AI Neural Network Models Animal Gaits for Robotics Advances
Researchers at Brown University have developed an artificial neural network that mimics the gait patterns of four-legged animals, offering new insights into how brains process complex movements and potentially advancing the capabilities of quadruped robots.
The work, published in Neural Computation, centers on “attractor networks,” a mathematical model used to understand how neural activity settles into patterns. Professor of applied mathematics Carina Curto explained that the team expanded the existing attractor framework to model dynamic behaviors, rather than static ones like memory recall. “We know the brain has to be able to flexibly and robustly maintain and change rhythms,” Curto said. “By tapping into the rules of attractor networks, we have created an artificial neural network that hints at how biological brains might simultaneously encode and transition between different patterns and rhythms.”
The network, comprised of just 24 artificial neurons, can generate five distinct quadruped gaits – bounding, pacing, trotting, walking, and pronking – and seamlessly transition between them, even responding to sudden changes in terrain without requiring parameter adjustments. This ability to shift between gaits, such as slowing to a walk on uneven ground and then accelerating on a flat surface, is crucial for efficient locomotion, as demonstrated by animals in the natural world.
Juliana Londono Alvarez, a postdoctoral researcher at Brown and the study’s lead author, emphasized the significance of this expansion of attractor network theory. “This paper shows that you can expand attractor networks beyond the static to include the dynamic,” she said. “Once you do that, you can observe how the same principles underlying memory encoding can also generate something dynamic, like these gaits.”
The research team collaborated with Katherine Morrison, professor of mathematical sciences at the University of Northern Colorado, who participated in the project during a residency at Brown’s Institute for Computational and Experimental Research in Mathematics (ICERM), a National Science Foundation-funded mathematics institute.
Beyond fundamental neuroscience, the researchers suggest the network could inspire a new generation of quadruped robots. Current quadruped robots rely on complex and resource-intensive programs, often requiring an internet connection to function. A robot built on the principles of Curto’s streamlined neural network could potentially operate offline, offering greater autonomy and efficiency. Londono Alvarez is currently in discussions with roboticists to explore adapting the network for use in their projects.
This research was supported by National Institutes of Health grants R01 EB022862, NSF DMS-1951165 and DMS-1951599, and National Science Foundation grant DMS-1929284, awarded during the mathematicians’ residency at ICERM’s 2023 Math + Neuroscience program.