Robot Health Estimation through Unsupervised Anomaly Detection using Gaussian Mixture Models

Abstract

Autonomous robots in complex environments are usually forced to act very conservatively, greatly limiting their potential. Taking more risky actions confidently requires the robot to have a deep understanding of its abilities, especially in its current state. The foundation for such a self-awareness is knowledge about current damages and the stress the different components of the robot are under. While the skills of a robot can be modeled in advance, the potential errors that might occur cannot easily be predicted exhaustively. Due to this, the robot is required to notice unforeseen changes in itself and judge their severity. This work presents a solution for this in the form of a Gaussian Mixture Model based framework for anomaly detection. The model requires only training data for a healthy robot, with no samples needed for expected problems and is able to correctly notice, localize and quantity various introduced damages and impairments. Transfer to new robots requires a user to only specify available sensor data for the robot's different components. It was implemented and tested on two different robots sharing no hardware, with different problems introduced into both systems. This approach lays the foundation for a general framework for adaptive self-aware robot decision making and planning.