Methods for Designing, Manufacturing and Controlling Micro-Scale Swimming Robotic Systems Built with DNA Nanotechnology

Small-scale swimming robots have the potential to revolutionize healthcare by enabling targeted drug delivery at nanoscale. But in order to make swimming robots ubiquitous in a clinical setting, we need to develop a commercially viable process for manufacturing custom designs at scale, and develop effective control schemes for these designs. We present a wideranging investigation into the tasks of designing, manufacturing and controlling small-scale swimming robots, highlighting some of the challenges that we face for each function. For each obstacle, we propose a novel solution, frequently combining DNA nanotechnology - a precision material for building nanostructures - with modeling approaches from artificial intelligence, including probabilistic graphical modeling and graph neural networks. We begin by considering the challenge of controlling simplified, millimeter-scale, single link magnetic robots using existing medical infrastructure, such as magnetic resonance imaging (MRI) machines. Next, we increase the complexity of the robot design under consideration to a two-link swimmer, but remain at the millimeter scale, where we use a robust, repeatable experimental platform for defining design rules for building faster swimming robots, while avoiding the confounding variable of polydispersity in off-the-shelf components (such as magnetic microspheres) that is frequently observed at smaller scales. Equipped with these initial investigations at the millimeter scale, we then proceed to incorporate DNA nanotechnology for building swimming robots at the nanoscale. We look for opportunities to streamline the DNA nanotechnology manufacturing process to make it more cost effective, first by accelerating the characterization step used to confirm the formation of the desired structure by reducing the amount of data that must be collected. Next, we propose to speed up DNA nanostructure simulation tools, which are currently only capable of generating predictions of steady state structure configurations over very small time scales (less than one second) by training graph neural networks, which have been shown to be able to quickly generate accurate predictions by using the graph to abstract physical relationships between nucleotides while leveraging efficient hardware implementations. And finally, we propose a novel, cost-effective and flexible approach to assembling DNA nanostructure components together to form a swimming robot, and present a probabilistic graphical model for predicting the yield of such assemblies. Ultimately, this work successively adds complexity in the swimming robot representation, and pursues implementations at increasingly smaller scales, in an effort to one day bring small-scale swimming robots out of the lab and into the clinic.