Anomaly Behavior Detection in Mobile Robot Navigation for Repetitive Tasks Using CNN-LSTM Autoencoder Reconstruction

The reliability of mobile robotic systems is important for the safety and efficiency of repetitive tasks performed in dynamic environments. However, mechanical faults or environmental changes cause these dynamic systems to suffer from anomalies, which translates into degradation of navigation. Herein, an entirely new approach is proposed on how to detect anomaly behavior during mobile robot navigation using a Convolution Neural Network - Long Short Term Memory (CNNLSTM) autoencoder network. The model was trained with sensor data from robot systems (Inertial Measurement Unit (IMU), odometry and velocity commands), used to reconstruct normal navigation behavior, and then evaluated for anomaly detection. The system was tested on simulation data, where different types of anomaly behaviors, such as wheel slipping and unexpected turning, were introduced. The results proved that the proposed method works effectively for the identification of anomalies, also gave information on the root cause, and achieved a considerably promising F1 score of $\mathbf{0 . 7 6 2 4}$. This paves the way for fault analysis in real time for the enhancement of robotic performance, even in high-demand environments such as warehouses and factories.