Learning the prediction error for improving an analytical-based prediction (object-model) system for manipulation tasks

Abstract

One of the main tasks in robotics today, is to bring robots closer to humans in everyday situations. This requires the robot to understand how its environment (objects, people, conditions) behaves. One method that tries to connect the environment to the robot is called object model. This proposed system (object model) is able to give the robot an understanding of the physics of the environment. Object models have been used to give robots the ability to understand and control object behavior. This information helps robots to be more capable for skilled manipulation tasks, by predicting how the object will react to external stimulus. The object model used as case of study in this paper, uses an analytical representation for describing object behavior. This analytical representation has the advantage of using meaningful object properties and quickly allowing the robot to manipulate the object without doing a lot of trial and error repetitions. A challenge of this approach is that it can be very difficult to derive a mathematical/mechanical model of the object behavior. Therefore, this model, in most cases, will not describe all the peculiarities and details of object behavior. As a result, predictions are good but not perfect. This paper proposes a method to improve the prediction performance of such system, by learning the error of the analytical model and using this to correct the original prediction. Our results show that such a system is able to improve the prediction performance of the system. A quantitative evaluation using cross validation is provided to demonstrate the ability of our system to reduce the error exhibited by the prediction system (object model).