The renowned physicist Richard Feynman famously stated, “What I cannot create, I do not understand.” I believe the same holds true for neuroscience’s understanding of how the nervous system controls the body.
The remarkable agility and robustness of animal movements arise from tight interactions between the central nervous system, the peripheral nervous system, the musculoskeletal system and the environment. Understanding the mechanisms underlying animal behavior therefore requires integrative approaches that consider the nonlinear interactions among all these components.
Neuromechanical models that integrate neural circuits, the body and the environment are becoming key scientific tools for neuroscience to test whether proposed biological mechanisms are sufficient to generate behavior. These models—implemented either as computer simulations or in physical robots—offer neuroscientists the chance to consider the important roles of the body and the environment in the study of animal motor skills and behaviors.
Neuromechanical models can complement animal experiments. Building a neuromechanical model is a rigorous intellectual exercise that requires explicitly representing the relevant components and interactions underlying a target animal behavior. Once implemented, it enables the typical iterations of a scientific methodology, including hypothesis design, synthetic experiments, predictions and comparisons with animal experiments.
This type of synthetic approach is particularly useful for demonstrating the sufficiency of specific mechanisms as opposed to their necessity. Animal experiments can often show that a particular control loop is necessary—for example, through ablation experiments—but can rarely demonstrate that it is sufficient to explain a given behavior. By contrast, a robot or simulation, in which different components of a control circuit can be selectively activated or deactivated, not only demonstrate sufficiency but help reveal the nonlinear interactions among all components involved in animal motor control.
T
he field of neuromechanical modeling finds its roots in early work in cybernetics and robotics; in the 1950s, W. Grey Walter developed tortoise robots that exhibited phototaxis—moving away or toward a light—and conditional learning. These robots were physical precursors of Braitenberg “vehicles,” a series of thought experiments that explored how behaviors could emerge from sensorimotor circuits in virtual agents. In the 1990s, Örjan Ekeberg and Sten Grillner combined neural circuits, an elongated body and a fluid dynamics model to develop a neuromechanical simulation of the lamprey. Their work—one of the first complete neuromechanical simulations of animal motor behavior—provided an understanding of how coupled neural oscillator circuits could be modulated for speed and heading control.Since then, several landmark studies have used neuromechanical models and simulations to demonstrate the sufficiency of hypothesized biological mechanisms. For example, one showed that measurements of optical flow are sufficient to mediate bees’ ability to avoid obstacles and regulate speed and landing. Another demonstrated that a relatively simple bilateral auditory circuit was sufficient to explain phonotaxis, or sound-seeking, in crickets. In my lab, we have shown that incorporating slow limb oscillatory circuits into a fast axial swimming network is sufficient to explain gait transitions in salamanders.
Collectively, these studies support the view that animal motor control relies on multiple interacting control loops, with substantial redundancy between central and peripheral mechanisms. Because several mechanisms operate simultaneously, their respective contributions are often difficult to disentangle without neuromechanical modeling. In particular, these circuits are often controlled by a combination of feedforward and feedback control loops, and their relative contributions in different animals likely depend on the mechanical stability of locomotion.
In addition, neuromechanical models offer interesting advantages in comparison with animal experiments: They are repeatable; they offer access to variables or quantities that would be difficult to measure in animals; and their morphology and environments can be systematically changed. They are particularly useful for experiments that cannot be performed on real animals for practical, financial or ethical reasons. For locomotion control, for instance, it is interesting to investigate all possible gaits that an animal could, in principle, perform (including gaits that animals do not exhibit), to identify the tradeoffs that an animal has to satisfy in terms of speed versus energy efficiency or sensing versus locomotion performance.
Sometimes I am asked whether it is better to implement neuromechanical models in a robot or in simulation. I always advocate for trying computer simulations first. Thanks to the rapid improvement of physics engines, such as MuJoCo and Isaac Lab, and the availability of various computational models of muscles and neurons, it is now possible to simulate neural circuits, the musculoskeletal system and the environment faster than ever possible before. Examples include simulations of the fruit fly, lamprey, zebrafish, mouse and human. Compared with physical robots, simulations are generally faster, cheaper, more accessible, easier to tune and more accurate for modeling some components, such as muscle-tendon pairs.
At the same time, going the extra mile to build a physical robot is often worth it, for two reasons. First, the robot benefits from real-world physics, which is important for situations in which the body-environment interactions are difficult to simulate numerically, such as complex fluid dynamics or complex terrains, including sand, mud and grass. Second, robots can interact with animals and real environments, enabling real-world experiments that provide realistic and rich sensory inputs, including natural visual, auditory and chemical cues. Recently, we modeled the optomotor response in zebrafish and demonstrated that it could help a fish-like robot maintain its position in a river, despite the complex water flow and visual inputs of the real world. The ability to directly interact—physically and socially—with animals, potentially even in their natural environments, has led to interesting studies investigating collective behavior in fish and cockroaches.
In the next few years, I foresee tighter interactions between neuroscience, numerical modeling and robotics. Interestingly, robotics is moving toward neuroscience and biology in several ways. In terms of morphologies, there is currently a boom of animal-like, quadruped and humanoid robots, several of which are commercially available for reasonable prices. In terms of control, the state-of-the-art locomotion controllers are now implemented as neural networks trained with deep reinforcement learning algorithms, similar to how animals learn to move.
As a result, we will have many opportunities to investigate important questions about animal motor control and behavior, such as agility, fault tolerance, multimodal sensorimotor integration, action selection and evolution and developmental processes. This will benefit not only neuroscience but also robotics and the development of more agile biologically inspired robots for field applications such as planetary robotics, environmental monitoring, facilities inspection, agriculture and transport.
