Harvard scientists developed a virtual rat model with an AI brain to study how brains control the movement of real rats. This innovative model accurately simulates neural activity observed in real rats, leading to similar behaviors.
Collaborating with Google's DeepMind AI lab, Harvard University researchers aim to understand better how brains control movement.
Advancements in Virtual Neuroscience
Partnering with Google's DeepMind AI lab, Harvard University researchers have developed a virtual rat model featuring an artificial brain capable of replicating natural movements. This model aims to enhance understanding of how brains regulate movement.
Despite significant progress in robotics, Interesting Engineering reported that mimicking the fluidity of animal and human motion remains a challenge. Diego Aldarondo, a Harvard graduate student involved in the project, highlighted hurdles in both hardware and software.
Aldarondo elaborated that challenges exist in both hardware and software domains. On the hardware side, researchers struggled to replicate animal bodies' flexibility, robustness, and energy efficiency in robots.
On the other hand, software hurdles involve developing efficient physics simulations and machine learning pipelines to train controllers to mimic human movement accurately.
He also pointed out the "sim-to-real gap," which complicates the transfer of controllers learned in simulation to real robots due to differences between physics simulators and real-world conditions.
Collaborating with Bence Ölveczky, a professor at the Department of Organismic and Evolutionary Biology, and other colleagues from Harvard and Google's DeepMind, Aldarondo led the development of a biologically accurate digital model of a rat.
Rats Are Becoming More Aggressive During The Pandemic, Warned CDC: Food-Borne Illnesses Can Possibly Spread (Photo : Oxana Kuznetsova on Unsplash)
Developing Virtual Rat Model
The researchers collaborated with Google DeepMind due to its expertise in training artificial neural networks (ANNs) capable of controlling biomechanical models of animals within physics simulators.
They utilized MuJoCo, a physics simulator replicating gravity and other physical forces. They devised a new pipeline called Motor Imitation and Control (MIMIC) to train the ANN to mimic rat behavior.
The researchers trained the ANN using detailed data obtained from real rats. Aldarondo highlighted the significance of this advancement for neuroscience, as it facilitates the development of computational models that simulate animal movement in physical simulations.
This enables predictions about the neural activity patterns expected in real brains. Employing the ANNs, the researchers constructed inverse dynamic models, believed to be utilized by our brains for guiding bodily movements and transitioning from the present body state to the intended state.
Aldarondo elaborated that in simpler terms, an inverse model determines the muscle activations necessary to attain a specific posture while considering the body's physics. This framework proves valuable in motor neuroscience, as it entails learning how to adapt to one's body's physical characteristics through interaction with the environment.
The information gathered from actual rats assisted the virtual model in understanding the forces necessary to generate the intended movement for reaching a specific state, even without direct training on them.
Upon measuring neural activity in both real rats and the virtual model, researchers discovered that the virtual model precisely anticipated the neural activity observed in real rats.
This marks the beginning of a new era in virtual neuroscience, wherein AI-simulated animals could be used to investigate neural circuits and potentially explore how they are affected by various diseases.
