Learning bioinspired joint geometry from motion capture data of bat flight. (10th April 2019)

• Record Type: Journal Article
• Authors: Bender, Matthew; Guo, Jia; Powell, Nathan; Kurdila, Andrew; Müller, Rolf
• Is Part Of: Bioinspiration & biomimetics. Volume 14:Number 3(2019:Sep.)
• Publication Date: 2019-04-10
• Subjects: manifold identification -- machine learning -- automated mechanical design -- bioinspired robotics; Biomimetics -- Periodicals; Biomedical materials -- Periodicals; Medical innovations -- Periodicals; Biomedical engineering -- Periodicals; 600
• DOI: 10.1088/1748-3190/ab0fba
• Languages: English
• ISSNs: 1748-3182

Abstract

Abstract: Bioinspired robotic systems are often designed by assuming that the kinematics of the biological system of interest are well approximated by an open kinematic chain subject to box constraints. These box-constrains are typically generated by designer interpretation of biomotion and anatomical studies or pragmatic fabrication constraints. In contrast to this standard design paradigm, this paper presents a methodology for learning joint geometry which restricts the range of motion of generic ball joints to only the reachable set observed during biomotion experiments. This reachable set is identified by constructing an analytical-empirical potential energy function over the experimental observations. The energy function effectively 'pushes' configurations close to the set of observations. This energy function is then thresholded to identify the zero-potential (ZP) configuration set, and joint geometry is constructed using the resulting contour. We construct an entire bat wing using this method, and demonstrate through motion capture experiments that the learned geometry successfully restricts motions to the ZP set.