UAV-DualCog

Task Quantity Distribution

The release balances all six task families at 1,024 samples each, covering both cognition axes in image and video settings.

This yields 3072 Agent-State and 3072 World-State questions. The four image tasks split evenly between the two states, and the two video tasks preserve the same balance for temporal reasoning.

Task Quantity Distribution. Each released task contributes 1,024 samples, yielding balanced task counts across the two cognition axes and the six task families.

Scene Coverage

Released scenes are first constructed as fused Stage 1 environments with aligned geometry, poses, and scene-level observations, and only then filtered into the benchmark-facing split. The current benchmark selects 12 test scenes from this reviewed pool.

The scene library below shows two benchmark-facing views per released test scene whenever public thumbnails are available, while the chart summarizes how landmarks, image samples, and video samples remain distributed across scenes. ENV 16 contributes the most landmarks (126).

Larger contributors such as ENV 16, ENV 7, and ENV 20 are complemented by smaller scenes including ENV 8, ENV 13, and ENV 17. Together they cover dense urban layouts, industrial zones, waterfront areas, and compact environments.

Previous 1 / 12 Auto-play: On Next

ENV 7 · Waterfront City

Released benchmark scene

Landmarks 86

Image Samples 688

Video Samples 344

Area 220580 m²

Scene Library. Two representative benchmark-facing snapshots are shown for each released test scene with available public thumbnails.

The following chart compares each scene's contribution to the landmark inventory and the image and video task pools.

Landmark Category

Released landmarks come from the Stage 2 review pipeline, where raw scene candidates are filtered, assigned coarse and fine-grained semantic labels, and linked to multiview RGB evidence before split selection. The released benchmark keeps 512 reviewed landmarks across eight coarse categories and 127 fine-grained subcategories; the broader reviewed asset pool spans 166 subcategories for future expansion.

The tables below summarize the dominant categories in the released split, and the landmark library grounds those labels back into the same multiview evidence used by the image-task branch. Building is currently the largest coarse category (34.4%), while Mid Rise Building is the single most common fine-grained subtype (15.6%).

The category distribution is broad but not flat. Buildings remain the largest coarse class, and vegetation is also strongly represented through frequently occurring tree landmarks, while public facilities and industrial infrastructure together account for a substantial fraction of the release. At the fine-grained level, the long tail is preserved: mid-rise buildings and deciduous trees dominate, but street furniture, signs, cranes, benches, and other urban objects remain visible as benchmark targets.

Landmark Category Distribution. The released split preserves all eight coarse categories while remaining long-tailed at the fine-grained subcategory level.

The landmark library below keeps the full eight side-view orbit for representative released landmarks, so the category taxonomy is grounded in the same multiview evidence used by the image-task pipeline.

Landmark Mid Rise Building · ENV 7High Rise Building · ENV 16Prefabricated Container Building · ENV 9High Rise Building · ENV 11Wooden Scaffolding Structure · ENV 20Mid Rise Building · ENV 10Mid Rise Residential Building · ENV 7Mid Rise Building · ENV 7Small Shed · ENV 9Sedan Car · ENV 23

Previous Auto-play: On Next

Mid Rise Building

Building · Mid Rise Building · ENV 7

light gray mid-rise building with red tiled roof and dormer windows

Occluded / Invalid

Occluded / Invalid

Landmark Library. Each showcased landmark preserves the full eight landmark-centric side views used by the released image-task branch.

Image Tasks

Image-task construction begins with reviewed Stage 2 landmarks, preserves their valid landmark-centric orbit views, and then adds task-specific egocentric captures such as the query observations used by world-state tasks. The image branch is therefore built from released scene assets and released landmark views rather than from detached benchmark-only screenshots.

In the current release, the image branch contains 4096 samples backed by 4840 source images, including 3816 valid landmark views and 1024 extra task-specific observation captures. Landmark descriptions are kept concise while remaining discriminative, averaging 10.6 words across the released split. Detailed rendering and prompt-constrained generation are described in the construction page.

The two histograms show that image-task supervision remains visually and linguistically well-conditioned. Most released landmarks retain rich multiview support, with the largest mass concentrated at eight valid side views, and the remaining landmarks still clustering around five to seven views. Description lengths are similarly concentrated: the released split peaks around 9 to 11 words, so descriptions stay compact while still carrying enough semantic detail to distinguish nearby landmarks.

Image-Branch Summary. The released image branch comprises 4096 samples, 3816 valid landmark views, and 1024 task-specific egocentric observations.

The histograms below make two properties explicit: most released landmarks preserve a high number of valid viewpoints, and landmark descriptions are intentionally concise rather than paragraph-length annotations.

Video Tasks

Video-task construction starts from reviewed landmarks and a predefined Stage 3 behavior library. In this library, atomic behaviors are the reusable low-level flight primitives, such as approach, orbit, rise, or mapping motions, each instantiated with controlled default parameter ranges. Composite behaviors are higher-level inspection patterns built by composing multiple atomic maneuvers into longer and more operationally meaningful UAV routines.

The released video branch therefore does not treat flight videos as unconstrained motion clips. Instead, it anchors both Flight Behavior Recognition and Temporal Localization and Landmark Visibility Counting and Interval Reasoning in the same hierarchical behavior definition, so that benchmark predictions can be interpreted against explicit flight-mode semantics rather than only against raw video appearance.

Behavior Hierarchy. The released video branch is grounded in a two-layer behavior library where reusable atomic maneuvers are composed into higher-level inspection classes.

The released video branch contains 2048 samples with 5h 27m 13s of total released footage, built from 5 composite templates and 10 atomic behavior classes. The distributions below summarize how those missions are instantiated and populated in the benchmark, while the detailed behavior library and mission-generation chain are documented on the Construction page.

The statistics below expose three complementary properties of the video branch. First, the behavior-mode counts are broadly even across the five composite inspection families, while the atomic library remains distributed across ten lower-level maneuvers rather than being dominated by only one or two motion primitives. Second, composite missions are physically longer than atomic ones, with most composite trajectories concentrated between 200 and 600 meters, whereas atomic trajectories are concentrated below 400 meters and especially in the 0 to 200 meter range. Third, the video-duration distribution mirrors that pattern: most atomic missions fall below 20 seconds, while most composite missions cluster between 10 and 30 seconds, with only a small long tail extending beyond 40 seconds.

The following distributions summarize how the released video branch is populated: the first chart counts released missions per behavior mode, and the latter two compare atomic and composite trajectories in terms of physical path length and video duration. In both latter views, atomic missions dominate the shortest bins, while composite missions shift toward the mid-range bins that correspond to longer inspection routes and longer observation windows. The underlying behavior hierarchy and default mission templates now live in the Stage 3 construction section so that the benchmark page can stay focused on released statistics and task-facing interpretation.