Learning and Identification of human upper-limb muscle synergies in daily-life tasks with autoencoders

Introduction Human motor control underlies all human activities, such as reaching and grasping an object. Although the task space seems simple and low-dimensional, the number of actually involved actuators, i.e., muscles, is significantly higher than the effective degrees of freedom. Due to this anatomical redundancy, muscles shall work in a synergistic manner which leads to a lowdimensional synergy space and ergo a synergy-based control [1,2]. In fact, several studies have focused on human upper-limb muscle synergies either of only local muscles [3], or in isometric tasks [1][3], or in specifically designed tasks [4]. In the present work, instead of such specific conditions, more general and daily-life cases are considered and analyzed. Systematic experiments are conducted in order to gather the relevant multi-joint motion trajectories and multi-muscle surface electromyograms (sEMG), as well as identify task-dependent muscle synergies.