Year: 2025
Type: dissertation
Abstract: Magnetically manipulated microrobots are traditionally controlled through methods such as open-loop controllers, closed-loop controllers, reinforcement learning (RL), and deep learning (DL). The advent of foundation models (FMs), particularly those based on transformer architectures, has significantly advanced various domains, especially natural language processing (NLP). These models have recently been explored for applications in robotic control and planning, leveraging their generalization capabilities to enhance performance across diverse tasks. However, these techniques have not yet been widely employed in micro robotics. In this thesis, we investigate the integration of FM-based methodologies into the control of magnetically manipulated microrobots. We focus on a novel approach that employs large language models (LLMs) as automated reward function designers within an RL framework. To facilitate this, we develop a simulator that accurately captures the system's dynamics using DL techniques. Subsequently, we train an RL policy within this simulated environment, utilizing the LLM-generated reward functions. Our results demonstrate that this policy effectively controls microrobots in both simulated scenarios and real-world magnetic manipulation setups. This research underscores the potential of integrating LLMs into RL frameworks to automate reward function resulting in a manipulation and environment agnostic method. Our findings suggest that FMs can provide highly generalized solutions in micro robotic control and planning tasks.
Source: Aaltodoc (Aalto University)
Author: Sattar, Usama
Language: English
Topic: Micro and Nano Robotics
Subfield: Condensed Matter Physics
Field: Physics and Astronomy
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7035962302