Advantages and Limitations of Reservoir Computing on Model Learning for Robot Control

In certain cases analytical derivation of physicsbased models of robots is difficult or even impossible. A potential workaround is the approximation of robot models from sensor data-streams employing machine learning approaches. In this paper, the inverse dynamics models are learned by employing a learning algorithm, introduced in [1], which is based on reservoir computing in conjunction with self-organized learning and Bayesian inference. The algorithm is evaluated and compared to other state of the art algorithms in terms of generalization ability, convergence and adaptability using five datasets gathered from four robots in order to investigate its pros and cons. Results show that the proposed algorithm can adapt in real-time changes of the inverse dynamics model significantly better than the other state of the art algorithms.