FireJet : a coupled visual perception system for water-jet guidance in industrial firefighting robots

Abstract

Industrial fires often involve flammable materials, high-pressure equipment, and complex spatial layouts, posing significant risks of fire escalation, explosion, and toxic release. Effective and timely suppression is therefore critical to preventing major accident scenarios and mitigating consequential losses. Although robotic firefighting systems offer a safer alternative to manual intervention in hazardous environments, their performance is fundamentally constrained by the accuracy and consistency of visual perception. Existing approaches typically treat fire detection, flame segmentation, and water-jet landing-point localization as independent tasks. This separation limits the consistency of the visual feedback available for downstream spray-direction adjustment To address this gap, this study proposes FireJet, a coupled visual perception system that integrates these three tasks in a unified video-processing architecture. FireJet uses an edge–cloud design in which YOLO-Fire performs lightweight fire detection on the robot, while first-frame-prompted SAM2 segmentation and water-jet landing-point localization run in the cloud. The coupled outputs provide spatial feedback that can be used by a downstream spray controller Experiments on DetectiumFire and the evaluated industrial-like scenarios showed consistent performance across the three perception tasks. These results support FireJet as an engineering-oriented perception module, while closed-loop targeting and suppression efficiency remain to be validated. Our project homepage is https://widemountfirejet.github.io/.