Understanding Sensorimotor Control Through Realistic Neuro-Biomechanical Simulation

Summary/Abstract The brain evolved to move the body, i.e., to implement sensorimotor control. Understanding the brain, then, is inextricably linked to understanding how it coordinates the joints and muscles of the body to generate competent behavior in dynamic and unpredictable environments in the service of goals. A tantilizing consequence of this is

that understanding sensorimotor control will shed light on overall brain organizational and computational principles, including, potentially, higher-level cognition, which evolved much more recently by adapting the circuits already in place to control movement. This promise has not yet been realized, however, because

standard approaches to studying motor control seek to reduce complexity by, for example, isolating simple circuits, studying artificial tasks, or constraining movements. These approaches thus avoid the sine qua non of motor control in biology: multiple interacting brain regions, multiple simultaneous goals, and multiple muscle

coordinations, all in the presence of many sources of noise and sensory delays. Here, we propose to embrace this complexity rather than reduce it and are enabled to do so through the use of 'virtual rat' models that comprise deep neural network controllers designed to be analogous to biological brains and biomechanically accurate

bodies that are instantiated in simulators with real physics. Using a high-throughput easy-to-use ‘virtual neuroscience’ platform that we are developing for our own use, and the use of the broader research community, we will train these models to imitate freely-behaving real animals such that they internalize the statistics of

naturalistic behavior and then train them to solve goal-directed tasks. This novel ‘deep neuroethology’ approach has two crucial features: highly biologically realistic behavior and the full transparency of a model. We will then apply this approach to generate and test longstanding hypotheses about motor control and learning. For

example, we will interrogate: (i) how the learning and execution of complex behavior are influenced by certain circuit motifs such as laterality, reciprocal inhibition between antagonistic muscle pairs, feedback architecture, sensor delays, cortical–subcortical interactions, and dopamine-mediated plasticity; (ii) how feedforward outputs

and feedback inputs – in the setting of noise and sensory delays – coordinate movement; (iii) how animals learn to adapt their behaviors quickly such that they can generalize across novel environments and tasks; and (iv) what roles the distinct neural representations and circuit motifs found throughout the hierarchy of the motor

system play in neural computation. The results of these studies will drive previously unachievable refinements to our theories of sensorimotor control and will thus spur new research directions in motor neuroscience and, potentially, in robotics and other fields. Finally, and perhaps most importantly, we will have demonstrated the

power of virtual neuroscience, inspiring future similar research programs, potentially using virtual animals of many species, to probe the mysteries of neural computation.