Detecting static and dynamic novelties using dynamic neural network

Abstract

This paper presents a dynamic neural network based novelty filter where a mobile robot explored in an environment and built a dynamic model of the robot normal sensory-motor values perceived from the environment. Afterwards, the acquired model of normality was used on the robot to predict expected values of the sensory-motor inputs during the patrolling. Novelties could be detected if the prediction error between model-predicted values and real observed values exceeded a certain novelty threshold. Because of high uncertainty while robot interactions with the environment, the network-prediction errors along the robot route were not normally distributed to define only one novelty threshold. Therefore region-specific novelty thresholds were estimated by the proposed novelty detection system. As a result, the robot is capable of selecting a local novelty threshold depending on where the robot currently occupied. To evaluate the proposed system, a set of real-world robotic experiments were carried out. Experimental results showed that the novelty filter was able to highlight unusual static and dynamic objects in the environment. Furthermore, the filter also produced reliable local novelty thresholds while the robot patrolled in the noisy environment.