Researchers from the New Jersey Institute of Technology working with colleagues in China have demonstrated that microbots can adjust to changing surroundings while moving through liquids by using neural networks. Their findings were shared in a journal article published by Communication Physics, highlighting a notable step in the study of autonomous micro-scale devices.
These floating microrobots hold promise for precise drug delivery and delicate medical tasks such as microsurgery. Yet their maneuverability has been a limiting factor until now, constraining their practical use in real-world medical settings where fluid flows and disturbances are common.
By combining neural networks with reinforcement learning, the teams showed that the microrobots, built from three components linked by pulling joints, can learn to navigate difficult conditions and steer in the desired direction within liquid environments. The learning process mirrors how a human learns to swim, gradually refining movements through trial, feedback, and adaptation to the surrounding currents and obstacles.
In the training phase, each action performed by a microrobot produced feedback about its accuracy and effectiveness. The robot then adjusted its behavior over time, using this feedback to improve future performances, effectively building a memory of how to respond to different environmental cues and disturbances.
After prolonged training, the microrobots demonstrated the ability to follow complex trajectories without being programmed with explicit instructions for every possible scenario. They also proved resilient to disturbances caused by the motion of surrounding fluids, maintaining course and precision even when external conditions shifted rapidly.
The researchers concluded that adaptive behavior is essential for deploying microrobots in complex, real-world environments where environmental factors cannot be fully anticipated. This capability paves the way for more reliable medical applications and other tasks where autonomous micro-scale agents must operate in dynamic, unpredictable fluids. This work appears in Communication Physics and represents a meaningful advance in the integration of learning systems with micro-scale robotics, underscoring the potential for smarter, more capable devices in the future.