Self-Supervised Task Learning For Robotic Underfloor Insulation

Abstract

Effective under-floor insulation (UFI) of residential buildings to reduce their energy consumption and CO2 emissions is a substantive challenge in retrofitting existing homes. Traditional UFI installation techniques require suspended floors to be taken up, damaging and disrupting the living space for several days. To deliver the low-disruption insulation of suspended floors, a robot has been developed for accessing and spraying thermal insulation to the underside of a suspended floor. Q-Bot has been installing UFI with a fleet of robots in a largely teleoperated mode for several years. In this mode, the operator is in complete control of all actions the robot takes, which is a significant cognitive burden. This paper addresses Q-Bot's steps toward automating the UFI installation process, reducing the operator's cognitive load and eventually freeing the operator to perform other tasks. Years of recorded experience are leveraged to train a simplified U-Net model in a self-supervised fashion, enabling robots to decide where to apply the insulation foam next. Results obtained from the on-site collected data show that the weighted symmetric cross-entropy loss function yields better spray-region prediction results than the base loss, Cross-Entropy. Our method can adapt to various operator preferences, generalise to novel building crawl spaces, and improve with more data.