PROJECT DOCUMENTATION
Project & data notes
Interactive maps and navigation
The vector viewer reuses the existing map renderer and routing/navigation core with isolated browser sessions. The scene guide contains 74 public places; its accepted walking graph remains empty. Navigation is demonstrated on the separate synthetic network using simulated position updates, with no agent movement or task scoring. The city atlas is an unchanged 12,288 × 18,432 Blender render displayed through 1,167 local pyramid tiles. It is high-resolution raster imagery, not a vector map. These source versions are presented separately and are not asserted to be geometrically registered.
Project positioning
CityAgentBench is a working title for a research project on embodied-agent evaluation in reconstructed urban environments. This page does not claim acceptance at a conference, a public benchmark release, or completed six-level results.
Demonstration provenance
The 65-second city tour is a native Blender architectural rendering. The 16.375-second UE clip is a real API-driven short pilot recorded on 2026-09-23, with waiting time preserved. The 12-second third-person clip is a re-render of logged trajectory data with interpolated poses and adjusted playback timing. Neither legacy pilot is a result under the current six-level protocol or its 1.80 m observation setup.
Chart methodology
The coverage plot is recomputed from coverage_grid_100m.tsv and ground_grid_100m.tsv: 3,308 aerial-footprint cells, 595 airborne-LiDAR cells, 727 vehicle raw-LiDAR cells, and 359 Dashahe backpack cells. Each cell covers 100 × 100 m. These counts describe the selected historical index, not the complete dataset or a new full-file audit. The older textual summary reports 598 ALS cells; the chart follows the 595 positive ALS cells in the actual TSV and does not silently equate the two.
Acquisition composition
The aerial chart counts 83,766 valid records from photos.tsv across nine sorties. Nadir views use pitch < −75°, matching the index generator; camera identifiers alone do not define view type. The source indices date to 2026-09-08.
Alignment and image provenance
Aerial footprints support coverage analysis, not pixel-level alignment. The Starlight Bridge panorama is frame 1130d/998; the backpack sample is 1204a/928. Their per-frame absolute poses remain unverified. Scene renders come from the 2026-09-24 reconstruction report. The six-level figure is a pure-vector taxonomy of levels, reasoning demands, and all 24 coverage categories. It contains no scene imagery or measured scores. A second, project-guided conceptual illustration is retained in the figure tabs. Its locations and paths are schematic, not measured task evidence.
Six levels and 24 categories
The framework follows the existing evaluation plan: direct perception, active evidence gathering, spatial reasoning, long-term memory and reuse, constrained planning and revision, and integrated autonomous exploration. L1–L3 are prioritized; L4–L6 remain research designs. The categories guide coverage, not fixed quotas. No new questions were authored for this website.
Availability
This research preview presents selected demonstrations and visualization assets. Source datasets, models, and private evaluation truth are not included. Author names, affiliations, manuscript links, and citation metadata will be added when supplied and verified.