Structured learning of rigid-body dynamics: A survey and unified view\n from a robotics perspective

摘要

Accurate models of mechanical system dynamics are often critical for\nmodel-based control and reinforcement learning. Fully data-driven dynamics\nmodels promise to ease the process of modeling and analysis, but require\nconsiderable amounts of data for training and often do not generalize well to\nunseen parts of the state space. Combining data-driven modelling with prior\nanalytical knowledge is an attractive alternative as the inclusion of\nstructural knowledge into a regression model improves the model's data\nefficiency and physical integrity. In this article, we survey supervised\nregression models that combine rigid-body mechanics with data-driven modelling\ntechniques. We analyze the different latent functions (such as kinetic energy\nor dissipative forces) and operators (such as differential operators and\nprojection matrices) underlying common descriptions of rigid-body mechanics.\nBased on this analysis, we provide a unified view on the combination of\ndata-driven regression models, such as neural networks and Gaussian processes,\nwith analytical model priors. Further, we review and discuss key techniques for\ndesigning structured models such as automatic differentiation.\n