Researchers from Santa Clara University, New Jersey Institute of Technology and the University of Hong Kong have been able to successfully teach small robots how to swim through deep reinforcement learning, marking a major leap in the evolution of precision swimming ability.
There has been tremendous interest in developing artificial micro-swimmers that can navigate the world similarly to naturally occurring microorganisms, such as bacteria. These microswims offer promise for a wide range of biomedical applications in the future, such as drug delivery and microsurgery. However, most synthetic micro-swimmers to date can only perform relatively simple maneuvers using kinematic, steady gaits.
In a study published by researchers in Communication PhysicsThey saw that young swimmers could learn – and adapt to changing conditions – through artificial intelligence. Much like human learning to swim requires reinforcement learning and feedback to stay afloat and propel in different directions under changing conditions, so must young swimmers, despite the unique set of challenges posed by physics in the microscopic world.
“Being able to swim on a small scale is in itself a challenging task,” said En Sean Buck, assistant professor of mechanical engineering at Santa Clara University. “When you want a precision swimmer to perform more complex maneuvers, designing their kinematic gaits can quickly become difficult.”
By combining artificial neural networks and reinforcement learning, the team successfully taught a simple young swimmer to swim and navigate toward any arbitrary direction. When a swimmer moves in certain ways, he receives feedback on how well a particular action is. Then the swimmer gradually learns how to swim based on his experiences interacting with the surrounding environment.
said Alan Tsang, associate professor of mechanical engineering at the University of Hong Kong. “It does this without relying on human knowledge but only on a machine learning algorithm.”
The AI-powered swimmer is able to adaptively switch between different kinetic gaits to navigate toward any target location on their own.
As evidence of the swimmer’s strong ability, researchers have shown that it can follow a complex path without being explicitly programmed. They also demonstrated the strong performance of the swimmer in navigating under turbulence arising from external fluid flows.
“This is our first step in meeting the challenge of developing microswims that can adapt like biological cells to navigate complex environments independently,” said Yuan Nan-yong, professor of mathematical sciences at the New Jersey Institute of Technology.
Such adaptive behaviors are essential for future biomedical applications of artificial micro-swimming in complex media with uncontrolled and unpredictable environmental factors.
said Arnold Mathiesen, an expert in micro-robotics and biophysics at the University of Pennsylvania, who was not involved in the research. “The integration of machine learning and precision swimming in this work will lead to further links between these two very active areas of research.”
Alternating walking and target navigation for young swimmers through deep reinforcement learning
The date the article was published
June 21, 2022
Disclaimer: AAAS and EurekAlert! Not responsible for the accuracy of newsletters sent to EurekAlert! Through the contributing institutions or for the use of any information through the EurekAlert system.