Cognitive evolution enables de novo design of untethered millimeter-scale robots
This paper introduces EvoAgent, a large-language-model agent that employs cognitive evolution to overcome complex design constraints, successfully creating a high-performance, untethered 10-mm robot with significantly improved thrust and operational robustness compared to traditional genetic algorithm baselines.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Tiny machines that can move on their own without being tied to a power source hold a promise of exploring places too small or dangerous for humans. Imagine a device the size of a grain of rice that could crawl through a cracked pipe to find a leak or navigate inside a human body to deliver medicine. The challenge, however, is that at this tiny scale, every part of the machine must fit into a space no larger than a few millimeters. The battery, the motor, and the legs all have to share this cramped volume, and changing one part often breaks another. If the motor is made slightly larger to produce more power, it might leave no room for the battery, or it could shift the machine's balance so it cannot walk. Designing these robots usually requires a difficult balancing act where engineers try to guess the right combination of parts, often relying on trial and error that can take a long time and many failed attempts.
A team of researchers at Beihang University and other institutions has developed a new way to solve this puzzle using a type of artificial intelligence that thinks like an engineer. They created a system they call EvoAgent, which acts as a digital designer that learns from its own mistakes and successes. Instead of just randomly trying different shapes and hoping one works, this system uses a large language model to understand the physics behind the design. Before the computer even runs a simulation to see if a robot can move, EvoAgent proposes a new design and explains the physical reasons for its choices. It considers why a previous design failed, such as a wire being too thick or a gap being too small, and uses that knowledge to propose a better version. This process creates a "family tree" of designs where the lessons learned from one generation are passed down to the next, allowing the system to evolve a working robot much faster than traditional methods.
The researchers tested this approach by trying to design two critical parts of a millimeter-scale robot: the electromagnetic motor that provides the power, and the leg structure that converts that power into movement. They compared their new AI-driven method against older, standard techniques that rely on random variations, similar to how natural selection works in biology. The results showed that the AI method was far more efficient. It found a motor design that produced more than twice the pushing force of the best design found by the older methods, while using the same amount of space. When the researchers built the robot based on the AI's design, it proved to be remarkably capable. The final machine, which is only 10 millimeters long and weighs just one gram, could move at a speed of 6.5 body lengths per second. It was also much more reliable, successfully moving under thirty different combinations of electrical current and vibration frequency, whereas the robot designed by the older method only worked under ten conditions.
What makes this achievement significant is not just the speed of the robot, but how the design was achieved. The AI system managed to navigate the complex web of constraints that usually trap human designers. It learned that certain changes, like increasing the number of wire turns in the motor, would boost power but also add weight, and it figured out how to balance these trade-offs without needing to test every single possibility. The system recorded its findings, turning specific failures into general rules that it could apply to future designs. For instance, it learned to avoid certain shapes that caused the robot to lose contact with the ground. This ability to learn from physical feedback and apply those lessons to new problems suggests a new way to build machines. The researchers built the robot, tested it, and confirmed that the performance predicted by the computer simulations matched the real-world results. The robot could not only move forward but also turn sharply and carry a load more than five times its own weight.
This work demonstrates that artificial intelligence can do more than just optimize numbers; it can understand the physical world well enough to invent new machines. By treating the design process as a conversation where the computer learns from its own history, the researchers were able to create a robot that operates with a wide margin of safety and performance. The robot's ability to function across a broad range of conditions means it is more likely to succeed in the unpredictable environments where such machines are needed. While the study focused on a specific type of robot, the method used to design it could be applied to other complex systems where parts must work together in tight spaces. The success of this project suggests that the future of robotics may not just be about building smaller machines, but about teaching computers how to think like the engineers who build them, learning from every step of the process to create designs that are both innovative and practical.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.