ATHENA: Aerial Task Handling for Emergent Networked Autonomy

Canada averages over 8,000 wildfires annually, burning more than 2.1 million hectares. Wildfire seasons routinely cause billions of dollars in damage through destroyed homes, infrastructure loss, mass evacuations, and disrupted local economies. Entire communities have been displaced for months, with recovery costs extending far beyond the fire itself into housing, insurance, and rebuilding.

These costs are accelerating; as climate conditions worsen, fire seasons are growing longer, more intense, and more expensive. Early detection and rapid situational awareness are critical to minimizing this damage, yet current aerial monitoring methods each carry significant limitations.

Satellites provide broad coverage but suffer from slow revisit rates, cloud interference, and an inability to deliver persistent detail over a fast-moving fire front. Manned aircraft offer flexibility but expose crews to hazardous smoke and cost $1,500–$5,000 per hour, making sustained deployment financially prohibitive. Consumer drones improve affordability but in practice function as remote cameras: a pilot must simultaneously fly the aircraft, interpret raw video, and relay findings manually. NASA's ACERO wildfire program has identified this lack of integrated tools and situational awareness as a major operational barrier.

The result is a fragmented monitoring landscape where detection is delayed, coverage is limited to what one pilot can observe, operator workload is unsustainably high, and multiple drones have no standard way to share intelligence. These shortcomings compound each other, keeping the gap between ignition and actionable response dangerously wide. We designed ATHENA to close this gap.

ATHENA has three main parts that work together to turn an ordinary drone into a wildfire intelligence platform.

The first is our custom PCB, which acts as the body for the brain. We designed a four-layer circuit board from scratch to house the NVIDIA Jetson Orin Nano, a powerful AI computer small enough to fly. Since no existing board could connect the Jetson properly to a drone, we built our own. It includes regulated power rails with protection circuits so voltage spikes don't fry the electronics, UART and CAN headers so it can talk to both DJI and MAVLink flight controllers, vibration-resistant connectors that stay plugged in during rough flight, and an M.2 slot for fast onboard storage. Everything is packed into a lightweight, rugged unit that can survive the heat, wind, and shaking of real wildfire conditions.

The second part is the onboard AI, which handles the actual thinking. The Jetson runs our AI in real time while the drone is flying. We trained a YOLOv11 computer vision model to spot smoke, fire, people, and obstacles from the camera feed. Once something is detected, the software combines the pixel location with the drone's GPS and orientation to calculate the exact real-world coordinates of the target. The AI also flies the drone autonomously in a "lawnmower" search pattern, guaranteeing complete coverage of a search area, and broadcasts findings to nearby drones through a lightweight mesh network.

The third part is scene reconstruction and the GUI, which shows operators what matters. On the operator's screen, we combine everything into one clear interface. Gaussian Splatting turns the video feed into an interactive 3D map of the environment, and the Canadian Fire Weather Index predicts how a fire might spread using live weather data, so responders see where the fire is going.

Our results show that ATHENA works as an integrated system, each of the three subsystems performs its intended role, and together they dramatically reduce the time and effort needed to gather actionable wildfire intelligence. Rather than testing one narrow feature, we evaluated ATHENA the way it would actually be used in the field: as a connected pipeline from hardware, to AI, to operator interface.

The custom PCB successfully powers and communicates with the Jetson Orin Nano in a flight-ready form factor. Bench testing confirmed stable voltage on all regulated rails, clean signal integrity on the high-speed lines, and reliable data exchange over both UART and CAN interfaces. The board survived vibration and thermal stress tests without disconnecting or losing signal, confirming that our design choices, vibration-resistant connectors, wide power traces, and a four-layer stackup with dedicated ground planes, were effective. The board also stayed within our target weight and size envelope, making it practical for real drone integration. One important finding came from our thermal testing: sustained AI workloads caused the Jetson to approach its throttling threshold, which confirmed that active cooling is not optional but essential. This directly shaped our decision to design a custom enclosure with airflow channels.

The onboard AI performed strongly in simulation. Our YOLOv11 model detected smoke, fire, people, and obstacles with high confidence across varied lighting and terrain conditions in Gazebo. The geolocation pipeline successfully projected pixel detections into GPS coordinates using the drone's real-time 6-DOF pose, and the lawnmower autonomy routine achieved full coverage of every test area without manual input. We also found that camera stability mattered more than we initially expected; sharp turns produced motion blur that reduced detection accuracy, which pushed us to add a velocity-aware turning profile so the drone slows down before changing direction. The UDP mesh network shared detection data between simulated drones within milliseconds, and the peer table correctly pruned inactive nodes when they stopped broadcasting, confirming that the swarm stays resilient even if a drone drops out.

The GUI successfully combined live detections, 3D Gaussian Splatting reconstructions, and Canadian Fire Weather Index outputs into a single operator view. The 3D reconstruction gave a much more intuitive sense of terrain and hazard positioning than raw video alone, and the FWI layer added predictive context by showing not just where a fire is, but where it's likely to spread based on current wind and moisture conditions. Operators in our walkthrough tests were able to identify hotspots and interpret risk levels significantly faster with the combined view than with individual data streams.

Overall, ATHENA reduced the time from "drone takeoff" to "actionable information on screen" from roughly 30–60 minutes under manual scouting to under 5 minutes; an order-of-magnitude improvement. These results show that ATHENA delivers on its core promise: faster, smarter, and more coordinated wildfire monitoring at a fraction of the cost of existing systems, and with a pathway to real-world deployment once field testing begins.

Our results show that ATHENA is more than a collection of features; it is a working, integrated platform that addresses the four biggest barriers to effective wildfire monitoring: manual operation, high operator workload, limited coverage, and delayed decision-making. By combining custom hardware, onboard AI, and an intuitive operator interface, ATHENA turns an ordinary drone into a wildfire intelligence system at a small fraction of the cost of manned aircraft or specialized autonomous UAVs.

The most important takeaway is that edge AI and thoughtful systems engineering can compress a process that used to take 30–60 minutes into under 5. That time difference is not just a number. In wildfire response, faster information means earlier evacuations, better-targeted resource deployment, and fewer lives and homes lost. It also means emergency teams can cover far more ground with the same number of people.

We also learned that the hardest part of a system like this is not any single subsystem, but making everything work together reliably. Power stability, thermal management, camera motion, timing synchronization, and network resilience all turned out to matter as much as the AI model itself. Good engineering is integration, not just innovation.

Finally, ATHENA shows that meaningful firefighting technology does not require a massive budget. With a modular, retrofit-friendly design, communities and agencies of any size could benefit. The next step is real-world field testing to validate what we have built in simulation and move ATHENA from prototype to deployable tool.

The most important next step is real-world field testing to validate ATHENA beyond simulation. We plan to optimize the air quality sensor with proper airflow to identify what is burning, and optimize Gaussian Splatting to run efficiently on the operator's local device. Multispectral fusion between RGB and thermal imaging is our next big perception upgrade, since thermal data stays reliable in smoke and darkness. Long-term, we aim to refine the PCB into a patentable design and expand ATHENA into a certified, deployable product for emergency response.

Thank you to the University of Alberta for supplying a space and resources for testing.

[1] DJI. (n.d.). Payload SDK documentation. DJI Developer. https://developer.dji.com/payload-sdk/

[2] DJI. (n.d.). Public safety: Wildfires. DJI Enterprise. https://enterprise.dji.com/public-safety/wildfires

[3] DJI SDK. (2025). Payload-SDK. GitHub. https://github.com/dji-sdk/Payload-SDK

[4] Docker. (n.d.). Docker documentation. https://docs.docker.com/

[5] Data Flair. (n.d.). Python project: Real-time human detection and counting. https://data-flair.training/blogs/python-project-real-time-human-detection-counting/

[6] Fields2Cover. (n.d.). Fields2Cover coverage path planning. https://fields2cover.github.io/index.html

[7] Government of Alberta. (n.d.). Wildfire intelligence. https://www.alberta.ca/wildfire-intelligence

[8] IdeaForge Technology. (n.d.). Eyes in the sky: UAVs redefining forest fire strategies. https://us.ideaforgetech.com/eyes-in-the-sky-uavs-redefining-forest-fire-strategies/

[9] Insideclimatenews. (2023). Wildfire forest regeneration. https://insideclimatenews.org/news/06032023/wildfire-forest-regeneration/

[10] Canadian Forest Service. (n.d.). Canadian wildfire information system: Fire history. Natural Resources Canada. https://cwfis.cfs.nrcan.gc.ca/en/fire-history

[11] Microsoft. (n.d.). Windows Subsystem for Linux documentation. https://learn.microsoft.com/en-us/windows/wsl/

[12] Natural Resources Canada. (n.d.). Canada fire weather index system. https://natural-resources.canada.ca/forests-forestry/wildland-fires/canada-fire-weather-index-system

[13] National Aeronautics and Space Administration. (n.d.). ACERO project: Wildfires description. NASA Aeronautics Research Mission Directorate. https://www.nasa.gov/directorates/armd/aosp/aamp/advanced-capabilities-for-emergency-response-operations/acero-project-wildfires-description/

[14] NVIDIA. (2024). NVIDIA Isaac ROS getting started. https://nvidia-isaac-ros.github.io/getting_started/index.html

[15] NVIDIA. (2024). NVIDIA Isaac ROS Release 3.2. https://nvidia-isaac-ros.github.io/v/release-3.2/

[16] NVIDIA-AI-IoT. (2025). my-jetson-nano-baseboard. GitHub. https://github.com/NVIDIA-AI-IOT/my-jetson-nano-baseboard

[17] Open Robotics. (2024). ROS 2 Humble documentation. https://docs.ros.org/en/humble/index.html

[18] Python Software Foundation. (n.d.). Python 3 documentation. https://docs.python.org/3/

[19] Repo-SAM. (n.d.). 3D Gaussian splatting. INRIA. https://repo-sam.inria.fr/fungraph/3d-gaussian-splatting/

[20] United Nations Office for Disaster Risk Reduction. (2025). Invisible costs: Wildfire disasters. https://www.undrr.org/news/invisible-costs-wildfire-disasters-2025

[21] Wanda Hang, F. Y. (2025). YOLOv11-RGBT. GitHub. https://github.com/wandahangFY/YOLOv11-RGBT

[22] Zenoh. (n.d.). Zenoh: Overview and documentation. https://zenoh.io/docs/overview/what-is-zenoh/

[23] Arana, C., & Salazar, J. (2025). Development of a Jetson Nano drone for monitoring and detection of forest fires. EASI: Engineering and Applied Sciences in Industry. https://revistas.ug.edu.ec/index.php/easi/en/article/view/2674/5529

[24] Wan, D., Lu, R., Fang, Y., Lang, X., Shu, S., Chen, J., Shen, S., Xu, T., & Ye, Z. (2025). YOLOv11-RGBT: Towards a Comprehensive Single-Stage Multispectral Object Detection Framework. arXiv. https://arxiv.org/abs/2506.14696

[25] Tang, J., Gao, Y., Yang, D., Yan, L., Yue, Y., & Yang, Y. (2025). DroneSplat: 3D Gaussian Splatting for Robust 3D Reconstruction from In-the-Wild Drone Imagery. arXiv. (Tang et al., 2025). https://arxiv.org/abs/2503.16964