FOP Networks for Learning Humanoid Body Schema and Dynamics

Robot inverse dynamics modeling is performed mainly via standard system identification and/or machine learning techniques. In this paper we part from the theoretical framework of First-Order Principles Networks (FOPnet), combining data-aided learning with basic knowledge to learn the model of a targeted robot. The framework, previously used for learning the dynamics of a fixed-base serial manipulator, is now extended to the learning of the kinematics and dynamics of tree-structured robots with floating base. Our approach leverages the principle of compositionality to separate the main problem into two partially independent modules. The first defines the robot's body schema by characterizing its morphology and topology. The second is dependent upon the latter and defines the inertial properties of the multi-body system. To demonstrate the capabilities of the approach, a simulated humanoid robot with 30 degrees of freedom is used. We discuss the implementation of our method and evaluate its estimation and generalization capabilities in comparison with other common machine learning approaches. Finally, we present experimental results on a 7- DoF manipulator.