Source-linked AI summary
Multilingual Agent System for Inclusive Wildfire Evacuation Guidance
Shruti Kulkarni, Lynn Tong, Aditi Namboodiripad, Chelyah Miller, Helen Lin, Peeyush Patel, Bogdan Bistriceanu, Diane Myung-kyung Woodbridge
TL;DR
BEACON addresses the limited accessibility and operational shortcomings of wildfire evacuation information, especially for people with limited English proficiency. It combines danger-aware, polygon-avoidant routing and personalized evacuation support with a context-aware multilingual agent. Preliminary qualitative evaluation demonstrated reliable language handling and avoidance of active fire polygons, while future work targets broader data, platform, and interaction support.
Problem
Over 80% of U.S. emergency alerts and evacuation orders are issued exclusively in English, while commercial routes and static checklists may not reflect emergency exclusions or household needs.
Method
BEACON integrates Watch Duty, NOAA, and user data with polygon-avoidant routing, personalized checklists, and an LLM-driven multilingual agent.
Results
Qualitative evaluation found reliable first-message language identification, interface localization, mid-session language switching, and avoidance of every active fire polygon in a preliminary Montrose County test.
Takeaways & Limitations
BEACON provides integrated evacuation guidance that connects multilingual conversation with route status and personalized checklist support.
Abstract
from arXiv · showhide
Wildfire seasons have become 84 days longer in the current days than in the 1970s, causing enormous threats to one's financial status and short- and long-term health. During the fire, public agencies send out emergency messages to provide warnings and orders. Although 26 million people in the US have limited English proficiency, over 80% of those messages are only delivered in English, which can cause disproportionate information distribution and awareness. In order to better serve marginalized communities during emergencies, the authors developed BEACON, a service that provides comprehensive and personalized evacuation guidance, including navigation routes, personalized checklists, and a chatbot in the language that a user uses. Our current system ingests data including fire perimeter information, evacuation order status, and shelter information from Watch Duty. When a user is within a certain proximity from the fire, the system utilizes real-time GPS locations and nearby weather data from the National Oceanic and Atmospheric Administration (NOAA) to predict fire danger levels. The assessment model refreshment are dynamically scheduled based on fire progress and trends using XGBoost. If the location has a likelihood of fire danger, the system sends alerts with evacuation routes outputted from a polygon-avoidant routing pipeline. The application provides a context-aware multilingual agent that users can communicate with and is tightly connected to other features of the application. In addition, based on data that the user entered, the system dynamically generates and checks off personalized reminder items to provide an organized evacuation plan. The system's user interface dynamically changes its language settings based on the language the user most recently used in either setting or chatbot conversation for all the application elements.
I. INTRODUCTION
BEACON addresses language and operational gaps in wildfire evacuation support by combining multilingual guidance with personalized routes and checklists. Its architecture connects mobile GPS, external emergency data, danger assessment, and user-facing evacuation assistance.
- Motivation: 84 days longer wildfire seasons and English-only alerts affecting over 80% of warnings heighten the need for accessible evacuation guidance.The paper reports 25.7 million U.S. residents with limited English proficiency.
- Operational gaps: Commercial GPS routes may cross fire perimeters or evacuation areas, while agency checklists often remain static and household-insensitive.The paper identifies needs involving elderly people, children, pets, livestock, mobility, medical conditions, and vehicle availability.
- System overview: BEACON connects a user’s GPS ping to danger assessment, polygon-avoidant routing, multilingual evacuation steps, and personalized checklists.Figure 1 presents this end-to-end data flow.
- System overview: The system ingests fire, evacuation, weather, and points-of-interest data to trigger alerts and generate evacuation routes when users approach danger.Sources include Watch Duty, NOAA, and Overpass.
B. Danger assessment
BEACON assesses wildfire danger using Hourly Wildfire Potential and adapts reassessment timing to recent danger patterns. The resulting score also constrains route selection in high-risk areas.
- HWP computation: Hourly Wildfire Potential combines wind gust, temperature, dewpoint, soil moisture, and snow water-equivalent fields on the HRRR grid.The user’s HWP is taken from the nearest grid point after spatial smoothing.
- HWP computation: HWP = 0.213 · G^1.50 · VPD^0.73 · (1 − M)^5.10 · S.VPD is computed from temperature and dewpoint, M from soil moisture availability, and S from snow water equivalent.
- Routing integration: Candidate routes are discarded when their length exceeds configurable fraction ϕ_max in cells where HWP exceeds τ_HWP.This incorporates assessed fire danger into routing constraints.
- Adaptive reassessment: XGBoost classifies danger as safe, elevated, high, or extreme and schedules the next HWP refresh using prior values, momentum, and 24-hour means.The refresh cadence is adaptive because HWP calculation is time- and resource-consuming.
C. Route Generation
BEACON generates shelter routes that avoid active fire, evacuation, potential-danger, and hazardous regions. Its routing pipeline also handles unavailable waypoints and missing safe routes with fallback guidance.
- Route constraints: BEACON uses OpenRouteService with GeoJSON hard constraints to return paths that avoid supplied polygons and points.This addresses the inability of evaluated platforms to render or avoid emergency exclusion regions.
- Route constraints: The avoid region A is the union of active fire perimeters, evacuation areas, potential fire dangers, and hazardous points.These regions form the core exclusion geometry for route computation.
- Path generation: The pipeline computes a shortest-W multi-stop path from origin o through optional waypoints to candidate shelter s without crossing A.Waypoints are selected by the user, and shelters come from Overpass and Watch Duty.
- Route validation: Fig. 2 renders fire and HWP regions alongside safe routes, while preliminary testing found the pipeline avoided every active fire polygon near Montrose County.Commercial baseline APIs returned paths crossing explicitly excluded regions for the same geometry.
- Graceful degradation: When waypoint-constrained searches fail, BEACON provides a route without waypoints; when no safe route exists, it sends emergency instructions through chat.The system also notifies users when a requested waypoint is unavailable.
D. The Multilingual Agent
BEACON uses an LLM-driven agent to coordinate routes, checklists, preferences, and multilingual interaction. Session context and persistent language settings let the agent respond in the user’s language while updating the broader interface.
- Agent functions: The LLM-driven agent selects among route, checklist, and preference features from chat keywords or classified intended uses.For example, reporting packed medicine checks the corresponding checklist item.
- Session context: Each session prompt includes identity, location, evacuation urgency, route status, personalized checklist context, and language rules.The context includes twelve household questions and checklist status.
- Multilingual interface: The most recently used language persists across chat and preferences and updates buttons, settings, route instructions, and checklists.Language changes are stored in the database and application.
- Session context: Qdrant stores embedding-based cross-session memory so recent conversation summaries and user profiles can be injected into later sessions.The agent uses this information as what it knows about the user.
III. RESULTS
The evaluation examined four end-to-end system properties through qualitative observations, using nine large Colorado fires from July to September 2025.
- The evaluation focused on automatic language detection, polygon-avoidant routing, context-aware conversation, and graceful degradation.
- Nine Colorado fires exceeding 1,000 acres were examined from July to September 2025.
A. Automatic language detection without configuration
BEACON identified the user’s language from the first message and replied in that language without manual selection. The interface also followed language switches during a session.
- The agent reliably detected the user’s language from the first message and responded in the same language without manual configuration.
- The mobile interface was displayed in the user’s language and followed switches from the most-recent message during the session.
B. Polygon-avoidant routing under realistic perimeters
The polygon-avoidant routing pipeline consistently avoided every active fire polygon in a realistic Montrose County, Colorado test. Preliminary baseline routing APIs produced paths that crossed excluded regions.
- The pipeline consistently returned paths avoiding every active fire polygon near Montrose County, Colorado.For the same input geometry, preliminary commercial routing APIs returned paths crossing explicitly excluded regions.
C. Context awareness from the first turn
The agent responded with full context on the first user turn because the session-start prompt included danger, route, household, and conversation information.
- A first-turn “Should I leave?” query produced a tailored response referencing perimeter distance, the pre-loaded route, and household-specific items.
D. Graceful degradation under operational failures
BEACON is designed to preserve useful evacuation assistance when routing, language, classification, memory, or onboarding inputs fail. Tested fallback paths retain degraded responses and preserve emergency-service escalation.
- A core design principle is that no single external-service failure should leave a user in immediate danger without a response.Each external data source or service is intended to continue in a degraded mode.
- When no shelter route is safe, the agent instructs the user to call 911 and dispatches a push notification.
- If a requested waypoint cannot fit near shelter candidates, the pipeline retries without it and informs the user that the waypoint was unavailable.
- Unsupported routing languages fall back to English route requests with Claude-translated turn-by-turn instructions.ORS natively supports 23 languages.
- Classifier, memory, and household-profile failures reduce context or intent handling, while observed degraded sessions still provided useful responses and preserved emergency escalation.
IV. CONCLUSION
The authors developed a multilingual agent that provides fire-related safe routes and advice for communities with limited English. They report a qualitative experiment demonstrating reliability and effectiveness, while identifying future expansion of data sources and platform support.
- The preliminary system provides multilingual safe routes and advice to communities with limited English during fires.
- A qualitative experiment demonstrated the reliability and effectiveness of the implemented features.
- Future work will add fire and live-shelter APIs and databases, develop an Android application, and explore speech-to-text support.