Meeting Abstract

16 A-4   10:45 - 11:00  Learning tube feet control for sea star locomotion Heydari, S*; Merel, J; McHenry, MJ; Kanso, E; University of Southern California ; DeepMind; University of California, Irvine; University of Southern California sinaheyd@usc.edu http://scf.usc.edu/~sinaheyd/

There is a growing effort to understand decentralized control mechanisms, particularly in application to robotic systems with distributed sensors and actuators. Sea stars, being equipped with hundreds of tube feet, are an ideal model system for studying decentralized sensing and actuation. The activity of the tube feet is orchestrated by a nerve net that is distributed throughout the body; there is no central brain. How the numerous tube feet are controlled and coordinated for locomotion through such a distributed nervous system and what they sense remains mostly unknown. Here we use a sea star inspired locomotion model to find effective decentralized control policies for locomotion. Namely, we develop mathematical models of the biomechanics of the tube feet and the sea star body and use a reinforcement learning algorithm to train individual tube feet optimal locomotion control policies. We provide every foot with a set of local and biologically-plausible sensory cues and train the feet to perform an effective control policy that is cloned among all feet. We then evaluate the performance of the trained policies and find the optimal sensory cues for each control task. We find that tube feet trained on local sensory cues achieve stable forward locomotion and axial and shear strain are the most effective sensory cues for learning this task. We also find that policies trained on a few tube feet can be generalized to a larger number of feet, making the training process significantly faster. These findings shed light on the neuromechanics of Echinoderms and offer a new paradigm for walking using soft actuators, with potential applications to autonomous robotic systems.