Teaching Autonomous Soft Machines to Swim

Self-oscillating gels are polymeric materials that change shape, driven by chemical reactions occurring entirely within the gel. The research team will develop a computational and machine learning program to discover how to configure self-oscillating gels so that they undergo deformations that result in swimming. The long term goal is to develop a general framework for controlling autonomous soft machines.

Caption: Chemical waves and elastic deformations in thin gel sheets. (a & b) Target and spiral waves in the BZ (Belousov-Zhabotinsky) chemical reaction. (c) Numerical simulations of the BZ reaction. (d) Resulting reference metrics and shapes for the gel sheet.(e) Experimental images of oscillating gels.

The graceful swimming of a jellyfish, the locomotion of a snail, and the beating of our hearts are inspiring examples of machines made from soft materials. While examples abound in nature, we lack the technology to fabricate our own soft machines. The great promise of soft machines is their ability to assume an infinite variety of shapes and produce large deformations. To truly exploit that potential, we need to develop control and actuation schemes for their many degrees of freedom.

Self-oscillating gels are polymeric materials that change shape, driven by chemical reactions occurring entirely within the gel. The research team will develop a computational and machine learning program to discover how to configure self-oscillating gels so that they undergo deformations that result in swimming. The long term goal is to develop a general framework for controlling autonomous soft machines.

The project team’s process will consist of first numerically solving a model for the dynamics in the forward direction: given initial data (parameters, initial conditions, etc.), compute the time- and space-dependent deformation. Next, deep learning will be used to identify low-dimensional structure in the gel dynamics and to predict initial data for given deformation fields. Lastly, the learned model will predict the initial data for a target deformation and the predictions will be validated against the PDE model and experiments.