Accepted in ICRAS 2026 Sensor Fusion • Gas Leak Detection

Gazzard: Gas Leak Detection using a Mobile Robotic Platform

Dulce Maria Yadao¹ • Mark Kenneth Biasbas¹ • Faustino Miguel Flores¹ • Carl Christian Gatchalian¹ • Lorin Angela Velasco¹ • Ma. Madecheen Pangaliman¹ • Anthony James Bautista²

¹Department of Electronics Engineering, ²Department of Mechanical Engineering
University of Santo Tomas, Manila, Philippines

🥇 Champion • FUSERS 2026 Video Presentation 🏅 People's Choice • ECE Open Lab Exhibit 🎤 Presenter • S&T Undergraduate Research Symposium 📌 Participant • 2026 NRCP Annual Scientific Conference & 93rd GMA Poster Exhibit

📄 Paper Coming soon 💻 Code 🗂️ Dataset 📋 BibTeX

📄 Abstract

In enclosed industrial facilities, CO₂ pipe leaks pose safety risks because the gas is colorless and odorless, allowing it to accumulate undetected in confined spaces. While various robotic systems have been deployed to automate gas leak detection, they face limitations: airflow patterns can mislead gas sensors, and vision-based systems struggle to detect invisible CO₂ plumes or confirm whether identified cracks are actively leaking. This study integrates a Senseair S8 NDIR CO₂ sensor with a zero-shot YOLO-World object detection model on a TurtleBot3 Waffle Pi platform running ROS 2 Humble, combining gas concentration monitoring with visual crack detection and Depth-Anything-V2 depth estimation to verify leak sources and reduce false alarms. Across four evaluation phases, ROS 2 communication latency averaged 24.23 ms, sensor percentage errors stayed within the ±3% tolerance (1.09–1.49%), and the detection model reached an mAP@50–95 of 0.8692 for pipe crack identification. Fusing visual detection with CO₂ concentration data increased leak identification accuracy from 46.67% to 90% and reduced false positives from 100% to 6.67%, showing that threshold-based sensor fusion is needed for reliable leak source confirmation.

CO₂ gas leak detection sensor fusion zero-shot detection YOLO-World NDIR sensor mobile robot

🤖 System

Gazzard is a ROS 2 Humble workspace running on a TurtleBot3 Waffle Pi. A distributed publisher–subscriber graph streams camera frames, CO₂ readings, and timing data between the robot and an onboard laptop, where YOLO-World and Depth-Anything-V2 run the detection pipeline.

Four-phase research process flow — Research process flow: ROS 2 configuration, gas sensor validation, object detection deployment, and gas leak source identification.

ROS 2 publisher and subscriber communication flow — ROS 2 communication flow between the publisher (robot) and subscriber (laptop) nodes.

Hardware

Robot platformTurtleBot3 Waffle Pi
CO₂ sensorSenseair S8 (NDIR, UART)
Sensor range400–10,000 ppm, ±40 ppm ±3%
CameraRaspberry Pi Camera v2.1 (8MP)
Onboard OSUbuntu 22.04 + ROS 2 Humble
Link5 GHz Wi-Fi

Software Stack

DetectionYOLO-World XL (zero-shot)
DepthDepth-Anything-V2
Class promptsdummy / pvc pipe / paper crack
Fusion ruleCrack + elevated CO₂ → leak source
ROS 2 latency24.23 ms avg (distributed)
UART response21.81 ms avg

📸 System in Action

Demo video — Gazzard locating and confirming a CO₂ leak source.

Operational GUI: detection, CO2 reading, depth map — Operational GUI: live camera feed with detection parameters and a *LEAK SOURCE FOUND* banner on the left, depth map on the right.

Experimental setup for evaluating the gas sensor — Gas sensor evaluation setup: varying dry ice mass at fixed distance, then varying distance at fixed mass.

Experimental setup for detecting the gas leak source — Leak source setup with PVC crack, paper-core substitute crack, and dummy crack.

📊 Key Results

Evaluated over 30 controlled trials (15 leak / 15 no-leak) using dry ice to simulate CO₂ leaks.

0.869

mAP@50–95
crack detection

90%

Leak ID accuracy
fused (from 46.67%)

6.67%

False positives
down from 100%

24.23 ms

ROS 2 latency
distributed nodes

Sensor Fusion vs. Detection Only

Method	Accuracy	Precision	Recall	FP rate
Detection only	46.67%	48.28%	93.33%	100%
Detection + CO₂	90.00%	92.86%	86.67%	6.67%

Vision alone flags every crack as a leak (100% false positives) because CO₂ is invisible. Adding the CO₂ threshold check recovers 14 of 15 true negatives. Wilson 95% CI for fused accuracy: 74.4–96.5% (n=30).

Per-Class Detection

Crack type	Substrate	mAP@50–95
Dummy crack	Smooth PVC	0.9560
PVC pipe crack	PVC pipe	0.9116
Paper core crack	Porous paper	0.7400

Inference Latency (CPU)

Component	Mean ± Std
YOLO-World XL	892.37 ± 60.61 ms
Depth-Anything-V2	1028.47 ± 66.53 ms
Total pipeline	1920.84 ± 100.58 ms
Throughput	0.52 ± 0.03 FPS

Zero-shot YOLO-World XL averages mAP@50–95 of 0.8692 across crack types; accuracy tracks substrate texture. Depth estimation, not detection, dominates pipeline latency — GPU deployment is recommended for real-time operation.

📝 BibTeX

If you find this work useful in your research, please consider citing:

@inproceedings{yadao2026gazzard, title = {Gazzard: Gas Leak Detection using a Mobile Robotic Platform}, author = {Yadao, Dulce Maria and Biasbas, Mark Kenneth and Flores, Faustino Miguel and Gatchalian, Carl Christian and Velasco, Lorin Angela and Pangaliman, Ma. Madecheen and Bautista, Anthony James}, booktitle = {Proc. International Conference on Robotics and Automation Sciences (ICRAS)}, year = {2026} }

🙏 Acknowledgment

This research was conducted at the Department of Electronics Engineering, Faculty of Engineering, University of Santo Tomas. We gratefully acknowledge and sincerely thank the SafeGuard project of Dr. Anthony James Bautista for funding this research and making this work possible.