University of Twente Student Theses


Exoskeleton controller design for cybathlon by parallel reinforcement learning

Mellema, Sibolt (2023) Exoskeleton controller design for cybathlon by parallel reinforcement learning.

[img] PDF
Abstract:Patients with Spinal Cord Injuries (SCI) experience physical disabilities due to the lost ability to walk. Wearable robotics, like an Exoskeleton can restore parts of the lost ability to walk. Challenges for an Exoskeleton is to perform different tasks such as stair ascending/descending, tilted paths or steppingstones. An exoskeleton gains the ability to perform the task by their controller. Current controller designs can be time-consuming due to complex systems, time-invariant and hard-to-measure quantities. Controller design by Reinforcement Learning may tackle the problems of time-invariant and hard-to-measure quantities. Parallel Reinforcement Learning shows the potential to tackle the time-consuming problem of the controller design. The focus in this work is to utilize parallel Reinforcement Learning for controller design. The software Isaac Gym provides a platform for parallel Reinforcement Learning, which uses a model of the Exoskeleton to move through the modelled environment. The design method for the controller has two stages. The first stage is to generate a model of the Exoskeleton worn by a human. The goal of this design is to prove the capabilities of locomotion for the Exoskeleton model. The second stage includes an environment with obstacles. The obstacles resemble different tasks which the Exoskeleton must complete. This stage includes four obstacles form the Cybathlon course for Exoskeletons (a race in which Exoskeletons completes different daily tasks). The design method includes an evaluation of the resulting Reinforced Learned controllers. The evaluation consists of two sections, a qualitative and a quantitative evaluation; to evaluate the gait of the exoskeleton and quantitative comparison between locomotion. The results have two designs for the Exoskeletons’ controller. The first design has a controller for locomotion on a smooth surface with human-like gait. The second design of the controller has the aim to overcome the obstacle. Unfortunately, the design did not show qualitative and quantitative results to overcome the obstacles. Both controllers have a training time under two hours, achieving a good controller within an hour. This works shows promising result for controller design with Reinforcement Learning for an Exoskeleton worn by humans. Further research can improve the design of the Reinforcement Learned controller by adjusting the reward function, variable setpoints for the body’s velocity or implementing Recurrent Neural Networks in the Reinforcement Learning method.
Item Type:Essay (Master)
Faculty:EEMCS: Electrical Engineering, Mathematics and Computer Science
Programme:Systems and Control MSc (60359)
Link to this item:
Export this item as:BibTeX
HTML Citation
Reference Manager


Repository Staff Only: item control page