Last Updated: 9 November 2024
Researchers from Harvard University, in collaboration with Google's advanced AI lab DeepMind, have developed a virtual rat model. This model is equipped with an artificial brain designed to replicate the neural activity observed in real rats, mimicking their movement behaviors.
The team leverages search-based mechanisms in their model to explore goal-oriented behaviors, enhancing our understanding of brain-controlled movements.
Harvard graduate student Diego Aldarondo pointed out that there are several obstacles in both hardware and software. Hardware limitations make it difficult to achieve the autonomy and energy efficiency found in animal bodies, while software challenges involve creating realistic physics models and machine learning pipelines. These pipelines must be trained with controllers that accurately mirror human movement.
A "reality gap" exists, Aldarondo explains, where controllers learned in simulations face difficulties when applied to actual robots due to inaccuracies in physics simulators.
In partnership with Professor Bence Ölveczky of Harvard’s Department of Organismic and Evolutionary Biology and colleagues at Harvard and Google DeepMind, Aldarondo is leading the development of a realistic digital rat model.
To advance this project, the team engaged Google DeepMind, known for its expertise in training deep artificial neural networks (ANNs) within simulation environments. They used the MuJoCo physics simulator to apply gravity and other physical laws, creating a new pipeline dubbed Motor Imitation and Control (MIMIC). This pipeline teaches ANNs to emulate rat behaviors.
Training involved using detailed real rat data, enabling predictions about neural activity patterns that might occur in actual brains.
These inverse dynamic models suggest how the brain calculates necessary muscle activations to achieve desired postures considering the body's physics.
This approach underscores the potential of motor neuroscience, where learning physical characteristics essential for interacting with the environment is crucial. The virtual model has been instrumental in understanding how significant forces need to be to produce targeted arm movements, without direct training on these movements.
Comparative studies between the neural activities of real rats and the virtual model have confirmed that the AI accurately predicts real rat neural behaviors, although broader applications in studying neural circuits and diseases using AI-simulated animals are still on the horizon.
