Researchers at Leipzig University have achieved a success in microrobotics. They were able to show for the first time that tiny, synthetic microswimmers perceive their environment directly via their own body shape and adapt autonomously in strongly changing fluid currents. The work now published in Science Advances thus establishes a new paradigm for autonomous microsystems whose control functions in difficult environments where conventional sensors fail. This opens up new perspectives for autonomous medical microrobots, for example for targeted drug delivery in the bloodstream.
The research team, led by Frank Cichos from the Molecular Nanophotonics Group at the Peter Debye Institute for Soft Matter Physics at Leipzig University, used reinforcement learning, a machine learning approach, to control microswimmers that navigated through complex flow fields. The microscopic particles learned successful navigation strategies with the help of algorithms, even though they had no direct sensory information about currents that counteracted their movement. As every movement of the particles already carried the signature of the flow, their bodies themselves served as sensors and thus as a data basis for the algorithm.
Professor Cichos emphasizes the overarching significance: "This work builds a bridge between biological inspiration and practical implementation. Motile microorganisms have evolved over millions of years to use their physical constitution for navigation. We now show that machine learning can discover similar strategies in synthetic systems within experimentally feasible timescales."
Physics becomes a learning and decision-making system for microswimmers
The researchers combine melamine particles coated with gold nanoparticles - so-called synthetic microswimmers (with a radius of about 1 micrometer) - with real-time optical control and machine learning algorithms. The particles are propelled by asymmetric laser heating. During the training phases, the particles learn to reach their destinations despite hydrodynamic disturbances caused by laser-induced currents.
"The experiments themselves were quite challenging," notes Dr. Diptabrata Paul, research associate at the Peter Debye Institute. "We had to achieve stable real-time control and train the learning algorithm at the same time - essentially we taught the microswimmers how to behave during navigation. The particles are exposed to currents that are up to four times stronger than their own propulsion speed, yet they learn to navigate successfully within around 50 training sessions."
The key insight lies in what researchers call "embodied intelligence" - the principle that physical structures and interactions with the environment can serve as computational resources for the algorithms. Instead of relying on miniaturized sensors and processors, the motion dynamics of the microswimmers themselves become information processors.
"This is fundamentally different from our usual idea of robot design," explains Paul. "Instead of trying to capture everything explicitly via sensors and then calculate reactions, the physical interaction between the body and its environment is used to obtain the required information. The learning algorithm discovers how to read and react to this embodied information."
Autonomous microsystems without sensors: a new paradigm
The work has important implications for applications where explicit sensing is impractical or impossible. "Think of the targeted administration of drugs in the human body," suggests Dr. Nico Scherf from the Max Planck Institute for Human Cognitive and Brain Sciences. "Conventional approaches are based on pre-programmed responses or external control, but physiological flows are complex and unpredictable. Microrobots that learn from their own dynamics could potentially move autonomously in the body." The research also opens up new avenues for swarm robotics: multiple microrobots could reveal collective embodied intelligence.
"We are really only at the beginning of exploring what is possible when we consider physical embodiment as a computational resource," summarizes Paul. "This work demonstrates the principle experimentally. The next challenge is to apply these ideas to more complex environments and tasks."
Research team
In addition to Dr. Diptrabrata Paul and Frank Cichos from Leipzig University, the research team also included Nikola Milosevic and Nico Scherf from the Max Planck Institute for Human Cognitive and Brain Sciences , who contributed their expertise in the field of machine learning optimization. All researchers are affiliated with the Center for Scalable Data Analytics and Artificial Intelligence Dresden/Leipzig ( ScaDS.AI Dresden/Leipzig ). The research was supported by the German Federal Ministry of Research, Technology and Space (BMFTR) as part of the ACONITE project and the Center for Scalable Data Analytics and Artificial Intelligence (ScaDS.AI) Leipzig.
Original title of the publication in Science Advances
,, Physical Embodiment Enables Information Processing Beyond Explicit Flow Sensing in Active Matter ", doi: 10.1126/sciadv.aec0783
