A NIST-led team has unveiled Safe Step, an artificial intelligence model that can dynamically direct building occupants to the safest exit during a fire by predicting how the blaze will spread and updating route guidance in real time. Described in the Journal of Building Engineering, the model is designed to drive a new generation of "dynamic emergency exit displays" being tested in smart buildings.
From shortest path to safest path
Traditional emergency egress algorithms guide evacuees toward the nearest exit using current conditions. Safe Step instead uses reinforcement learning trained on NIST's Fire Dynamics Simulator and the Fire Data Generator to forecast fire growth and choose the route that minimises cumulative toxic gas exposure — measured as fractional effective dose (FED).
How it deploys
"Fires can grow and spread," said Hongqiang "Rory" Fang, NIST research associate and the paper's first author. "Our model forecasts how the fire is evolving and can help update emergency exit displays to direct people toward the safest exit." In real-world use, Safe Step does not run a fire simulation in the moment — it relies on live sensors measuring temperature and air quality to continuously adjust its recommendations.
What comes next
The current model handles a single-story floor plan; a multilevel version with multi-agent coordination for crowd flow is in the works. NIST estimates Safe Step-style technologies could reach commercial buildings in five to ten years pending regulatory and reliability work. The release lands amid a broader push for AI in public safety applications, alongside other foundation-model debuts like OpenAI GPT-5.5 and Anthropic's Claude Fable 5.
Reporting based on coverage from NIST, TechXplore and Journal of Building Engineering.