AI Mimics Animal Gaits, Paving the Way for More Agile Robots
Researchers at Brown University have developed an artificial neural network capable of reproducing the diverse gait patterns of four-legged animals. This breakthrough, published in Neural Computation, offers new insights into how the brain controls complex movements and could significantly advance the development of more autonomous and adaptable quadruped robots.
Understanding the Complexity of Movement
The ability to seamlessly transition between different gaits – from a gradual walk to a rapid trot or a bounding leap – is a hallmark of animal locomotion. Understanding the underlying mechanisms that allow for this flexibility has long been a challenge for neuroscientists and roboticists alike. “We know the brain has to be able to flexibly and robustly maintain and change rhythms,” explains Carina Curto, a professor of applied mathematics at Brown University.
Attractor Networks: A Key to Decoding Gait
The Brown University team’s research centers around attractor networks, a mathematical framework used to model neural activity. These networks naturally settle into specific patterns and the researchers expanded this framework to encompass dynamic behaviors. Specifically, they utilized a Hopfield network, traditionally used for modeling static brain behaviors like memory recall, to create a model capable of generating a range of dynamic gaits.
From Theory to Practice: Generating Diverse Gaits
The resulting artificial neural network, comprised of just 24 artificial neurons, can generate five distinct quadruped gaits: bounding, pacing, trotting, walking, and pronking (a leaping gait). Crucially, the network can transition between these gaits rapidly and seamlessly without requiring any parameter adjustments. This efficiency mirrors the adaptability observed in biological systems.
Implications for Robotics
Current quadruped robots often rely on complex and computationally expensive programs, frequently requiring an internet connection to function optimally. The Brown University team’s research suggests a path towards creating robots that can operate more autonomously and efficiently. A robot inspired by this streamlined neural network could potentially function offline, making it suitable for a wider range of applications.
“This paper shows that you can expand attractor networks beyond the static to include the dynamic,” said Juliana Londono Alvarez, a postdoctoral researcher at Brown and the study’s lead author. “Once you do that, you can spot how the same principles underlying memory encoding can also generate something dynamic, like these gaits.” Londono Alvarez is currently exploring collaborations with roboticists to adapt the network for real-world robotic applications.
Funding and Collaboration
This research was supported by grants from the National Institutes of Health (R01 EB022862) and the National Science Foundation (DMS-1951165 and DMS-1951599). Additional support was provided by NSF grant DMS-1929284 during a residency at Brown’s Institute for Computational and Experimental Research in Mathematics (ICERM). The project also involved collaboration with Katherine Morrison, professor of mathematical sciences at the University of Northern Colorado.