Kinematics and sensorimotor control of walking in fruit flies

Abstract

Animals move with remarkable agility, fluidly adjusting their limbs as they seamlesslymove from one behavior to another. Although the movements may look effortless, they are generated by complex feedback circuits which integrate proprioceptive and exteroreceptive inputs with internal state to precisely move muscles. Understanding this circuitry would be not only scientifically satisfying, but also aid in the design of agile robots and medical interventions for motor impairements. In this thesis, we provide some insight into this circuitry by proposing new frameworks forquantifying animal motion and modeling the sensorimotor loop. In Chapter 2, we introduce Anipose, an open-source toolkit for robust markerless 3D pose estimation. We apply Anipose to quantify the 3D kinematics of walking in Drosophila melanogaster and discover a hitherto undescribed degree of freedom: the rotation of the femur limb segment in each of the legs. In the course of applying Anipose to various datasets, we found that the manual annotations took up a disproportionate amount of time when setting up 3D tracking. Taking a step towards addressing this problem, in Chapter 3 we introduce BKinD-3D, a method for discovering 3D keypoints from video without requiring any annotations. Together, these contributions form a firm foundation for 3D animal pose estimation. Next, we turn our attention to modeling the feedback circuitry underlying movement. InChapter 4, we argue for building multiscale models of sensorimotor loops to elucidate the role of proprioceptive feedback during behavior. We follow this prescription in Chapter 5 as we build a model of the sensorimotor loop of walking in Drosophila melanogaster. Our model includes sensory and actuation delays as well realistic kinematics based on data obtained from tracking with Anipose. When we combine these two features, we discover that flies move as fast as their physiology permits for robust walking. We conclude with future directions for improving frameworks for 3D pose estimation andmodeling sensorimotor loops in Chapter 6.