Learning-based control and localization of magnetic soft millirobots

Abstract

Soft millirobots have promising biomedical applications due to mechanical compliance, absorbing excess forces without additional computational effort, and multifunctionality. Especially with wireless multimodal locomotion capabilities, magnetic soft millirobots (i.e., ≤1 cm) have emerged as potential minimally invasive medical robotic platforms as they can access confined and hard-to-reach spaces in the human body (e.g., distal vascular regions), and carry out medical applications, such as on-demand drug delivery, sensing, and embolization, in a target location. For such potential biomedical applications, the adaptivity of the robot control is essential to ensure the continuity of the operations, as task environment conditions show dynamic variations that can alter the robot's performance. However, fabrication-, material-, physical-interaction-dependent variations, and complex kinematics with virtually infinite degrees of freedom arising from the nature of their soft material structure limit the applicability of the conventional modeling and control methods. The main objective of this dissertation is to establish a data-efficient adaptive multimodal locomotion framework for the targeted application scenarios of magnetic soft millirobots. To this end, a probabilistic learning approach leveraging Bayesian optimization (BO) and Gaussian processes (GPs) is introduced to address the controller adaptation challenge. First, the efficacy of the BO to fabrication variabilities is shown on three different robots fabricated following the same steps. Next, through augmented tests on benchmark datasets, it is shown that transferring the posterior mean learned by one robot as the prior mean to the other robots and test cases improves the learning performance of BO by achieving quicker gait adaptation. Afterward, the controller adaptation method employing the proposed transfer learning approach is demonstrated in various task spaces with varying surface adhesion, surface roughness, and medium viscosity properties. To further improve the adaptation performance by including multimodal locomotion, the sim-to-real transfer learning method is developed in the third study. In this regard, a data-driven simulation environment is designed, and its accuracy is demonstrated by comparing the simulated results to the physical experiments. Leveraging the simulated experience and BO based transfer learning, it is demonstrated that sim-to-real transfer learning provides efficient locomotion learning. Furthermore, the adequacy of the automated locomotion adaptation through the Kullback-Leibler divergence-based domain recognition approach is shown to changing environmental conditions. As the secondary objective, a new localization method using electrical impedance tomography is introduced. The applicability of the proposed approach is demonstrated for stationary and moving cases in environments with and without any obstacles. With these contributions, this thesis proposes a domain-adaptive locomotion learning framework enabling the soft millirobot locomotion to quickly and continuously adapt to environmental changes while exploring actuation space.