⚡ electrical engineering

Time-indexed FDS-derived Fire-risk Fields for Indoor UAV Emergency Routing in Building Fires

This study proposes a fire-risk field enrichment framework that converts transient Fire Dynamics Simulator outputs into time-indexed risk maps for indoor UAV emergency routing, demonstrating significant reductions in cumulative sampled risk compared to standard geometric pathfinding in building fire scenarios.

Original authors: zhijun Gao, mingqi Yu, xin Wang

Published 2026-10-01

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: zhijun Gao, mingqi Yu, xin Wang

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

When a fire breaks out inside a building, the danger is not just the flames themselves, but the invisible and rapidly changing environment they create. Heat rises, thick smoke fills corridors, toxic gases spread, and visibility drops to zero, often turning a clear hallway into a deadly trap within minutes. For firefighters and rescue teams, knowing which path is safe is a matter of life and death, yet traditional maps only show walls and doors, not the shifting clouds of heat and poison. To solve this, researchers are turning to small, unmanned aircraft that can fly ahead of human responders, scouting the interior without risking a single life. However, for these machines to be truly useful, they need more than a simple map of obstacles; they need a way to "see" the invisible dangers of a fire and choose a route that avoids them, even if that route is slightly longer.

A team of researchers at Shenyang Jianzhu University has developed a new method to give these drones this critical awareness. Instead of treating a fire as a single point of danger or a static red zone on a map, they created a system that translates complex computer simulations of fire behavior into a dynamic, time-based risk map. By using a powerful fire modeling tool called FDS, which simulates how heat, carbon monoxide, and smoke move through a building over time, the team generated a series of detailed snapshots of the fire's evolution. They then converted these physical measurements into a unified "risk field," a digital layer that assigns a danger score to every part of the building at every second of the fire's progression. This allows a drone's navigation computer to understand that a short path might lead straight into a cloud of toxic gas, while a slightly longer detour keeps it in safe air.

The researchers tested this approach in a simulated building fire, comparing their new risk-aware navigation system against a standard routing method that only looks for the shortest geometric path. In a fixed scenario where a drone needed to travel from a starting point to a goal, the standard system chose a route that passed through increasingly dangerous areas as the fire grew. In contrast, the new system recognized the rising threat and selected a different path that avoided the worst conditions. The results were striking: the new system reduced the total amount of risk the drone was exposed to by nearly 99.67 percent compared to the standard method. This massive reduction was achieved not by flying faster or stronger, but by simply choosing a smarter path that traded a little extra distance for a massive gain in safety.

To ensure this method works in a realistic setting, the team did not stop at computer calculations. They took the planned flight paths and replayed them inside a sophisticated robotics simulation environment. In this virtual test, the drone's position was set along the planned route to check if the path was physically possible and free of collisions with walls or furniture. The replays were successful, confirming that the routes generated by the risk-aware system were not just theoretically safer, but also practically flyable within the building's layout. Furthermore, the researchers tested the system with different starting and ending points, simulating various emergency scenarios. In almost every case, the new system found a safer route, reducing risk exposure by anywhere from 38 percent to nearly 100 percent, depending on how much space was available to maneuver around the fire.

The study also explored what happens when the fire changes while the drone is flying. By updating the risk map at discrete time steps using precomputed data, the system was able to recalculate the path in an offline simulation, further lowering the risk by an additional 28 percent. This demonstrates that the method can adapt to a changing environment, a crucial feature for real-world emergencies where conditions evolve rapidly. The researchers were careful to note that their work is currently a simulation-based proof of concept. They used pre-computed fire data rather than real-time sensors, and their robotics tests were a "set-pose" replay, meaning the drone was virtually placed along the path rather than flying autonomously with its own flight controls. While this does not yet prove the system works in a real burning building, it establishes a vital bridge between fire science and robotics.

This work represents a significant step forward in how we prepare for indoor fire emergencies. It moves beyond the idea that a safe route is simply one that avoids walls. Instead, it shows that safety is a dynamic calculation involving heat, gas, and visibility, all of which change by the second. By converting complex fire physics into a format that a drone can understand and use to make decisions, the researchers have provided a blueprint for future rescue systems. These systems could one day allow unmanned aircraft to scout dangerous buildings, find safe paths for human teams, and locate victims without ever exposing a firefighter to the lethal conditions of a smoke-filled room. The simulation results suggest that with the right data, a drone can navigate a fire not just by seeing where it can go, but by understanding where it should not.

Technical Summary: Time-indexed FDS-derived Fire-risk Fields for Indoor UAV Emergency Routing in Building Fires

Problem Statement

Indoor building fires present a critical challenge for emergency response due to rapidly evolving hazardous conditions, including heat, smoke, toxic gas accumulation, and visibility loss. While Small Unmanned Aerial Vehicles (UAVs) offer a promising solution for pre-entry reconnaissance to reduce firefighter exposure, current routing methodologies often fail to account for the dynamic nature of fire environments. Existing studies frequently represent fire as a static point source, a prescribed hazard zone, or a single static risk map. These simplifications do not capture the temporal and spatial evolution of temperature, carbon monoxide (CO) concentration, and visibility within a building's geometric layout. Consequently, a route that is geometrically collision-free may still traverse regions that are thermally severe, toxic, or visually degraded, rendering them unsafe for UAV operations. Furthermore, a gap exists between fire simulation outputs (typically used for evacuation analysis) and robotics-oriented path planning (which relies on obstacle grids), limiting the practical deployment of fire-informed routing strategies.

Methodology

The authors propose a physics-informed fire-risk field enrichment framework that converts transient Fire Dynamics Simulator (FDS)/PyroSim outputs into route-level risk information for indoor UAV routing. The methodology is organized into five stages:

Fire-Field Extraction: FDS/PyroSim simulations generate time-indexed outputs for temperature, CO concentration, and visibility fields at a specific UAV flight height ( $z = 1.7$ m) over a 150-second fire evolution period.
Risk Normalization and Fusion: The physical fields are normalized into three risk channels ( $R_T$ , $R_{CO}$ , $R_S$ ) using specific clipping functions based on baseline values and the 95th percentile of the field data. These are fused into a single integrated fire-risk field $R(i, j, t)$ using a weighted sum (0.30 for temperature, 0.35 for CO, 0.35 for visibility).
Cost-Augmented Routing: The framework employs a cost-augmented $A^*$ algorithm. Unlike the Standard $A^*$ baseline, which minimizes geometric path length, the proposed planner incorporates the normalized fire-risk value into the traversal cost function: $g(n') = g(n) + d(n, n') + \lambda_{risk}R(n', t)$ . This allows the planner to trade increased path length for reduced hazard exposure.
Dynamic Replanning: An offline dynamic replanning strategy is tested by updating the risk map at discrete time intervals ( $t = 30, 60, 90, 120, 150$ s) and re-planning the path from the current position to the goal.
ROS2-Gazebo Replay: To bridge the gap between fire simulation and robotics, the planned trajectories are replayed in a ROS2-Gazebo environment using a "set-pose" protocol. This involves directly assigning the UAV model's pose along the planned path to verify spatial consistency, obstacle feasibility, and collision status. A qualitative fire/smoke visualization layer derived from FDS data is also added to the Gazebo scene for interpretation.

Key Contributions

The paper outlines five primary contributions:

Enrichment Workflow: Development of a workflow to convert transient FDS/PyroSim fields (temperature, CO, visibility) into time-indexed, route-level fire-risk fields specifically for indoor UAV reconnaissance.
Physics-Based Representation: Moving beyond simplified point-source or static hazard representations by utilizing a sequence of physics-based fire-simulation fields as routing-relevant hazard information.
Multi-Channel Fusion: Implementation of a multi-channel risk fusion model at the UAV flight height, while retaining the obstacle occupancy map as a hard geometric constraint.
Cost-Augmented Formulation: A routing formulation that embeds time-indexed fire-risk values into the path traversal cost, enabling explicit risk-length trade-off analysis under evolving fire conditions.
Simulation Bridge: The use of ROS2-Gazebo set-pose replay as an intermediate layer to check map-level route consistency and visualize fire/smoke scenes alongside UAV poses, facilitating the transition from fire-safety simulation to robotics-side assessment.

Results

The framework was evaluated using a single-fire indoor building scenario with five time points ( $t = 30$ to $150$ s) and five different start-goal configurations.

Risk Reduction (Fixed Task): For a fixed start-goal task, the fire-risk-aware planner reduced the cumulative sampled risk by 99.67% compared to the Standard $A^*$ planner (0.285914 vs. 86.485638 normalized risk units).
Robustness (Multi-Start-Goal): Across five different start-goal configurations at $t=150$ s, the risk-aware planner achieved risk reductions ranging from 37.9% to 99.9%, depending on the availability of safer alternative corridors.
Dynamic Replanning: Offline dynamic replanning further reduced the cumulative sampled risk by 28.32% compared to a static path plan.
Replay and Geometry: All planned trajectories were successfully replayed in ROS2-Gazebo with zero recorded collisions or set-pose failures. Path geometry screening confirmed that while risk-aware paths were often longer and more tortuous, they maintained geometric connectivity and sufficient obstacle clearance.

Significance and Claims

The paper claims that transient fire simulation outputs can provide actionable, route-level support for indoor UAV fire-response reconnaissance. The significance of the work lies not in validating combustion models or UAV flight controllers, but in linking fire-safety simulation with robotic decision-making at a decision-support level.

The authors emphasize that their framework addresses the engineering gap where fire simulations and robotics planners typically operate in separate domains. By converting fire fields into sampled route costs, the study makes the effect of evolving hazards measurable at the path level. The ROS2-Gazebo interface serves as a practical intermediate step, allowing for the verification of spatial consistency and the visualization of fire risks alongside UAV trajectories before more expensive steps like hardware-in-the-loop testing.

The authors remain modest regarding the current scope, explicitly stating that the work represents a simulation-based decision-support step rather than a complete rescue system. They acknowledge limitations, including the use of precomputed (non-real-time) fire fields, the reliance on set-pose replay rather than dynamic flight simulation, and the testing on a single scenario. Future work is identified as necessary to address online risk updating, reduced-order fire-field estimation, and dynamic UAV simulation.