Prediction of behavior of a noncommunicative robot based on observations of its movement

Abstract

AbstractAn adversary robot can move around in a free space “patrolling an area” appearing to have a random “behavior” if observed through a “lens.” The “behavior” of the robot is associated with its observed/internal state. Stochastic techniques or machine learning techniques, which can be used to predict the future states of an adversary robot demand high computation and thus, power. Advanced RNN techniques like long short‐term memory, which can make predictions using long‐term dependencies of the sequence data, also have extremely high computing and memory requirements. Thus, the systems involved in using these former techniques are not appropriate to be deployed in fields which require inconspicuous monitoring. In this work, we present two novel methods using the “behavior” to predict the future states of an adversary robot which can be deployed using a low computing device/small single board computer/robot in a non‐GPS environment without any internet connection. The proposed two methods are based on a list and a graph, respectively. Unlike the most available literature, where the prediction is based on “time,” our model predicts the future state using the states of the system. These methods work for both single and multiple‐looped paths taken by the adversary robot. The results show that the prediction error which depends on the predicted move and the actual move taken by the robot decreases as a logarithmic function and is comparatively lower when the adversary robot follows a path having multiple loops.