Body State Recognition for a Quadruped Mobile Robot

The body states must be tracked by the onboard software on the robot to make good decisions. A human can pick up this machine or if the robot encounters anomalies (e.g. fall over) during locomotion, the state changes must be identified to execute the necessary responses. The authors of this paper developed a machine learning model which can recognize four states (normal, pick-up, fall over, poked) of a Sony AIBO robot. A deep neural network classifier with these predictors achieved 98% accuracy on unseen data and actual test runs on the robot proved the practical use with real-time execution speed. These properties made the proposed method a good candidate for adaption to other legged robots.