Premium
Towards autonomous inspection of concrete deterioration in sewers with legged robots
Author(s) -
Kolvenbach Hendrik,
Wisth David,
Buchanan Russell,
Valsecchi Giorgio,
Grandia Ruben,
Fallon Maurice,
Hutter Marco
Publication year - 2020
Publication title -
journal of field robotics
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 1.152
H-Index - 96
eISSN - 1556-4967
pISSN - 1556-4959
DOI - 10.1002/rob.21964
Subject(s) - sanitary sewer , robot , ground truth , support vector machine , artificial intelligence , visualization , software deployment , computer science , set (abstract data type) , engineering , visual inspection , computer vision , waste management , programming language , operating system
The regular inspection of sewer systems is essential to assess the level of degradation and to plan maintenance work. Currently, human inspectors must walk through sewers and use their sense of touch to inspect the roughness of the floor and check for cracks. The sense of touch is used since the floor is often covered by (waste) water and biofilm, which renders visual inspection very challenging. In this paper, we demonstrate a robotic inspection system which evaluates concrete deterioration using tactile interaction. We deployed the quadruped robot ANYmal in the sewers of Zurich and commanded it using shared autonomy for several such missions. The inspection itself is realized via a well‐defined scratching motion using one of the limbs on the sewer floor. Inertial and force/torque sensors embedded within specially designed feet captured the resulting vibrations. A pretrained support vector machine (SVM) is evaluated to assess the state of the concrete. The results of the classification are then displayed in a three‐dimensional map recorded by the robot for easy visualization and assessment. To train the SVM we recorded 625 samples with ground truth labels provided by professional sewer inspectors. We make this data set publicly available. We achieved deterioration level estimates within three classes of more than 92% accuracy. During the four deployment missions, we covered a total distance of 300 m and acquired 130 inspection samples.