Frontiers in Robotics and AI (Nov 2022)
On the dynamics and control of a squirrel locking its head/eyes toward a fixed spot for safe landing while its body is tumbling in air
Abstract
An arboreal mammal such as a squirrel can amazingly lock its head (and thus eyes) toward a fixed spot for safe landing while its body is tumbling in air after unexpectedly being thrown into air. Such an impressive ability of body motion control of squirrels has been shown in a recent YouTube video, which has amazed public with over 100 million views. In the video, a squirrel attracted to food crawled onto an ejection device and was unknowingly ejected into air by the device. During the resulting projectile flight, the squirrel managed to quickly turn its head (eyes) toward and then keeps staring at the landing spot until it safely landed on feet. Understanding the underline dynamics and how the squirrel does this behavior can inspire robotics researchers to develop bio-inspired control strategies for challenging robotic operations such as hopping/jumping robots operating in an unstructured environment. To study this problem, we implemented a 2D multibody dynamics model, which simulated the dynamic motion behavior of the main body segments of a squirrel in a vertical motion plane. The inevitable physical contact between the body segments is also modeled and simulated. Then, we introduced two motion control methods aiming at locking the body representing the head of the squirrel toward a globally fixed spot while the other body segments of the squirrel were undergoing a general 2D rotation and translation. One of the control methods is a conventional proportional-derivative (PD) controller, and the other is a reinforcement learning (RL)-based controller. Our simulation-based experiment shows that both controllers can achieve the intended control goal, quickly turning and then locking the head toward a globally fixed spot under any feasible initial motion conditions. In comparison, the RL-based method is more robust against random noise in sensor data and also more robust under unexpected initial conditions.
Keywords
