Researchers from the National University of Singapore have used a large language model (LLM), GPT-4, to train microrobots to swim in viscous fluids. Zhuoqun Xu and Lailai Zhu used a few-shot learning approach and a minimal unified prompt of five sentences to guide two distinct microrobots, the three-link swimmer and the three-sphere swimmer. The LLM-based decision-making strategy outperformed traditional reinforcement learning methods in training speed. The study highlights the potential of LLMs in deducing physical principles underlying robotic locomotion and designing locomotory gaits, particularly in microrobotic swimming.
Microrobots Trained to Swim Using Large Language Model
Researchers Zhuoqun Xu and Lailai Zhu from the Department of Mechanical Engineering at the National University of Singapore have utilized a large language model (LLM), GPT-4, to train two prototypical microrobots for swimming in viscous fluids. The researchers adopted a few-shot learning approach and developed a minimal unified prompt composed of only five sentences. This concise prompt successfully guided two distinct articulated microrobots, the three-link swimmer and the three-sphere swimmer, in mastering their signature strokes.
The Role of Machine Learning and Artificial Intelligence in Robotic Systems
Machine learning and artificial intelligence have recently become a popular paradigm for designing and optimizing robotic systems across various scales. Recent studies have showcased the innovative application of large language models (LLMs) in industrial control and in directing legged walking robots. In this study, the researchers utilized an LLM, GPT-4, to train two prototypical microrobots for swimming in viscous fluids. The LLM-based decision-making strategy substantially surpassed a traditional reinforcement learning method in terms of training speed.
The Influence of Biological Strategies on Robotic Movement
A fundamental characteristic of living organisms is their capacity for locomotory motion, encompassing a spectrum of movements such as running, crawling, flying, slithering, and swimming. These motility patterns, evolved through the long process of natural selection and adaptation, inspire the design of bionic articulated robots. They achieve locomotion by actuating their movable components, similar to natural joints or hinges and links, such as walking legs, flapping wings, and undulating fins.
The Use of Large Language Models in Robotic Control
The recent surge in large language models (LLMs) or foundation models in general has spawned a new machine learning paradigm, in-context learning (ICL), for decision-making and robotic manipulation. Unlike reinforcement learning or other machine learning techniques, ICL precludes updating neural network weights, instead learns to solve new tasks during the inference phase via receiving text prompts that incorporate exemplar task demonstrations. GPT-4, an LLM pre-trained on internet-scale datasets, has been successfully utilized for controlling HVAC and for enabling terrestrial robots to walk in simulated environments.
The Application of LLMs in Microrobotic Swimming
In this work, the researchers explored the ability of LLMs to deduce physical principles underlying robotic locomotion and subsequently leverage the understanding to design locomotory gaits. The specific focus was on microrobotic swimming at a vanishing Reynolds (Re) number, a scenario physically constrained by the dominance of viscous fluid forces over inertia forces. The researchers explored the LLM, GPT-4, to navigate Purcell’s three-link swimmer and another prototypical model, Najafi-Golestanian (NG’s) three-sphere swimmer, in inertialess viscous fluids.
The Design of a Unified Minimal Prompt Framework
The researchers designed a unified minimal prompt framework to coordinate the interaction between GPT-4 and both swimmers. Despite its training lacking physical data, GPT-4 enabled the microswimmers to learn efficient strokes, thereby overcoming the low-Reynolds-number physical constraints.
Viscous Hydrodynamic Environments for Microswimmers
The Purcell’s swimmer comprises three slender cylindrical links of length a and radius b connected in sequence by two planar hinges. These hinges allow the links to rotate relative to their neighbors. Propulsion is attainable by varying over time, the relative angles between every two adjacent links. In contrast, NG’s swimmer constitutes three spheres with radius a aligned colinearly along their common axis. The spheres are linked sequentially by two extensible links with lengths that vary over time.
In the article titled “Training microrobots to swim by a large language model“, authors Zhuoqun Xu and Lailai Zhu discuss their innovative research on training microrobots. Published on January 22, 2024, the study explores the potential of large language models in the field of microrobotics.
