The new model imitates animal movements with high accuracy.
Scientists from Harvard University, together with Google's DeepMind artificial intelligence laboratory, have developed a virtual rat model with an artificial brain capable of imitating movements similar to natural ones. The model was created to better understand how the brain controls movement.
Breakthrough in robotics
Modern robotics, despite all its achievements, still cannot reproduce the natural movements of animals and people. Diego Aldaondo, a Harvard graduate student and a participant in the project, noted that the main problems are both hardware and software aspects.
Hardware and software problems
Aldaondo explained, “On the hardware side, researchers have encountered challenges in creating robots that have the flexibility, strength, and energy efficiency of animal bodies.” On the software side, the main hurdles are developing effective physics simulations and machine learning algorithms to train controllers that mimic human movements.
Simulation and reality
There is also a problem known as the simulation-reality gap, caused by differences between physical simulations and the real world. This makes it difficult to transfer controllers trained in simulation to real robots.
Creating a virtual rat model
Together with Professor Bence Olwiecki of the Department of Organic and Evolutionary Biology, as well as other scientists from Harvard and Google DeepMind, Aldaondo developed a biomechanically realistic digital rat model.
Collaboration with Google DeepMind
The researchers teamed up with Google DeepMind as the platform developed tools for training artificial neural networks (ANNs) that can control biomechanical animal models in physics simulators. The team used MuJoCo , a physics simulator that simulates gravity and other physical forces, and developed another platform, Motor IMItation and Control (MIMIC), to train an ANN to behave in a rat. High-resolution data recorded from real rats was used to train the ANN.
Significance for neuroscience
Aldaondo noted, “This is important for neuroscience because it allows the development of computational models that reproduce the movements of animals in physical simulations and predict the pattern of neural activity that would be expected from real brains.”
The world of virtual neuroscience
Using ANN, the researchers were able to create inverse dynamic models that scientists believe our brain uses to control body movements and achieve the desired state.
Aldaondo explained: “In more bodily terms, one can think of the inverse model as creating the muscle activations necessary to achieve the desired posture, taking into account the physics of the body. This concept is useful for motor neuroscience because motor coordination involves learning to account for the physical properties of the body through experience with the world."
Model accuracy
Data from real rats helped the virtual model learn the forces needed to achieve the desired movement, even if it had not been specifically trained on them. When neural activity was measured in both real rats and a virtual model, the researchers found that the virtual model accurately predicted the neural activity of real rats.
New Horizons
This opens up a new frontier of virtual neuroscience, where artificially created animals can be used to study neural circuits and their disturbances in diseases. Olwiecki, an expert in teaching rats complex behaviors, now aims to use virtual models to solve problems faced by real rats.
“We want to start using virtual rats to test these ideas and help advance our understanding of how real brains create complex behavior,” Olwiecki added in a press release.
The results of the study were published in the journal Nature.
Scientists from Harvard University, together with Google's DeepMind artificial intelligence laboratory, have developed a virtual rat model with an artificial brain capable of imitating movements similar to natural ones. The model was created to better understand how the brain controls movement.
Breakthrough in robotics
Modern robotics, despite all its achievements, still cannot reproduce the natural movements of animals and people. Diego Aldaondo, a Harvard graduate student and a participant in the project, noted that the main problems are both hardware and software aspects.
Hardware and software problems
Aldaondo explained, “On the hardware side, researchers have encountered challenges in creating robots that have the flexibility, strength, and energy efficiency of animal bodies.” On the software side, the main hurdles are developing effective physics simulations and machine learning algorithms to train controllers that mimic human movements.
Simulation and reality
There is also a problem known as the simulation-reality gap, caused by differences between physical simulations and the real world. This makes it difficult to transfer controllers trained in simulation to real robots.
Creating a virtual rat model
Together with Professor Bence Olwiecki of the Department of Organic and Evolutionary Biology, as well as other scientists from Harvard and Google DeepMind, Aldaondo developed a biomechanically realistic digital rat model.
Collaboration with Google DeepMind
The researchers teamed up with Google DeepMind as the platform developed tools for training artificial neural networks (ANNs) that can control biomechanical animal models in physics simulators. The team used MuJoCo , a physics simulator that simulates gravity and other physical forces, and developed another platform, Motor IMItation and Control (MIMIC), to train an ANN to behave in a rat. High-resolution data recorded from real rats was used to train the ANN.
Significance for neuroscience
Aldaondo noted, “This is important for neuroscience because it allows the development of computational models that reproduce the movements of animals in physical simulations and predict the pattern of neural activity that would be expected from real brains.”
The world of virtual neuroscience
Using ANN, the researchers were able to create inverse dynamic models that scientists believe our brain uses to control body movements and achieve the desired state.
Aldaondo explained: “In more bodily terms, one can think of the inverse model as creating the muscle activations necessary to achieve the desired posture, taking into account the physics of the body. This concept is useful for motor neuroscience because motor coordination involves learning to account for the physical properties of the body through experience with the world."
Model accuracy
Data from real rats helped the virtual model learn the forces needed to achieve the desired movement, even if it had not been specifically trained on them. When neural activity was measured in both real rats and a virtual model, the researchers found that the virtual model accurately predicted the neural activity of real rats.
New Horizons
This opens up a new frontier of virtual neuroscience, where artificially created animals can be used to study neural circuits and their disturbances in diseases. Olwiecki, an expert in teaching rats complex behaviors, now aims to use virtual models to solve problems faced by real rats.
“We want to start using virtual rats to test these ideas and help advance our understanding of how real brains create complex behavior,” Olwiecki added in a press release.
The results of the study were published in the journal Nature.
