Application of Large Language Models in Magnetically Manipulated Microrobots

Abstract

Large language models (LLMs), machine learning systems trained on massive internet-scale datasets, have recently been applied across diverse domains-including robotics-due to their ability to generalize, reason over complex data, and adapt to a wide range of tasks with minimal fine-tuning. In robotics, LLMs have enabled advancements in planning, instruction following, and high-level decision-making. However, their potential in microrobotic systems, such as magnetically manipulated microrobots, remains largely unexplored. Compared to macroscale robotics, this domain poses unique challenges, including complex nonlinear dynamics, sparse observations, and limited data availability, while offering diverse applications ranging from biomedical solutions to environmental cleaning. In this work, we present the first demonstration of LLMs applied to physical microrobotic control by automating the design of reward functions within a reinforcement learning (RL) framework. We train an RL policy using LLM-generated rewards and deploy it to control the motion of a ferromagnetic particle (<tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\varnothing \sim \mathbf{5 0 0} \mu \mathrm{m}$</tex>) using solenoids at the air-water interface. Furthermore, we evaluate the impact of LLM model scale, prompt design, and reward configuration on learning performance. The results show that LLM-based reward design enables effective training and experimental deployment, opening a new direction for LLM-driven control in microrobotics.