EVAC-AI: USC wants AI agents to coordinate fire, traffic and Waymo to evacuate Los Angeles, but it's still a research project

By USC Viterbi School of Engineering · November 11, 2026.

Researchers at the USC Viterbi School of Engineering have announced EVAC-AI, an agentic artificial intelligence framework to coordinate the evacuation of Los Angeles in real time during wildfires. An agentic system is one in which AI doesn't just recommend but connects data sources and continuously proposes or carries out actions. The project is funded by the U.S. National Science Foundation (NSF) through its FIRE program and will run for four years. It is led by Ruolin Li, an assistant professor of civil engineering.

The starting point, according to the university itself, is the 2025 Los Angeles fires: late warnings, alerts sent to the wrong neighborhood and gridlocked roads. The team's diagnosis is that fire-spread models don't account for the road network, while evacuation models may not consider ember-driven ignitions (embers that start spot fires ahead of the visible front and cut off roads).

EVAC-AI is designed with four components. The first translates probabilistic fire forecasts into information about which roads are passable; the team plans updates every five minutes at the census-tract level. The second is a "mixed autonomy" model that treats centrally dispatched autonomous vehicles, such as Waymo's, as a potential rescue resource, with the constraint of guaranteeing coverage for households without a car. The third is messaging: instead of a single mass alert, it will explore segmented, staggered instructions to avoid overwhelming mountain and canyon routes and to leave room for emergency crews. The fourth is the two-way agentic core: emergency managers receive recommendations, and residents provide data on their situation (vehicle, elderly people, pets) that feeds back into the system.

An app or interface for residents is also planned, described by the university as something similar to ChatGPT but focused on evacuation questions.

Our take: the project's core thesis is right to treat evacuation as a single coupled system rather than separate problems. That's where AI genuinely adds value: integrating heterogeneous sources faster than a human team can under pressure. Staggering departure orders is also an almost self-evident idea that current systems don't exploit well.

But the enthusiasm should be tempered. What has been published is a launch announcement for a four-year project: there are no results, no validated simulations and no trials with emergency services, and the article speaks in the conditional ("could," "will explore"). This is a funded aspiration, not a demonstrated capability. Tough questions also remain that the announcement doesn't answer. What happens if the agent wrongly says a route is safe? Who is accountable for a bad recommendation? How do you keep the resident chatbot from hallucinating in a life-or-death situation? For industry context, this is exactly the debate running through AI agents: the advantage won't come from giving them more autonomy, but from being able to audit them and assign responsibility. In this case, the sensible approach is for the system to recommend while a human keeps the final decision.

There's another risk: equity. The team explicitly mentions older adults, people with disabilities and households without vehicles, and the coverage constraint in the autonomous-fleet model points in that direction. If it holds up, it would be a good example of AI designed to protect those who are worst off. If it depends on an app and on data reported by residents, it could leave out precisely those with the least digital access.

The horizon that interests us is clear. In the short term, climate change will keep making fires worse and cities will have to adapt; AI will be a management tool, not a miracle fix. In the long run, a society with better prediction, coordination and abundant resources should be able to drastically reduce disaster deaths. EVAC-AI is a small step in that direction and worth following, but what will matter is what it proves in simulation and in real-world drills, not the announcement.