Seven in one stroke: the brain of a fly will teach you how to control an exoskeleton with the agility of an insect

A researcher from South Korea has created a neural network of 160,000 artificial neurons that mimics the work of a fly's brain. The scientist connected the system to virtual bodies, and it turned out that she was able to control them, gradually improving her skills of interacting with her avatars. The data obtained can be used to create motion control systems for robots and exoskeletons that can maneuver closer to insects. According to Russian experts, this is one of the promising areas of research necessary for the further development of artificial intelligence. New technologies have already borrowed a number of principles from nature. For example, developments in the field of computer vision were created taking into account the principles of the nervous system of animals.

The Korean fly

A Korean researcher known on social media under the nickname Yakshawan and who did not want to reveal his real name told "Izvestia" about his unique experiment with a neural network simulating the brain of a fruit fly. The scientist connected two virtual bodies to it to test whether the system could learn how to control them. According to him, AI is gradually mastering this process, which in the future may open up the possibility of creating motion control systems for robots, exoskeletons and prosthetics with maneuverability comparable to insects.

— The purpose of this project is to find out whether the nervous system, which has evolved for the body of a fruit fly, is able to adapt to new sensory signals and movements if placed in a completely different body. I created a trainable virtual nervous system based on the real central nervous system of a male fruit fly and connected it to two virtual bodies: a four—legged robot and a humanoid avatar of an anime-style girl," Yakshawan explained.

The scientist observes how the nervous system independently finds ways to make meaningful movements. Although both models initially use the same neural network structure, the nature of learning differs significantly depending on body shape. The relatively simple four-legged robot is already showing steady progress in the direction of purposeful movement. The girl's avatar, which has significantly more joints and a body shape more different from the structure of a fruit fly, is still having difficulty forming a meaningful movement towards the goal. At the same time, there is slow but noticeable progress in this model.

"I plan to continue my training and observation to find out which movement strategy the nervous system of the fruit fly will eventually be able to detect on its own," said the author of the experiment.

According to him, in the future, it is possible to create models based on biological neural networks, which were effectively formed during evolution, and use them instead of traditional motion control systems in various machines. At the same time, a separate artificial intelligence will be able to be responsible for making higher-level decisions. This approach, for example, can be used in robotics, "brain—computer" interfaces, and devices that enhance human physical capabilities.

Infrared study: Underwater solar power plant and robot neurosurgeon

"Izvestia" collected the top 5 science news for the week

Practical application

At the moment, only a small number of organisms have been able to completely reconstruct the structure of the connections of the nervous system at the synaptic level. Among them, the fruit fly has one of the most complex nervous systems. The model includes not only the brain, but also the ventral nerve cord with motor circuits. It has more than 160,000 neurons and a huge number of synaptic connections. This makes it a convenient object for studying the full path — from receiving sensory information to forming a motor signal. At the same time, the model is small enough that such experiments can be carried out on a regular user computer, the scientist said.

However, it is not yet possible to determine exactly how the model or even the real brain of the fly perceives what is happening. It also remains unknown whether they possess what is commonly called consciousness.

Photo: Yakshawan Personal Archive

As Elena Kantonistova, associate professor of the Department of Big Data and Information Retrieval at the HSE Faculty of Computer Science and winner of the Yandex ML Prize 2026, explained to "Izvestia", developers have now achieved high results in the field of large language models. At the same time, in the field of "physical" AI, which is used, for example, to control robotic arms or exoskeletons, the results are much more modest. Therefore, according to her, it is important to develop this area.

— Such systems are necessary for the further development of robotics. This is also important for people with disabilities, such as those with impaired leg movement. With the help of such AI architectures, it is potentially possible for such patients to regain the ability to move. Technologies borrowed from nature are already being actively used. For example, many computer vision systems work on the basis of so—called convolutional neural networks that reproduce the principles of the nervous system of cats, she said.

Experiments with the capabilities of the fly's brain began after scientists managed to fully describe its nervous system. Based on this data, it was first reproduced in a virtual environment where the model could fly, feed, and perform other actions. The fly neural network was also trained to play computer games, neurophysiologist Mikhail Lebedev told "Izvestia".

— This system has enough elements to work as a neural network. Such experiments will show which tasks it is best used for. Most likely, these will be tasks similar to those performed by a fly: to see a stimulus and fly towards it, or to distinguish between edible and inedible," the scientist said.

At the same time, as Yakshawan emphasized, his work is not an academic study, but a personal experimental project and has no specific practical purpose.