A Step Toward Learning to Control Tens of Optically Actuated Microrobots in Three Dimensions

Automated manipulation of a large number of independently controlled microrobots in three dimensions can potentially lead to a major breakthrough in the formation of functional microstructures, both engineering and biological, in a scalable and reliable manner. In this paper, we provide the first foundational step toward realizing this objective by developing a reinforcement learning-based model predictive controller. Our controller (approximately) optimizes the trajectories of the microrobots through importance sampling, while ensuring that the trajectories do not cause collisions with workspace obstacles, such as other randomly dispersed microobjects or already-formed microstructures. Simulation experiments with tens of microspheres, actuated using holographic optical traps, demonstrate the feasibility and usefulness of our method.