Improving Performance of EMG-driven Ankle-Foot Orthosis Through Passivation & Reinforcement Learning
- Publication Type:
- Thesis
- Issue Date:
- 2023
Open Access
Copyright Clearance Process
- Recently Added
- In Progress
- Open Access
This item is open access.
The human musculoskeletal system is known to weaken as we age, with deterioration more prominent at the limb joints. As the lowest joint of the body, the ankle joint must support the full weight of the body during all activities and movement tasks, making it very susceptible to injury and cartilage degeneration over time. The use of robotic rehabilitation devices for assisting in physiotherapy is not a new concept but has been steadily improving over the past decades, and discussion of how improvements can be made through electromyography, control systems, and artificial intelligence is the focus of this project.
Utilising passivity within control systems has shown to remove redundancy requirements and allows less conservative controller design, improving performance. Passivating any system through an automated script that utilizes linear matrix inequalities and optimisation techniques allows this efficient design, while also minimizing perturbation of the original system to retain reliable simulation and control. An optimal input-output pairing is also revealed through this analysis, allowing improved robotic design. Originally this technique was assumed to be applicable to all systems but was found to be more restrictive during research. A lemma was proposed to encapsulate the limitations which stemmed from the relative degrees of the transfer functions within the system.
Reinforcement learning developed an adaptive PID and adaptive admittance controller that could better handle uncertain interactions within a human-robot interfacing system. The adaptive PID controller was successful in producing better reference tracking abilities than traditional PID controllers in changing environment situations. The adaptive admittance controller was also shown to be a viable approach, but was generally outperformed by a switching controller.
Reinforcement learning was also used to try and classify motions from electromyography signals to act as the input to a rehabilitation robot. Using secondary data, convolutional neural networks were able to perform this classification to a high accuracy, and produce a feature dataset that could be used to classify motions. The aim of the reinforcement learning was to identify the essential features needed for classification, and prune the rest to create a more efficient categorisation system. Results showed reinforcement learning was not well suited for the task of state space reduction, and more traditional machine learning techniques provided high enough accuracy results that reinforcement learning was deemed unnecessary.
Overall, this project aimed to combine admittance control, reinforcement learning, and passivity-based process control to improve safety and ease-of-use of wearable ankle-foot orthoses.
Please use this identifier to cite or link to this item: