Learning steps: models of intent-driven lower limb prosthesis use
File(s)
Author(s)
Hodossy, Bálint
Type
Thesis
Abstract
Walking is often perceived as a simple and effortless process. However, it is generated by a complex and finely tuned neuromechanical system. Artificial devices that aim to seamlessly support or replace parts of it must adapt to the changing goals and environments of their users. Virtual environments provide an accessible and low-risk way to accelerate the early-stage development of these technologies. To capture the challenge inherent in this task, these simulations must aim to reproduce the complexities of our movement, our environment, and the dynamic control required assist it.
This thesis approaches this problem from two sides. First, we propose locomotion intent estimation models that regress biosignals to the desired future walking path. We investigate the sensitivity of our intent estimators to common error inducing factors such as electrode shifts, and propose ways to mitigate their effect. Second, we apply physics-based animation methods to create full-body motion corresponding to arbitrary walking path trajectories. We then use control policies optimised with reinforcement learning to trial prosthesis-assisted gait in non-steady-state locomotion settings. Conditioning the prosthesis' control on its user's walking intent is shown to mitigate the need for compensatory movements from the human agent.
The application of our gait synthesis agents to test prototype passive devices is also explored. We create a simulated model of a compliant prosthetic foot, and procedurally adapt our motion synthesis to tackle varying slopes and tripping hazards. We identify and quantify potential benefits provided by the compliant design and contrast it against data from an external experimental study.
In conclusion, we identify a series of challenges and solutions for dynamically simulating and evaluating intent-driven lower limb prosthetic interventions. Our methods contribute to a framework for refining assistive device design, facilitating the transfer of the next-generation assistive devices from the concept stage to real world systems.
This thesis approaches this problem from two sides. First, we propose locomotion intent estimation models that regress biosignals to the desired future walking path. We investigate the sensitivity of our intent estimators to common error inducing factors such as electrode shifts, and propose ways to mitigate their effect. Second, we apply physics-based animation methods to create full-body motion corresponding to arbitrary walking path trajectories. We then use control policies optimised with reinforcement learning to trial prosthesis-assisted gait in non-steady-state locomotion settings. Conditioning the prosthesis' control on its user's walking intent is shown to mitigate the need for compensatory movements from the human agent.
The application of our gait synthesis agents to test prototype passive devices is also explored. We create a simulated model of a compliant prosthetic foot, and procedurally adapt our motion synthesis to tackle varying slopes and tripping hazards. We identify and quantify potential benefits provided by the compliant design and contrast it against data from an external experimental study.
In conclusion, we identify a series of challenges and solutions for dynamically simulating and evaluating intent-driven lower limb prosthetic interventions. Our methods contribute to a framework for refining assistive device design, facilitating the transfer of the next-generation assistive devices from the concept stage to real world systems.
Version
Open Access
Date Issued
2024-11-01
Date Awarded
01/05/2025
License URL
Advisor
Farina, Dario
Sponsor
UKERC (Organization)
UK Research and Innovation
Grant Number
EP/S02249X/1
810346
Publisher Department
Department of Bioengineering
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)