2026-06-04 米国国立標準技術研究所（NIST）
◆従来の避難シミュレーションでは、人間行動を単純化した仮定に基づくことが多かったが、本モデルは実際の行動データや環境条件を学習し、煙の拡散や混雑状況の変化に応じた避難行動をより現実的に予測できる。研究者らは、この手法によって建築物や公共施設の避難計画を改善し、非常時の安全性向上につなげられると期待している。また、AIは避難経路の設計や建築基準の評価にも活用可能であり、高齢者や身体的制約のある人々を含む多様な利用者への配慮も検討できる。今回の成果は、防火工学と人工知能を融合した新たな避難安全評価技術として、建築設計や災害対策分野への応用が期待される。
NIST researchers developed a new AI model that can identify safe evacuation routes during a fire. The model can be used with new electronic exit signs, called dynamic emergency exit displays, to show whether an exit is safe to use. Credit: A. Kim/NIST
＜関連情報＞
- https://www.nist.gov/news-events/news/2026/06/new-ai-model-shows-how-evacuate-fires-one-safe-step-time
- https://www.sciencedirect.com/science/article/abs/pii/S2352710225033698
強化学習を用いた建物火災避難のための、居住可能性に基づく経路計画モデルの開発 Development of a tenability-based path planning model for building fire evacuations using reinforcement learning
Hongqiang Fang, Wai Cheong Tam, Ruggiero Lovreglio, Md Ismail Siddiqi Emon, Michael Xuelin Huang
Journal of Building Engineering Available online: 30 December 202
DOI:https://doi.org/10.1016/j.jobe.2025.115132
Highlights
- A path planning model, SafeStep, adapted to evacuees’ tenability is developed.
- A reward function incorporating evacuees’ toxic gas exposure via FED is formulated.
- The proposed model is benchmarked against the Dijkstra’s algorithm approach.
- The proposed model can be used to facilitate dynamic directional exit sign control.
Abstract
To enhance the safety of building fire evacuations by accounting for the dynamic and cumulative impacts of hazardous fire conditions on evacuees, we developed SafeStep, a novel reinforcement learning-based path planning model that integrates occupant tenability into evacuation decision-making. Specifically, the model employs the Fractional Effective Dose (FED) of toxic gases to quantify dynamic fire impacts, and a FED-derived reward function is formulated to guide the RL agent toward safer and more efficient paths. A Deep Q-Network (DQN) is adopted to optimize evacuation path planning within complex building geometries. SafeStep is benchmarked against traditional path planning algorithms, Dijkstra’s algorithm (DA), through two test cases. In one case, the results show that SafeStep provides safer evacuation paths, achieving approximately a 62 % reduction in FED exposure compared to DA. In the other, it successfully identifies viable evacuation routes in scenarios where the DA-based model fails. To further assess its applicability, a case study in a complex building geometry shows that SafeStep can consistently generate evacuation paths with lower FED across arbitrary starting positions. These findings indicate that SafeStep effectively addresses key limitations of traditional path planning algorithms, which often fail to account for evolving fire dynamics and cumulative fire effects. As such, the proposed model has strong potential to support smart building technologies, such as dynamic directional exit signs, to enhance evacuation safely and efficiently in real-world fire emergencies.
