Loading...

Control of a Lower Limb Exoskeleton Robot using a Reinforced Model Predictive Controller (RL-MPC)

Moosavi Zargar, Kazem | 2025

12 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 58849 (08)
  4. University: Sharif University of Technology
  5. Department: Mechanical Engineering
  6. Advisor(s): Vossoughi, Gholamreza
  7. Abstract:
  8. Spinal cord injury is a leading cause of lower-limb motor impairment and affects many people each year. Lower-limb motor deficits are a common consequence of spinal cord injury; patients with such deficits require repeated physiotherapy exercises to recover their gait cycle. The use of wearable exoskeleton robots for rehabilitation has increased in recent years; however, assistance only promotes the user’s improvement of their movement trajectory if it is provided when the user actually need it. Therefore, implementing an "assist-as-needed" strategy in partial rehabilitation substantially improves the effectiveness of therapy. In this study, the objective was to develop, design, and implement a model predictive controller augmented with deep reinforcement learning. The controller was designed for the knee joint of a wearable robot to assist knee flexion–extension movements. By estimating an online user capability index, the controller detects the user’s need for assistance and provides help accordingly, so that the controller behavior adapts along a continuum between zero force control (to minimize interaction forces) and position control (to correct the trajectory). Using an extracted dynamic model of the human–robot interaction, an adaptive and reasonably robust MPC was designed and validated in simulation. A human capability index was defined and a DRL agent was trained to estimate this index online; the estimator was then tested with synthetic capability profiles. The combined RL–MPC controller was subsequently implemented on the wearable robot of the Mechatronics Research Laboratory at Sharif University of Technology. Experimental results showed a maximum absolute tracking error of 3.3° in the complete impairment case, a maximum mean absolute interaction force of 3.6 N on the shank in the full capability case, and acceptable online estimation of the user capability index in tests. Results were also compared with previous laboratory studies reporting 8° and 6.7 N, respectively, and overall demonstrate that the RL–MPC controller can provide online assist as needed partial rehabilitation in real world conditions
  9. Keywords:
  10. Exoskeleton ; Rehabilitation ; Model Predictive Control ; Deep Reinforcement Learning ; Assiste-as-needed ; Lower Limb Exoskeleton

 Digital Object List

 Bookmark

No TOC