Online Multi-task Learning of Robotic Manipulators Dynamics Models with Echo-State Networks
Dynamics modeling is crucial in the field of robotics for achieving fundamental operations such as torque-based control, operational-space control and trajectory simulation. Although there exist analytical physics-based approaches, their derivation can be problematic since they require knowledge about the manipulators’ physical properties, which can be hard or even impossible to derive. A machine learning alternative is to employ regression approaches for approximating dynamics models exploiting only the sensors’ data-streams. Therefore, in this paper is presented a model learning approach, based on Echo-State Networks which can be applied online for learning both inverse and forward dynamics models solely from data. The method exploits the abilities of self-organization, Echo-State Networks and multivariate Bayesian regression. The proposed approach has been implemented within a computed-torque controller on a physical robotic system in order to achieve online learning of inverse dynamics models and used for executing trajectories. Moreover, it has been evaluated on datasets captured from multiple robotic manipulators and has been compared with other machine learning algorithms. The results – in the case of inverse dynamics learning – exhibit state-of-the-art performance on generalization ability, prediction accuracy and convergence while the proposed method outperforms the state-of-the-art in terms of adaptability. In case of forward dynamics modeling it outperforms the prediction accuracy of the state-of-the-art algorithms both on short and long-term predictions.