Anomaly detection with robotics

Sammanfattning

The aim of this paper is to compare anomaly detection algorithms in robotic environments and develop a real time solution. The results of the simulated and real robot experiments were used in order to choose the most suitable algorithm in this context, were various types of anomalies were introduced. The algorithms were selected between a fixed set of candidates based in an initial study and attending to their performance on the experiments considering quantitative metrics results, as well as a qualitative analysis of outputs contrasted with the expected results. The selected algorithm was PaDiM, as its adaptability allowed the detection of different types of anomalies both in simulated an real scenarios. The final real time system uses camera and a pre-trained PaDiM model to detect the presence of anomalies that may present a safety risk to the corresponding robotic operation.