Tiny robots learn to swim
A medicine that arrives exactly where it is needed. No detour through the whole body, no side effects elsewhere. Researchers at Leipzig University have now taken an important step towards this goal.
They have succeeded in doing something that has never been done before: Tiny artificial particles, so-called microswimmers, navigate independently through changing liquid currents. And they do this without any conventional sensors. Instead, an algorithm uses the movement of the particles themselves as a source of information. The results were published in the journal Science Advances.
The body as a compass
The microswimmers are around one micrometre in size and therefore a thousand times smaller than a millimetre. They consist of tiny plastic beads coated with gold nanoparticles. A laser heats them up unevenly and sets them in motion.
The special thing about them is their learning process. The particles are trained using reinforcement learning, an artificial intelligence method based on learning through reward and feedback. The algorithm observes their movement and gradually adapts their behavior. After around 50 training rounds, the particles can handle currents that are four times stronger than their own propulsive force.
"The experiments themselves were quite challenging," says Dr. Diptabrata Paul from the Peter Debye Institute for Soft Matter Physics at Leipzig University. "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."
Intelligence without electronics
The central idea of the work is called "embodied intelligence" by the researchers. Instead of using sensors, the movement of the particles themselves serves as a source of information. Any deviation from the course reveals something about the flow - the body becomes a sensor. "This is fundamentally different from our usual idea of robot design," explains Paul. Instead of measuring flows with sensors and then calculating reactions, the researchers use the physical interaction between the particle and its environment itself as a source of information.
The principle is based on nature. For millions of years, motile microorganisms have used their body shape to orient themselves in liquids. The Leipzig researchers have now shown that machine learning can also discover similar strategies in artificial systems - and within experimentally feasible time frames.
The research could open up new applications in the long term. For example, the targeted administration of drugs in the body. Here, flows are complex and difficult to predict. In future, micro-robots capable of learning could move autonomously within them. "We are really only at the beginning," says Paul. The next challenge is to transfer the principle to more complex environments and tasks.
Publication:
Diptabrata Paul et al. ,Physical embodiment enables information processing beyond explicit flow sensing in active matter.Sci. Adv.12,eaec0783(2026).
