Dynamic Resonance Frequency Identification for Energy-Efficient Movement of Legged Microbots

Advances in robotics and manufacturing processes have enabled the development of extremely small robotic devices, even as tiny as a penny. In this domain, legged microrobots (a.k.a. \textit{microbots}) offer numerous potential applications and characteristics to explore. However, controlling such small multi-legged robots presents significant challenges in achieving the desired behavior. Primarily, due to the robot's small size, it can only operate with a tiny battery, therefore, an extremely computationally efficient controller is needed. Tiny robots are also susceptible to damage and control methodologies are needed that also ensure longevity. This thesis work presents a novel approach for creating a control algorithm for a multi-legged system that dynamically identifies and operates a microbot at its resonance frequency of movement. At the resonance frequency, a microbot's leg oscillations achieve maximum amplitude with minimal energy, resulting in optimal locomotion efficiency. After imposing a sinusoidal control for actuating each of the robot's legs, a two-part controller is developed. The controller is then trained using soft actor-critic-based reinforcement learning on a custom model of the mClari microbot. A successful simulation-based demonstration of the learning-based controller is shown by varying the underlying surface (such as wet, soft, etc.) of the microbot where the controller dynamically identifies and switches to the corresponding resonance frequency, achieving efficient and adaptive locomotion.