Using Artificial Demonstrations Emulating Human Movement Variability for a Learning-Based Exoskeleton Flow Controller

Robotic exoskeletons hold great potential for reducing physical effort and mitigating work-related musculoskeletal disorders. Nevertheless, designing exoskeletons that assist users without disrupting their natural movement remains challenging, as these devices must align seamlessly with human motor intent. Human motor control involves intricate interactions among the nervous system, musculoskeletal system, and environment, leading to movements that are inherently variable. This letter introduces a novel approach using a stochastic optimal control model of human movement to generate artificial demonstrations preserving this variability, facilitating training by eliminating the need for real user demonstrations. Using a learning-based method, the human motion is predicted in real-time, enabling an adaptive flow controller to dynamically modulate assistance. Specifically, we increase assistance during periods of high variability, when users should tolerate deviations. Experiments with the ABLE7D upper-limb exoskeleton during reaching movements demonstrated that this approach effectively increases interaction comfort, and matches the performance of methods trained on numerous real demonstrations. Furthermore, our findings indicate that non-personalized artificial demonstrations do not compromise user comfort, supporting the feasibility and contributing towards a plug-and-play exoskeleton experience.