- 89 views
- 127 downloads
Machine-learned Adaptive Switching in Voluntary Lower-limb Exoskeleton Control
-
- Author / Creator
- Faridi, Pouria
-
The overall goal of this work was to design an intelligent method to reduce the cognitive and physical burdens associated with walking using lower-limb exoskeletons after paralysis. Lower-limb exoskeletons with many operating modes (i.e., walking patterns) can be challenging to work with. Manufacturers have allocated a switch button, allowing their users to select an operating/walking mode by switching through a list of available modes. This approach however, consumes a lot of time and energy from users, as they have to switch many times to get their desired mode at each switching instance. The work in this thesis used temporal-difference (TD) learning from the field of computational reinforcement learning (RL), that requires no previous modeling and/or a training dataset, to reduce the switching-related issues. Through the use of biologically-inspired general value functions (GVFs), an adaptive controller (referred to as adaptive switching method) was designed to reduce the number of required switching actions on the part of the user and limit it to a single switching action (one time hitting a switch button) at each switching instance. The adaptive switching method used the environmental and contextual representations to create predictions on the future usage of each operating/walking mode, specific to each individual. Using TD learning, the predictions about the GVFs related to each operating/walking mode were updated and adapted to the exoskeleton users’ (the experimenters) preferences.
Three users each performed three unique experimental scenarios, wearing the exoskeleton and using the adaptive switching method. The scenarios were designed to be most representative of the real-world situations. Adaptive switching method created a ranking mechanism in the switching list, ranking the operating modes based on their likelihood of being used next, from top of the list to the bottom. The order of the operating modes in the switching list was updated at each time step. Learning parameters (e.g., learning weights) were initialized to zero and built upon users’ switching behavior. Predictions were quickly learned and formed the ideal order of the modes in the switching list based on the users’ walking patterns. In the case of uncertainties (i.e., when more than one operating mode could be utilized), the machine-learned method (adaptive switching) was able to predict all of the likely mode utilizations and ranked the desired modes at the top of the switching list. When a change in the users’ behavior was seen, the adaptive controller was able to quickly adapt to that changing behavior, unlearn the previous behavior and learn the new walking pattern. The adaptive controller did not force the users to select a mode, but optimized their switching actions. This work demonstrated that the developed machine-learned controller can adapt to different walking behaviors and changing environments, without the need for offline training. It created an avenue for personalized walking and smart, optimized human-robot interactions.
This proof-of-concept work is the first demonstration of GVF prediction and learning in lower-limb exoskeleton control. The outcomes of this work contribute to the fields of neuroscience, robotics, computing science and engineering, and sets the path for further investigation of biologically-inspired learning methods in wearable robots and human-robot interactions. -
- Graduation date
- Spring 2023
-
- Type of Item
- Thesis
-
- Degree
- Master of Science
-
- License
- This thesis is made available by the University of Alberta Libraries with permission of the copyright owner solely for non-commercial purposes. This thesis, or any portion thereof, may not otherwise be copied or reproduced without the written consent of the copyright owner, except to the extent permitted by Canadian copyright law.