Optimizing upper-limb prosthesis control

Curious Optimization: Flavor of the research

Life is based on interactions. Some of these interactions are between passive objects and people, like when you toss a ball. Some of these interactions are between intelligent objects and people, like when you control a prosthetic arm or cooperate with an assistive exoskeletal leg. Some of these interactions are based on drastically different time-scales of being, like when you interact with a tree. Some of these interactions are between complicated and intelligent individuals, like when you talk with a friend. And some of these interactions are between groups of individuals and groups of poeple, where information sharing becomes complicated and the rules of rationality mathematically break down. I'm interested in exploring how components of each of these systems can be efficiently optimized to make interactions better.

For example, think about a car that offers an environmentally friendly mode as well as a sports mode. Designers have essentially tuned different parameters to give the best response they can, but geared towards different values. Can we use a similar mentality to design a prosthesis control mapping that requires less effort and yet results in faster, more accurate control? Can we design better sequences of rehabilitation that rehabilitate people faster, in alignment with their individual skills and resources (such as how close or far they are from a physiotherapist)? Can we design better interactions in a complex web of healthcare services between allied health professionals? Can we design better laws or incentive programs that enable communities / stakeholders / nations to achieve goals they mutually aspire to achieve? Key aspects in all of these questions include 1) interactions are occuring, and often with delays between actions and consequences, 2) some parts of the system can be changed via tunable parameters or societal constructs, while others have to obey the laws of physics; and 3) at least one and typically more things are valued (such as time, effort, money, future generations, etc.).

Finding algorithms that can address realistic problems in numerically efficient ways is a key pillar of our team's research. We leverage numerical optimal control theory with recent advances in artificial intelligence and processing power to develop efficient solutions to realistic problems. One area that we're particularly excited about at the moment is Curious Optimal Control, or implicit dual control as it is formally called. Curious optimal control acts as an artificially intelligent agent, interacting with other stakeholders to learn more about the environment or the individual needs and preferences of the people it interacts with. Things get complicated fast, and so while curious optimal control has been considered for decades, the "curse of dimensionality" has typically limited its real-word potential. We have recently developed numerically efficient approaches to this appealing paradigm that run quickly and can handle real-life questions. We're excited to apply this paragidm to increasingly complicated ecosystems of questions