A Variational Infinite Mixture for Probabilistic Inverse Dynamics Learning

Probabilistic regression techniques in control and robotics applications have\nto fulfill different criteria of data-driven adaptability, computational\nefficiency, scalability to high dimensions, and the capacity to deal with\ndifferent modalities in the data. Classical regressors usually fulfill only a\nsubset of these properties. In this work, we extend seminal work on Bayesian\nnonparametric mixtures and derive an efficient variational Bayes inference\ntechnique for infinite mixtures of probabilistic local polynomial models with\nwell-calibrated certainty quantification. We highlight the model's power in\ncombining data-driven complexity adaptation, fast prediction and the ability to\ndeal with discontinuous functions and heteroscedastic noise. We benchmark this\ntechnique on a range of large real inverse dynamics datasets, showing that the\ninfinite mixture formulation is competitive with classical Local Learning\nmethods and regularizes model complexity by adapting the number of components\nbased on data and without relying on heuristics. Moreover, to showcase the\npracticality of the approach, we use the learned models for online inverse\ndynamics control of a Barrett-WAM manipulator, significantly improving the\ntrajectory tracking performance.\n