NIST researchers develop AI model called Safe Step to predict fire spread and redirect building occupants to safer evacuation routes

NIST researchers and colleagues created Safe Step, an AI model using reinforcement learning that forecasts how fire evolves and recommends the safest evacuation route. The model can be integrated with dynamic emergency exit displays to guide occupants in real time.

Unlike traditional shortest-path algorithms that only consider current building conditions, Safe Step uses the fractional effective dose (FED) of toxic gases—a fire safety metric representing cumulative hazard exposure over time—to account for how fire hazards change as an occupant moves toward an exit.

In test cases, Safe Step consistently identified safe routes where traditional algorithms would not. For example, it can direct an occupant away from the nearest exit if fire growth makes that route increasingly dangerous, toward a more distant but safer exit.

Researchers estimate technologies like Safe Step could appear in five to 10 years, pending regulatory approval, reliability testing, and integration with existing safety systems. Current limitations include single-story floor plans; next steps include multilevel buildings and multi-agent modeling to handle congestion and coordinate firefighter access.