- 39 views
- 121 downloads
Learning to Partner: Exploring Real-Time Adaptive Feedback via Temporal-Difference Machine Learning for Improved Human-Prosthesis Collaboration
-
- Author / Creator
- Parker, Adam SR
-
Modern myoelectric artificial limbs are sophisticated devices with many of the degrees of freedom of biological limbs.
These devices have great potential to provide function for people with amputations, assisting them in participating in a greater number of activities and tasks of daily living.
While tremendous advancements have been made in the control of myoelectric prostheses, the interface between user and device only allows users to scratch the surface of the capability of modern prosthetic devices.
Importantly to the interface, feedback of any kind is not widely commercially available from prosthetic devices.A potential path towards improving user interactions with prosthetic limbs in the current age of artificial intelligence is to view the device not only as a tool being used but as a partner assisting the user in their daily life.
That is the goal of this work: to \emph{apply real-time machine learning methods to wearable assistive robotics to promote collaborative partnerships with users}.
Most prior work using machine learning in prostheses has been focused on how users control the device.
Research has been increasing in providing feedback from devices to users with a focus on communicating sensation, but has not yet begun to explore how to use machine learning to provide and curate feedback.
Here the focus is on the application of machine learning to the feedback pathway---signals from the device to the user.
This is an important step to enabling bi-directional communication between agents in order to achieve strong collaborative interactions.First, this dissertation champions viewing a direct-to-body device such as an upper-limb prosthesis as a partner.
It outlines the value of viewing the interaction as a partnership and introduces a framework for the value and evaluation of strong partnerships between a human and a machine.
Following that, a set of experiments demonstrate the ability of real-time machine feedback learning methods to learn something of value to a human user in a prosthetic domain for the first time.
These experiments also show that machine-learned feedback can be successfully acquired and adapted in real time during human-robot interaction.
The final two experiments explore the human side of interactions with a device that is adapting the feedback it provides over time.
Evidence in this dissertation suggests that signals coming from devices that adapt as a user interacts with them could be the key to encouraging the user to engage more deeply with their device and initiate positive long-term interactions.
Along with these findings, the methods described in this dissertation for understanding the human side of human-prosthesis interaction are important contributions of this work.Overall this dissertation demonstrates for the first time the reasons and benefits of viewing upper-limb prostheses, and indeed many other assistive technologies, as partners to users rather than resigning them to simply being tools.
Temporal-difference learning methods are shown to be capable of learning and adapting feedback signals being sent to users in real-time and are a strong path towards closing the loop between humans and machines to enable collaborations.
Combined with methodological contributions that use rich data from both machines and humans, this dissertation proposes a shift in the way we think about human-device interaction in rehabilitation.
This shift towards thinking about and creating collaborations between humans and assistive technology, especially in upper-limb prostheses, holds the potential for unlocking greater functional gains wherever humans and devices come together. -
- Graduation date
- Fall 2024
-
- Type of Item
- Thesis
-
- Degree
- Doctor of Philosophy
-
- License
- This thesis is made available by the University of Alberta Library 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.