A research demo of a multi-agent urban-planning pipeline — sense a street, negotiate its needs, govern what gets built — rendered on real geography with deck.gl + MapLibre.

Pipeline

flowchart LR
    S["Sense<br/>R1–R4 records,<br/>heat / vulnerability fields"]:::sense --> C["Conflict<br/>Detection"]:::conflict
    C --> E["Equity<br/>Review"]:::equity
    E --> O["Orchestrator<br/>budget allocation"]:::orch
    O --> D["Design<br/>trees, benches,<br/>canopy"]:::design
    D --> F["Feedback<br/>agents re-route,<br/>heat exposure"]:::feedback
    F --> L["Lifecycle<br/>as-built review"]:::lifecycle
    classDef sense fill:#2F2F31,stroke:#889A51,color:#EDEDEA
    classDef conflict fill:#2F2F31,stroke:#DC312F,color:#EDEDEA
    classDef equity fill:#2F2F31,stroke:#C9A932,color:#EDEDEA
    classDef orch fill:#2F2F31,stroke:#E07041,color:#EDEDEA
    classDef design fill:#2F2F31,stroke:#527855,color:#EDEDEA
    classDef feedback fill:#2F2F31,stroke:#889A51,color:#EDEDEA
    classDef lifecycle fill:#2F2F31,stroke:#C9A932,color:#EDEDEA

Loading

Three cases run through the same pipeline, each on frozen OpenStreetMap geometry with a locally packaged CARTO Dark Matter vector style and 3D-extruded buildings:

Case	Site	Role
Chicago	New ERA Trail corridor segment (OSM way 624189839), 40 m analytical envelope	live movement simulation
London	Mitre Yard brownfield (OSM way 49601059)	multi-constraint case
Suzhou	pocket retrofit parcel (OSM way 741252447)	low-complexity anchor

The interface is a workflow demonstrator. Its R1-R4 records, heat/vulnerability fields, live movement metrics and controlled implementation deviation are frozen illustrative inputs or outputs, not resident observations or completed Experiment 1 results.

Run

npm install
npm run dev        # http://localhost:5173

What is computed (not scripted)

• Conflict Detection — pairwise distances in metres (spatial), overlapping construction windows (timing), and the budget line (funding).
• Equity Review — every protected group checked against an adjustable access floor.
• Orchestrator — greedy budget allocation, protected needs first, with floor enforcement that reallocates from non-protected needs.
• Design — trees and benches scale with the money that actually reached them; the heat field contracts with built canopy.
• Feedback — pedestrians re-evaluate routes by role-weighted utility; heat exposure is accumulated per frame from moving agents.
• Lifecycle review — the as-built alignment is compared against the approved design and flags oversight.

Drag the budget and equity floor sliders at any step — the whole pipeline recomputes and re-renders in real time. Export run downloads the full pipeline state, decision trace and live metrics as JSON.

Structure

file	role
src/scenario.ts	the three cases in real lng/lat (needs, zones, paths)
src/pipeline.ts	pure pipeline: conflicts / equity / orchestration
src/sim.ts	pedestrian & vehicle simulation + per-frame metrics
src/mapLayers.ts	deck.gl layer factory (heat, conflicts, greenway…)
src/copy.ts	stage copy shown in the UI
src/App.tsx	app shell: map, camera choreography, panels

Research artifacts

Beyond the interactive demo, this repository carries the study's reproducibility materials (Python 3.9, pip install -r requirements-python39.txt):

directory	contents
world_model/	Three-city world-model build pipeline: immutable Overpass/USGS/city open-data snapshots (raw/), per-city metric GeoPackages (vector/, raster/), provenance metadata, and the build scripts that turn frozen OSM ways into Experiment 1 base inputs
preregistration/	Experiment 1 preregistration package: run manifests and SHA256 freeze manifests, scenario packages, analysis and validation scripts (validate_preregistration.py, experiment1_runner.py), agent prompt schemas and API configuration, Rhino/Grasshopper export contracts, SOPs and ethics templates

Site geometry is frozen to the three OSM ways above; raw snapshots record their retrieval timestamps. OSM extracts are © OpenStreetMap contributors (ODbL); USGS 3DEP and city open-data sources are credited in world_model/metadata/.

Basemap: CARTO Dark Matter (© OpenStreetMap contributors, © CARTO).

License

Code is released under the MIT License. Spatial data snapshots and derived layers retain the licenses of their original providers (OSM/ODbL, USGS public domain, city open-data portals) — see the notice in LICENSE and the provenance records in world_model/metadata/.