Closing the gap between animal movement and robotic control

Animals move with a level of precision and adaptability that robots struggle to match. In Carnegie Mellon University’s Department of Mechanical Engineering, researchers are developing a new AI-driven approach to uncover how brains and bodies work together. By turning complex biological systems into models that can be tested and refined, the team seeks to understand and replicate animal performance in robotic systems.

One focus of The Biohybrid and Organic Robotics Lab are neuromechanical models that simulate how neural signals and physical movement continuously inform one another. These models are powerful, but difficult to build because with countless parameters, even the smallest miscalculation can lead to large gaps between simulated behavior and what researchers observe in real animals.

“Biological systems are incredibly complex,” said Camila Fernandez, Ph.D. candidate in the department of mechanical engineering. “We’re trying to model something where everything affects everything, and it’s not always clear which piece we need to adjust when outcomes don’t match predictions.”

To address this, the team developed a systematic, data-driven framework that removes much of the guesswork. Traditionally, improving these models relied heavily on experts manually tweaking parameters, running experiments, and then comparing results in a very time-consuming cycle. The new method introduces a reinforcement learning algorithm that acts like a digital twin of the model with a built-in guide.

“It’s like having a coach for your model,” explained Fernandez. “It tells us, ‘This parameter is underperforming, so you should focus your refinements here.’”

The system can quantitatively identify which parameters are most responsible for any discrepancies between real animal data, the original neuromechanical model, and the reinforcement learning driven digital twin.

“Notably, the system only adds complexity where absolutely necessary,” said Vickie Webster-Wood, associate professor of mechanical engineering. “Instead of making the entire model more complicated, the system identifies the specific parts that require more detail while keeping everything else as simple as possible. This approach reduces computational cost while improving accuracy.”

So far, the team has validated their approach using computational models and robotic analogs. While a direct, real-time comparison between animals and physical robots, rather than through a model, remains a challenge, the team plans to explore how they can apply this framework to physical robots. Ultimately, the goal is to accelerate discovery.

“There’s still so much we don’t understand,” Fernandez said. “We’re essentially looking at a black box and trying to figure out how it works. If we can speed up how we build and refine models, we can ask better, more targeted questions, and get closer to the answers.”

This research was published in npj Robotics.