Evacuation Simulator

Inspiration

We wanted to know how simulating people moving

What it does

Evacuation Simulator is a web app for sketching buildings and estimating how long people take to get out.

Draw polygonal rooms, corridors, and stairs; place doors, exits, obstacles, and occupant groups
Work across multiple floors with linked stairs (ascent slower than descent)
Overlay PNG floor plans and trace over them.
Place hazards, a flood or fire, and watch how they change routes and outcomes.
Hit Run and watch agents evacuate in real time (streaming animation as the sim computes)
Inspect times, waits, congestion hotspots, and per-occupant routes

Built-in scenarios include offices, dual towers, a stadium, the RMS Titanic, and The Backrooms.

How we built it

Frontend: React + TypeScript + Vite, with a Konva canvas editor for polygons, placement tools, multi-floor tabs, floor-plan overlays, and playback (pause, speed, reset).

Backend: FastAPI + Pydantic + SQLAlchemy/SQLite. Buildings and simulation snapshots live in the DB. Runs stream frames as NDJSON so the UI can animate before the full timeline finishes.

Simulation engine (pure Python):

Builds a visibility-graph navigation network (openings + reflex-corner waypoints)
Routes with Dijkstra, with queue-aware exit assignment at spawn
Steps agents in discrete time (~0.25 s) with body radius, wall containment, and aperture-limited doors/exits
Models capacity at openings and stairs, plus fire / flood / smoke with different spread rules

Challenges we ran into

Wall tunneling: discrete timesteps made agents slip through thin walls until we tracked current-space membership and clamped motion through door apertures only.
Valid paths in polygons: Euclidean chords often leave the room; we needed a visibility graph with reflex corners and obstacle clearance so routes stay inside spaces.
Congestion that feels real: abstract flow rates weren’t enough — body radius vs opening width, slot assignment, and jams outside throats took iteration.
Hazards + playback: expanding fire/flood/smoke had to stay in sync with pause, scrub, and speed changes via per-frame plume state.
Flood physics lite: room-scoped fill with gravity down stairs (then upward) was much harder than a simple radial blob, but far more believable.
Large runs staying responsive: streaming NDJSON frames so the browser starts animating while the engine is still computing.

Accomplishments that we're proud of

A full loop: draw → save → simulate → stream → animate → stats
Polygonal multi-floor geometry with visibility-graph routing (not a crude grid-only model)
Bottlenecks driven by body size and opening width, so door width actually matters
Three hazard systems with distinct behaviors (radial fire, doorway flood with gravity, room-scoped smoke)
Showcase maps that make demos memorable (Titanic, stadium, dual offices, Backrooms)
Progressive streaming so long sims feel interactive
Clear assumption docs so the tool stays useful and honest

What we learned

Crowd egress is as much about geometry and capacity as about shortest path
Small discrete-time bugs (containment, apertures) dominate perceived realism
Streaming results early matters more for UX than perfect offline batch compute
A sharp editor UX (snap, undo, multi-floor, overlays) is half the product; the sim alone isn’t enough

What's next for Evacuation Simulator

• Mid-run dynamic replanning when hazards cut routes (instead of mostly fixed spawn routes) • Richer behaviour models (panic, following, mobility differences — we already reserve a behaviour field) • Better hazard physics and optional validation against real drills / published benchmarks • Accessibility-focused scenarios and reporting • Stronger import from real CAD/BIM floor plans • Collaborative editing and side-by-side scenario comparison (“wider doors vs second stair”)