at 95.4, Distance Comparison at 97.6, Collaboration Recognition at 99.5, and Motion at 89.1. By contrast, model failures are concentrated in aerial–aerial Object Matching, Camera Transformation, and aerial–ground Shared Association. The benchmark therefore argues that the dominant bottlenecks are genuine spatial reasoning failures rather than merely hard output formats (Zhang et al., 26 Jun 2026).

This diagnostic role places SpatialUAV in a broader research landscape rather than in isolation. Its task areas align closely with several active UAV research threads. GPS-denied self-positioning and cross-view geo-localization are represented by CEUSP, which targets dense urban UAV self-positioning via context-enhanced cross-view retrieval (Xu et al., 17 Feb 2025), by G2CL, which incorporates geographic neighborhoods into precise self-positioning losses (Li et al., 22 Feb 2025), and by SWA-PF, which combines semantic segmentation with a 4-DoF adaptive particle filter for GNSS-denied localization (Yuan et al., 17 Sep 2025). Geometry-aware vision-language navigation appears in SpatialFly, which aligns 2D semantic tokens with implicit 3D geometric priors for UAV VLN in urban environments (Jiang et al., 22 Mar 2026). UAV-satellite collaborative reasoning is developed further by SatAgent, which lifts UAV features into BEV space and aligns them with satellite features for multi-view spatial reasoning (Zhang et al., 30 Jun 2026).

The benchmark’s collaboration and motion categories also intersect with system-level UAV research. Decentralized cooperative search in cluttered GPS-denied environments is addressed through an observation-oriented decentralized POMDP framework for multi-UAV exploration and target finding (Zhu et al., 2021). Informative path planning in continuous 3-D space with multiresolution mapping and altitude-dependent sensing is treated in UAV terrain monitoring (Popovic et al., 2017). Risk-aware urban route planning based on weighted A* and dynamic local avoidance appears in low-altitude autonomous navigation (Castelli et al., 2016). Language-grounded mission generation over satellite imagery is exemplified by UAV-VLA and its hybrid successor UAV-VLPA*, which combine VLM or GPT-style semantic parsing with TSP-based global ordering and A* local refinement for large-scale aerial routing (Sautenkov et al., 9 Jan 2025, Sautenkov et al., 4 Mar 2025).

Taken together, these connections suggest that SpatialUAV is not a benchmark for one narrow subproblem, but a unifying evaluation instrument for perception, collaboration, localization, reasoning, and motion under low-altitude aerial viewpoints. Its importance lies less in proposing a single algorithm than in making visible the gap between current multimodal models and the full spatial competence required by real UAV systems (Zhang et al., 26 Jun 2026).