Predicting sit-to-stand motions with a deep reinforcement learning based controller under idealized exoskeleton assistance

Abstract

Maintaining the capacity for sit-to-stand transitions is paramount for preserving functional independence and overall mobility in older adults and individuals with musculoskeletal conditions. Lower limb exoskeletons have the potential to play a significant role in supporting this crucial ability. In this investigation, a deep reinforcement learning (DRL) based sit-to-stand (STS) controller is developed to study the biomechanics of STS under both exoskeleton assisted and unassisted scenarios. Three distinct conditions are explored: 1) Hip joint assistance (H-Exo), 2) Knee joint assistance (K-Exo), and 3) Hip-knee joint assistance (H+K-Exo). By utilizing a generic musculoskeletal model, the STS joint trajectories generated under these scenarios align with unassisted experimental observations. We observe substantial reductions in muscle activations during the STS cycle, with an average decrease of 68.63% and 73.23% in the primary hip extensor (gluteus maximus) and primary knee extensor (vasti) muscle activations, respectively, under H+K-Exo assistance compared to the unassisted STS scenario. However, the H-Exo and K-Exo scenarios reveal unexpected increases in muscle activations in the hamstring and gastrocnemius muscles, potentially indicating a compensatory mechanism for stability. In contrast, the combined H+K-Exo assistance demonstrates a noticeable reduction in the activation of these muscles. These findings underscore the potential of sit-to-stand assistance, particularly in the combined hip-knee exoskeleton scenario, and contribute valuable insights for the development of robust DRL-based controllers for assistive devices to improve functional outcomes.