A collaboration between Harvard and Google DeepMind has produced a virtual rat controlled by an artificial neural network, and the results are forcing neuroscientists to reconsider how much of complex behavior can be captured with surprisingly simple models. The team recorded neural activity from real rats running through an open arena, then trained a deep learning model to predict the forces acting on the rat’s body from that brain data alone. That network was then dropped into a physics simulator, where it drove a biomechanically accurate rat body through the same kinds of movements the real animals made. The striking part is that the model learned to produce realistic locomotion without any explicit rules about how rats should move. It figured out the coordination problem on its own.
What makes this notable is the direct comparison the researchers could draw between the artificial network’s internal activity and the real neural recordings. When the virtual rat performed a movement, the activation patterns in the AI model often resembled the patterns seen in the animal’s brain. That kind of alignment is rare. Most AI models that claim to be ‘brain-inspired’ are loose metaphors at best. Here, the network was trained on actual neural data and then tested against held-out activity, giving the team a concrete way to ask whether the model had learned anything about how biological systems solve motor control.
The implications go beyond a neat demo. A system that can translate neural activity into physical behavior in real time could eventually help researchers test hypotheses about motor disorders, prosthetic control, or even how the brain adapts to injury. The virtual rat is also a sandbox. Scientists can lesion the model, alter the body’s parameters, or change the environment and watch how the network compensates — experiments that would be impossible or unethical in live animals. That’s the real promise: not replacing animal research, but giving neuroscientists a tool to run experiments that currently have no in vivo equivalent.
Still, this is a first step, and the model has limits. The virtual rat was trained on a relatively constrained set of behaviors in an open arena, and the physics simulation, while good, is not a rat. The team acknowledges that the network doesn’t yet capture the full complexity of sensory feedback or the role of spinal circuits, which handle a lot of movement control below the level of the cortex. But the fact that a generic sequence model could extract this much structure from raw neural data suggests that the gap between artificial and biological motor control may be narrower than anyone assumed. The next question is whether this approach scales — to more complex behaviors, longer timescales, and eventually, to animals with richer behavioral repertoires.
