Original title: GPEvac: GNN-Based PPO for Adaptive Evacuation Routing During Shooting Events
Daniel Perkins, Subhadeep Chakraborty · September 2026
Mass shootings create dangerous chaos, yet current escape plans are either too rigid for new buildings or too slow to compute safe paths. Researchers created GPEvac to guide people to safety during active shootings. The system treats a building like a connected map of points and paths. It uses an edge-first message-passing method and a special virtual point to track both nearby hallways and distant exits. A reinforcement learning policy then scores the best escape options. Because of this design, one single trained policy works across different building shapes and sizes. In detailed computer simulations, GPEvac beat other smart systems by greatly reducing people's total exposure to danger. It calculates full escape routes in only 14.73 ms on a standard computer processor, fast enough to use live security camera feeds.
Imagine you are directing traffic at a festival where a sudden hazard appears—say, a fire near one stage. You have walkways that connect different areas, and people are already crowded in some paths. The usual plan would be to tell everyone to follow one fixed route out. But that does not work because the safest path changes as people move and the hazard shifts. In this paper, the system builds a map of the festival grounds, watches where the hazard is and where crowds are, and reroutes each person in real time using that updated map. It does all the thinking in under 15 milliseconds.
Active shooter events are terrifyingly common, and current advice ('run, hide, fight') is too vague to actually save lives. This system computes a safe evacuation path for every person in a building, tailored to the shooter's location and crowd movement, in the time it takes to blink. It could be plugged into live surveillance to guide people to exits without running into danger.
The abstract does not specify how the system performs in simulations of real-world conditions like smoke, panic, or communication failures. It also claims transferability to other domains but provides no experimental evidence for that claim. The 14.73 ms runtime is on a CPU, but real deployment with live cameras and many people may require more testing.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
GPEvac: GNN-Based PPO for Adaptive Evacuation Routing During Shooting Events — Daniel Perkins, Subhadeep Chakraborty
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.