Researchers led by Anthony Czubarow, with Raffaello D'Andrea among the authors, have used deep model-based reinforcement learning to shape a ferrofluid, a liquid that contains magnetic nanoparticles, according to a paper on arXiv. Such fluids could serve as soft robot bodies that move through tight spaces and handle objects.
The learned policies set the current of a single fixed electromagnet from camera images, steering the moving fluid towards shapes that would not hold still on their own. Trained directly on the physical setup, they repeatedly formed polygons, asterisks and pentominoes.



