Research notes
Mirai Guide: reading walkable floors from the map raster
Why booth-gap inference failed, and how color segmentation of the venue map yields a navigation grid for A*.
Point-to-point walking navigation inside a venue — booth to booth, entrance to anywhere — needs to know where the floor is. The first model inferred walkability from the negative space between booth rectangles. Measurement showed that model is wrong in ways navigation would expose, and the replacement reads the floor directly from the map image.
The evidence against booth-gap inference
Measured across three production venue rasters:
- Rendering seams read as aisles. Nearest-neighbor gaps between booths form two distinct populations — 1–3px seams and ~31px real corridors — and 30–42% of booth-to-booth paths squeezed through gaps narrower than 8px on every map. Acceptable for a decorative cart line, wrong for directions.
- No hall boundary. Many landmarks — entrances, exits, information desks — sit outside the hall outline. Of measured routes crossing one venue’s bottom wall, about 41% passed through solid wall rather than a door. The entrance is the most natural start point for navigation, so this is fatal.
- Missing obstacles. One venue has structural pillars inside its zones that exist in no anchor table.
- Non-rectilinear geometry. Another venue has a curved wall and a two-level walkable space joined by door gaps. An axis-aligned-rectangle obstacle model structurally cannot express it.
The raster already knows
All the production maps turned out to be indexed-color PNGs — vector-exported floor plans with flat color regions sharing one palette design system across venues: hall floor, exhibitor-zone floor, concourse, booths, and transparency outside the building each hold a stable palette color. Masking the walkable colors yields the true floor, corridors and pillars included, with real corridors starting around 8–11px and sub-4px runs identifiable as anti-aliasing artifacts.
So the decision: segment the walkable floor by color, store it as a compact navigation grid in the published bundle, and route with A* over the grid. This is deterministic image processing — reproducible, free, instant, able to backfill every existing map without human re-review — not an LLM extraction.
What still needs a human
Two things the raster cannot supply. Cross-map connectivity — which exit on one floor leads to which entrance on another — is semantic and is authored as a small portal graph by hand. And there is no meters-per-pixel scale, so distances and ETAs need separate calibration.
One boundary held deliberately: the existing cart visit-plan line keeps its own corner-visibility planner, verified at zero fallbacks and zero booth crossings across all 7,058 booth pairs on all eight maps. Navigation is a separate engine; making one serve both jobs was rejected as wasted work.