Terrain Classification Based on Sensed Leg Compliance for Amphibious Crab Robot

Abstract

Characterization of terrain in real time allows autonomous legged robots to modify their gait to better suit their environment and recognize hazardous conditions. We present a novel approach for terrain classification for legged robots with passively compliant components, in which terrain information is gathered by measuring component deflection during operation by means of low cost Hall effect magnetometers and embedded magnets. The effectiveness of this approach is demonstrated on a hexapod robot designed to operate in the surf zones of beaches. Datasets of sensor measurements corresponding to three types of granular terrain are collected in both a laboratory and outdoor beach environment, and are used to train support vector machine (SVM) classifiers. Results show that, using time domain features of measurements from each leg, a mean accuracy of 99.3% and 92.4% through 10-fold cross-validation can be achieved for the laboratory and beach datasets, respectively. For the beach dataset, this accuracy is found to be slightly greater than for a classifier trained on measurements from an on-board inertial measurement unit (IMU) (91.6%), while an accuracy of 95.1% can be achieved by combining the information from both sensor modalities as inputs to a single classifier. The effect of the Earth’s magnetic field on the measurements and considerations for data collection are also discussed.