Adaptive Magnetic Control using Stein Variational Gradient Descent computed Distribution of Object Parameters

Published: 11 Sept 2023, Last Modified: 11 Sept 2023DiffPropRob IROS 2023 PosterEveryoneRevisionsBibTeX
Keywords: Variational Inference, Adaptive Control, Stein Variational Gradient Descent, Optimization, Magnetic manipulation, eddy currents
TL;DR: Fitting a distribution over plausible dynamic parameters improves the robustness to stochastic dynamics compared to fitting single best dynamic parameter for nonmagnetic object magnetic manipulation.
Abstract: This paper extends recent work in magnetic manipulation of conductive, unknown nonmagnetic objects. Previous work shows six degrees of freedom manipulation for unknown conductive spheres by adaptively fitting the best current object parameters for use in control. In this work, we expand upon the object parameter by iteratively fitting a distribution over object parameters using Stein Variational Gradient Descent. Then we improve the control optimization to find the best control cost in expectation under the object parameter distribution. We show this controller is much more robust to noise and varying dynamics than single best parameter adaptive control.
Submission Number: 7