Predicted fire hazards reduced exposure, but fewer people were found.
A virtual office lets us change how a fire starts and spreads, then test how a planner responds to hazards it cannot directly see.
Simulate many fires on an office floor, train a fast hazard predictor, and test whether a rescue planner can use it to navigate with less heat and smoke exposure.
Edited demonstration briefThe result
Lower heat and smoke exposure did not improve rescue coverage.
A learned hazard model gave the planner a view beyond nearby sensors. It reduced heat and smoke exposure, but found fewer people than the sensor-based comparison.
Synthetic building and simplified point drone; no real rescue or experimental fire validation.
Recorded simulation
Heat and smoke spread through the simulated office.
Move through the recorded fire. Switch between heat and visibility.
Recorded outputs, replayed in your browser. No new simulation is running.
Simulated fires expose hazards beyond the planner’s nearby sensors.
The world we created
NIST’s Fire Dynamics Simulator computes heat, smoke visibility and carbon monoxide as fire moves through rooms, doors and vents. The campaign varies ignition, openings and ventilation, and adds noisy simulated sensor readings.
What this enables
A learned model predicts hazard maps beyond nearby sensors. A scripted planner uses those maps alongside sensor readings and a safety rule. It is the hazard predictor that learns, rather than the rescue policy.
Varied fires paired sensor readings with hidden hazards.
Vary ignition, doors and ventilation. Record heat, smoke and toxic gas alongside noisy sensor readings. Failed simulation runs are excluded.
A learned model estimated hazards beyond nearby sensors.
Train a model to estimate how hazards develop from the available observations.
Unseen fires exposed the tradeoff between exposure and coverage.
Replay decisions in unfamiliar simulated fires and compare exposure with rescue coverage.
The fire simulator creates training targets and unseen test environments. Replaying decisions makes both the gains and the tradeoffs measurable.
Original research artifacts
The original files retain the recorded measurements and model details.