USC develops AI framework to coordinate Los Angeles wildfire evacuations

Wildfire evacuation guidance in Los Angeles could eventually reflect not just where flames are moving, but which roads remain open and who needs a ride. USC on November 11 described EVAC-AI, a four-year, NSF-funded research project to connect those decisions. Led by assistant professor Ruolin Li, the team is developing a framework, not introducing a finished evacuation service.

The National Science Foundation’s Focused Interdisciplinary Research in Engineering program is funding the work. Its intended users span emergency managers, local agencies, residents and mobility providers. Li’s team wants to anticipate evacuation bottlenecks before they become crises, linking fire behavior, transportation and people’s movements rather than treating each as a separate planning problem.

A fire forecast becomes a road-access forecast

The project targets a mismatch that Li describes between existing models. Fire forecasts can predict a blaze’s spread without accounting for the road network beneath it. Evacuation models, meanwhile, may miss ember-driven spot ignitions: new fires that can make a road inaccessible ahead of the visible fire perimeter. A route’s usefulness can therefore change before the main blaze reaches it.

EVAC-AI would use probabilistic fire modeling, including the possibility of those spot ignitions, to translate changing fire conditions into road-access information. The researchers envision updates every five minutes at the census-tract level. That is a proposed update schedule, not a demonstrated operating speed. The goal is to let transportation decisions change alongside the fire instead of relying on a fixed picture.

Our goal with EVAC-AI is to approach evacuation as one interconnected system rather than a collection of separate problems

Ruolin Li, USC assistant professor and project lead

Changing who leaves, and when

Los Angeles presents a particular coordination challenge. Fires can begin in hillside vegetation near densely populated communities, while mountain and canyon roads offer limited ways out. USC describes how alerts that send whole neighborhoods onto those roads at once can create traffic and delays. The project’s messaging research addresses that surge, not just the selection of individual routes.

Researchers will explore targeted, staggered evacuation instructions rather than relying only on a mass alert. That includes deciding who receives a message, what it contains and when people are encouraged to leave. The intended result is less pressure on the road network while keeping routes available for emergency operations. Whether that approach delivers those benefits remains a research question.

Residents would feed information back

At the center would be an agentic AI system connecting the project’s components. USC describes a two-way design: emergency managers would receive information and recommendations, while residents could report circumstances that affect their evacuation. Those responses would feed into the larger picture, making residents a source of operational information rather than only recipients of instructions.

The planned resident-facing app would offer a conversational interface for evacuation questions. Someone could report having no car, caring for an elderly relative or needing to leave with pets. The system could combine those needs with changing fire, road and evacuation information to tailor its guidance. USC compares the interaction to ChatGPT, but with questions focused specifically on wildfire evacuation.

Transportation assistance is another research component. The team will explore centrally dispatching commercial autonomous vehicles, citing Waymo’s fleet as an example, to collect residents without transportation or with accommodation needs. Its proposed model includes a requirement to cover households without personal vehicles. The researchers want to examine how those pickups could fit into evacuation traffic without adding unnecessary congestion; the proposal is not a confirmed Waymo partnership.