Virtual mice with AI brains could mean better robots

Science has produced a better mouse, controlled by an artificial intelligence (AI) brain.

Developed by researchers at Harvard University and Google's DeepMind AI Lab, the virtual rodent is able to accurately mimic the movements of a real mouse, and represents a major milestone in our understanding of how the brain controls complex, coordinated movements.

This is a feat that even today's most advanced robots would struggle to replicate, and the researchers believe their discovery could lead to significantly greater agility in future robots.

The study was led by Bence Orbetsky, a professor in Harvard's Department of Organismic and Evolutionary Biology, who trained the artificial neural network using high-resolution data recorded from real mice. His lab is dedicated to studying the mechanistic mechanisms behind how the brain controls limb movement.

“We found that neural activity was better predicted by network activity in virtual rodents than by features of real rodent movements, consistent with both regions implementing inverse dynamics,” the researchers wrote.

According to a study published in the journal NatureA simulation of the virtual rat was created using the MuJoCo physics simulator, incorporating realistic forces such as gravity to mimic real-world conditions. The artificial neural network controlling the virtual rat's movements was trained with an inverse dynamics model, allowing it to predict the neural activity of a real rat with high accuracy.

The report noted that the findings could help scientists interpret neural activity in a variety of behaviors and relate it to principles of motor control. Principles derived from the virtual rat study could lead to improved robotic control systems.

“These results show how physical simulations of biomechanically realistic virtual animals can help interpret the structure of neural activity across behaviors and relate it to theoretical principles of motor control,” the study reads.

“We've learned a lot from the challenge of building embodied agents: AI systems that not only need to think intelligently but also translate those thoughts into physical actions in complex environments,” said Matthew Botvinick of Google Deepmind. Harvard Gazette. “We suspect that adopting the same approach in the field of neuroscience could help provide insights into both behavior and brain function.”

The team employed cutting-edge techniques, including AI, deep reinforcement learning, and 3D motion tracking, to enable virtual rats to reproduce a wide range of natural behaviors, including those that they were not explicitly trained to perform. This approach may establish the new field of “virtual neuroscience,” providing an accessible platform to study the neural underpinnings of natural behaviors to better understand how the brain controls movement, the paper notes.

This area of ​​research is essential for the development of advanced prosthetics and brain-machine interfaces, such as the technologies being developed by Neuralink and Precision Neuroscience.

The findings from this study may lead to new treatments for movement disorders by recreating neural circuits. Furthermore, the study points out that the virtual rat provides a transparent model to study neural circuits and the effects of diseases on these circuits.

Going forward, the researchers plan to give the virtual rats the autonomy to solve challenges encountered by real rats, further advancing their understanding of the brain's algorithms for skill acquisition.

Orbetzky said: Decryption.

Editor: Ryan Ozawa.