Loading...

Design and Implementation of an Intelligent RL-based Controller for the Lower-limb Exoskeleton to Reduce Interaction Torque

Abbasi, Mohammad Reza | 2021

259 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 54297 (08)
  4. University: Sharif University of Technology
  5. Department: Mechanical Engineering
  6. Advisor(s): Vosoughi, Gholamreza; Moradi, Hamed
  7. Abstract:
  8. Exoskeletons, or wearable robots, are electromechanical devices that have become the focus of academic and industrial research in recent years, and their applications in power augmentation have increased. One of the most important challenges of these applications is controlling the robot and synchronizing it with the user. The close interaction of the robot with the user and the change of movement pattern between different users and different gait cycles show the importance of estimating the user's movement intention, but the need for online method of estimating the movement intention and the complexity of accurate dynamic modeling has caused researchers to use reinforcement learning (RL) in robotics. The benefits of reinforcement learning include online learning of optimal policy, model-free learning through experience, and not being limited to a specific movement. However, the use of reinforcement learning methods is associated with challenges such as safety, convergence speed, and sample inefficiencies that will be addressed in this study.In this research, a control strategy based on reinforcement learning is proposed for power augmentation application. The purpose of this method is to estimate the motion intention of the user and reduce the interaction force between the user and the robot. First, in the offline phase, the frequency and amplitude features are extracted with the help of adaptive frequency oscillators. Then dynamic movement primitives (DMP) are used to extract the DMP parameters. Subsequently, a gait trajectory database consisting of 89 examples is augmented using a Gaussian process regression to produce a new 625-example gait dataset.In the online phase of the proposed strategy, the frequency and amplitude are extracted and given to the regression module to produce initial DMP parameters. Moreover, using a deep autoencoder the dimensions of the policy parameter space are reduced. Finally, reinforcement learning is carried out in this reduced space to ensure safety, speed, accuracy and quality of the final solution. The simulation results indicate that the RL algorithm converges after 100 gait cycles with a RMS of tracking error and interaction torque of 0.8 (deg) and 0.54 (N.m). Also in the first experiment with a slow pace, the convergence of the RL algorithm happens after 200 gait cycles with an RMS of tracking error and interaction torque of 1.44 (deg) and 1.05 (N.m). Finally, for the second experiment with a moderate pace the RL algorithm achieves convergence in 200 cycles and the RMS of tracking error and interaction torque are 1.95 (deg) and 1.49 (N.m). A comparison between the results of the proposed strategy with that of the zero-force algorithm demonstrates a 9% and 37.5% reduction in RMS of tracking error and interaction torque, respectively
  9. Keywords:
  10. Lower Limb Exoskeleton ; Autoencoder ; Reinforcement Learning ; Power Augmentation ; Interaction Torque ; Intelligent Control

 Digital Object List

 Bookmark

No TOC