Advancing the Acceptance and Use of Wheelchair-mounted Robotic Manipulators
Laura Petrich
Supervisor: Dr. Martin Jägersand
Laura Petrich
Supervisor: Dr. Martin Jägersand
Robotic assistive technologies focus on using robots to maintain and improve functional capabilities by assisting individuals living with disabilities and those who need extra help due to aging. The overarching goal of this thesis is to promote the acceptance and use of wheelchair-mounted assistive robotic manipulators in order to improve functional independence.
This site serves as a point of entry for access to the most up-to-date information and version of my thesis.
To access a PDF of my thesis, please click here.
Wheelchair-mounted robotic manipulators have the potential to help the elderly and individuals living with disabilities carry out their activities of daily living independently. While robotics researchers focus on assistive tasks from the perspective of various control schemes and motion types, health research tends to concentrate on clinical assessment and rehabilitation. This difference in perspective often leads to the design and evaluation of experimental tasks that are tailored to specific robotic capabilities rather than solving tasks that would support independent living.
In addition, there are many studies in healthcare on which activities are relevant to functional independence, but little is known about how often these activities occur. Understanding which activities are frequently carried out during the day can help guide the development and prioritization of assistive robotic technology. By leveraging the strength of robotics (i.e., performing well on repeated tasks) these activities can be automated, significantly improving the quality of life for our target population.
Our first research goal is to investigate daily task frequency data in order to provide deeper insights and meaningful guidelines for future research developments in the field of assistive robotic manipulation. These guidelines are meant to shift focus towards better supporting the needs and performance requirements of the target population. While we have established that assistive robotic manipulators can help individuals regain functional independence, their performance is restricted by the underlying control system. Aside from having direct control over the robots motion, each manipulator comes with a limited preprogrammed set of modes (e.g., drinking mode) that might not fully encompass the individual's needs. For mainstream use and acceptance, it is paramount to provide the target population with a way to refine the control and customize it to their personal needs. One method for learning new behaviours and skills on the fly is interactive shaping.
Interactive shaping is a method of transferring knowledge from a human to a learning agent by having the human teacher provide signals of approval (or disapproval) from instantaneous observations of the robots behaviour. This could be of great value in an assistive setting as it does not require the teacher to have expert domain knowledge. The teacher only needs to give feedback on whether the robot's previous action was "good" or "bad" in order to teach it new behaviours and skills, regardless of task difficulty. Our second research goal is to investigate whether interactive shaping can be used to teach robotic manipulators new autonomous tasks. To this end, we adapt TAMER, a framework for learning from human reward signals, to a seven degree of freedom robotic manipulator and carry out a proof-of-concept user study.
The work in this thesis is meant to open an avenue to better align research with the needs of individuals that will eventually leverage the technology in their daily life. To do so, we first bridge the gap between robotics and healthcare research to align both with respect to the target population's needs. Taking it one step further, we then adapt TAMER to allow the users themselves to add new autonomous behaviours to their wheelchair-mounted robotic manipulator. Together, these results introduce a new level of long-term autonomy for individuals living with disabilities.
Chapter 3 of this thesis is based on work which was published and presented as Petrich, L., Jin, J., Dehghan, M., and Jägersand, M., “A Quantitative Analysis of Activities of Daily Living: Insights into Improving Functional Independence with Assistive Robotics,” in 2022 IEEE International Conference on Robotics and Automation (ICRA). Please see our accompaniment video to the left and published paper.
The World Health Organization developed the World Health Organization Disability Assessment Schedule (WHODAS2.0) from the International Classification of Functioning, Disability and Health as a standardized, cross-cultural measure of functioning and disability across all life domains. WHODAS2.0 is used to measure the impact of health conditions, monitor intervention effectiveness, and estimate the burden of physical and mental disorders across all major life domains. This graph shows the key life domains of functioning with physical manipulation activities relevant to robotics highlighted in bold.
Our frequency analysis illustrates the most common daily living tasks found in the NTCIR video data (red bars) and MIT internet of things sensor data (blue bars). Some events are detected more reliably by the embedded sensors used in the MIT dataset, while others show up only in the image data. For example, sensors detect quick events more reliably than the low frame rate lifelogging video data. In contrast, the MIT sensor dataset is limited to detecting daily tasks only where sensors are placed, therefore missing events (e.g., outdoor activities) that are captured in the video data. By combining results from both datasets, we were able to obtain a more accurate quantitative measure of daily task significance.
High priority daily living activities that would have a large impact on the target community. The qualitative column reflects task priority preferences stated by the target population in surveys. The quantitative column highlights key results from our life-logging data analysis reflecting tasks that occur frequently throughout the day.
In this proof of concept work, we are interested in investigating whether the TAMER framework can be extended to high-dimensional robot manipulators. This is an alternative solution to the reward shaping problem prevalent in robotic reinforcement learning. The goal is to provide end users with a simple, yet reliable, method of teaching new autonomous behaviours to learning agents in the real world while carrying out tasks online.
The average number of mode switches participants made during trials of the reaching task. The black bar is the standard deviation of the number of mode switches.
These results support our hypothesis that as the input dimensionality increases, the number of times the user needs to switch modes while controlling the robotic arm decreases. One of the benefits of the TAMER control interface is that it removes the need to switch between modes (i.e., the number of mode switches is zero across all trials).
The number of user inputs required on average to complete the task. The black bar represents the standard deviation. Results suggest that training a TAMER agent requires far fewer user inputs in order to complete the task.
The average completion time for the reaching experiment. The black bar represents the standard deviation. Results show that task completion time decreases as the input dimensionality increases, reflecting a difference in control difficulty. The TAMER agent takes longer to complete the reaching task, although some trials performed comparably with the one DoF control interface. This is likely due to the exploratory nature of the TAMER agent and we hypothesize that task completion time will improve over time as it learns more of the state space.
We gratefully acknowledge the support of the Natural Sciences and Engineering Research Council of Canada (NSERC) and Alberta Innovates.
The University of Alberta, its buildings, labs, and research stations are primarily located on the traditional territory of Cree, Blackfoot, Métis, Nakota Sioux, Iroquois, Dene, and Ojibway/Saulteaux/Anishinaabe nations; lands that are now known as part of Treaties 6, 7, and 8 and homeland of the Métis. The University of Alberta respects the sovereignty, lands, histories, languages, knowledge systems, and cultures of First Nations, Métis and Inuit nations.