A Drosophila connectome prior enables adaptable embodied control
This study demonstrates that using the adult *Drosophila* brain connectome as a fixed structural prior for embodied control enables superior adaptability and robustness across diverse locomotion tasks, body morphologies, and environmental complexities compared to standard artificial neural networks.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer
Intelligence is not just a thing that happens inside a brain; it is a conversation between a body and the world. For an animal to move, its nervous system must constantly translate what it feels and sees into actions that keep it balanced and moving forward. This process, known as embodied control, is incredibly difficult because the world is never static. A creature must adjust its steps when the ground becomes uneven, turn sharply when a predator appears, or speed up when it senses food. While engineers have built robots that can walk over rough terrain, they often rely on complex computer programs that are trained specifically for one body shape or one type of environment. If you change the robot's legs or the texture of the floor, the program often fails. This raises a fundamental question for scientists: does the specific way a biological nervous system is wired give it a special ability to adapt that artificial systems lack? To find out, researchers turned to the fruit fly, a creature with a nervous system that has been mapped in exquisite detail, down to every single connection between its nerve cells.
The researchers took the complete map of connections from an adult fruit fly's brain, known as a connectome, and used it as the fixed "brain" for a computer simulation of a digital fly. This digital fly had a body with six legs, sensors to feel the ground, and muscles to move. The team did not rewire the brain map; instead, they kept the connections exactly as they are in nature and only taught the fly how to read its sensors and how to send signals to its muscles. They then put this digital fly through a series of tests, asking it to walk across flat ground, climb over blocks of increasing height, turn in different directions, and change its speed. They compared this fly, which used the natural brain map, against a digital fly controlled by a standard artificial intelligence program designed to learn from scratch.
When the flies walked on flat ground at a steady speed, both performed equally well. This showed that the natural brain map was not inherently superior for simple, unchanging tasks. However, the moment the situation changed, the difference became clear. When the digital fly had to climb over blocks that were nearly half a millimeter high, the fly with the natural brain map crossed them successfully far more often than the artificial one. It also moved more smoothly, with less jerky motion, and was better at tracking changes in speed, such as when asked to accelerate or slow down in a specific pattern. When the fly had to turn, the natural brain map allowed it to reorient its body more precisely and with less wasted movement. The artificial program struggled to make these adjustments, often overshooting its target or losing its balance.
To understand why the natural map worked better, the researchers looked inside the simulation to see how information flowed. They found that when the fly encountered a difficult obstacle, the natural brain map spontaneously activated specific pathways that connected the sensors in the legs to the motor commands. These pathways, which include neurons that carry signals upward from the body and others that carry commands downward from the brain, seemed to work together in a coordinated way that the artificial program could not replicate. The natural system appeared to have a built-in flexibility, allowing it to recruit the right parts of the network for the job at hand without needing to be explicitly told to do so. Furthermore, the researchers discovered that the natural brain map was more robust to small errors or changes in its settings, suggesting that its structure creates a stable foundation for learning new behaviors.
The most striking part of the study came when the researchers tested whether this advantage held up when the body itself changed. They took the exact same fixed brain map from the fly and connected it to the bodies of very different simulated animals: a four-legged ant, a two-legged robot, a three-legged tripod, and a six-legged hexapod. In every case, the brain map that evolved for a fly helped the new bodies move better than the artificial programs did. The advantage was especially clear when the animals had to navigate complex mazes. As the mazes became more difficult, with more turns and dead ends, the fly-brain controller became increasingly better at finding the way compared to the artificial controllers. Even though the fly's brain was designed for a six-legged insect, it helped a two-legged robot stand upright and a four-legged ant run faster and more directly. The study suggests that the specific wiring of the fly's brain contains a general principle for adaptability that can be transferred across different body shapes.
This work does not claim that we can simply copy a fly's brain and put it into any robot. The simulations showed that the natural wiring provides a strong starting point, or a "prior," that makes learning new movements easier and more reliable. The artificial programs still needed to learn how to read the sensors and move the muscles of their new bodies, but they learned much faster and performed better when they started with the fly's brain map than when they started from scratch. The results suggest that evolution has shaped nervous systems not just to perform specific tasks, but to provide a flexible structure that can handle the unexpected changes of the real world. By using the fruit fly's brain as a guide, the researchers have shown that the way neurons are connected matters just as much as the neurons themselves, offering a new blueprint for building machines that can adapt to a changing world.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.