Reseachers at Harvard University, in collaboration with Google’s Deep Mind AI lab, have created a virtual rat model with an artificial brain that can mimic movements like its natural counterpart.
The researchers created the model to understand better how brains control movement.
For all its advances, the modern-day field of robotics has yet to emulate the natural motion of animals and humans. In an interaction with Interesting Engineering (IE), Diego Aldarondo, a graduate student at Harvard who was involved in the work, said that challenges were both in hardware and software.
Aldarondo explained, “On the hardware side, researchers have found it difficult to build robots with the flexibility robustness and energy efficiency of animal bodies.”
While on the software side, he mentioned that the hurdles are developing efficient physics simulations and machine learning pipelines to train controllers to emulate human movement.
“Another challenge, known as the sim-to-real gap, arises from differences between physics simulators and the real world, making it difficult to transfer controllers learned in simulation to real robots,” Aldarondo added.
Along with Bence Ölveczky, a professor at the Department of Organismic and Evolutionary Biology, and other researchers at Harvard and Google’s DeepMind, Aldarondo developed a biomechanically realistic digital model of a rat.
Building the virtual rat model
The researchers teamed up with Google DeepMind since the platform has developed tools to train artificial neural networks (ANNs) that can help control biomechanical models of animals in physics simulators.
The team worked with MuJoCo, a physics simulator that simulates gravity and other physical forces, and developed another pipeline, Motor IMItation and Control (MIMIC), to train the ANN on rat behavior.
The researchers used high-resolution data recorded from real rats to train the ANN.
“This is significant for neuroscience as it enables the development of computational models that recapitulate animal movement in physical simulation and predict the structure of neural activity that we ought to expect from real brains,” Aldarondo elaborated in the email to IE.
World of virtual neuroscience
Using the ANNs, the researchers were able to build inverse dynamic models, which scientists believe our brains use to guide bodily movements and enable us to move from the current state of the body to the desired state.
“In more bodily terms, one can think of an inverse model as producing the muscle activations required to achieve a desired posture, subject to the physics of the body. This framework is useful for motor neuroscience, as coordinating movement involves learning how to account for the physical properties of one’s body through experience interacting with the world,” Aldarondo added in the email.
The data from real rats helped the virtual model learn the forces needed to produce the desired movement to reach a desired state, even if it was not explicitly trained on them. When neural activity was measured for both the real rats and the virtual model, the researchers found that the virtual model accurately predicted the neural activity of real rats.
This opens up a whole new world of virtual neuroscience, where AI-simulated animals could be used to study neural circuits and, perhaps, how they are compromised in disease.
Ölveczky, an expert at teaching rats complex behaviors, is now keen to use the virtual models to help solve problems real rats encounter.
“We want to start using the virtual rats to test these ideas and help advance our understanding of how real brains generate complex behavior,” Ölveczky added in the press release.
The research findings were published today in the journal Nature.
ABOUT THE EDITOR
Ameya Paleja Ameya is a science writer based in Hyderabad, India. A Molecular Biologist at heart, he traded the micropipette to write about science during the pandemic and does not want to go back. He likes to write about genetics, microbes, technology, and public policy.