Wearable devices, such as exoskeletons, are becoming increasingly common and are being used mainly for improving motility and daily life autonomy, rehabilitation purposes, and as industrial aids. There are many variables that must be optimized to create an efficient, smoothly operating device. The selection of a suitable actuator is one of these variables, and the actuators are usually sized after studying the kinematic and dynamic characteristics of the target task, combining information from motion tracking, inverse dynamics, and force plates. While this may be a good method for approximate sizing of actuators, a more detailed approach is necessary to fully understand actuator performance, control algorithms or sensing strategies, and their impact on weight, dynamic performance, energy consumption, complexity, and cost. This work describes a learning-based evaluation method to provide this more detailed analysis of an actuation system for our XoTrunk exoskeleton. The study includes: (a) a real-world experimental setup to gather kinematics and dynamics data; (b) simulation of the actuation system focusing on motor performance and control strategy; (c) experimental validation of the simulation; and (d) testing in real scenarios. This study creates a systematic framework to analyze actuator performance and control algorithms to improve operation in the real scenario by replicating the kinematics and dynamics of the human–robot interaction. Implementation of this approach shows substantial improvement in the task-related performance when applied on a back-support exoskeleton during a walking task.
