Adaptive and Autonomous Switching: Shared Control of Powered Prosthetic Arms Using Reinforcement Learning

Powered prosthetic arms with numerous controllable functions (i.e., grip patterns or movable joints) can be challenging to operate. Gated control— a common control method for myoelectric arms and other human-machine interfaces—allows users to select a function by switching through a static list of possible functions. However, switching between many controllable functions often entails significant time and cognitive effort on the part of the user when performing tasks. One way to decrease the number of switching interactions required of a user is to shift greater autonomy to the prosthetic device, thereby sharing the burden of control between the human and the machine. Previous work has demonstrated that reinforcement learning (RL), and specifically general value functions (GVFs), has the potential to reduce the time and switching cost of gated control methods. In the current work, we extend previous studies by advancing an RL method termed adaptive switching for use during real time control of a prosthetic arm. Adaptive switching uses contextual factors to build up predictions about the use of functions during a task. Based on these predictions, adaptive switching will continually optimize and change the order in which functions are presented to the user during switching. We also combine adaptive switching with another machine learning control method, termed autonomous switching, to further decrease the number of manual switching interactions required of a user. Autonomous switching uses predictions, learned in real time through the use of GVFs, to switch automatically between functions for the user. Over the course of several studies, we collected results from subjects with and without amputations, performing simple and more challenging tasks with a myoelectric robot arm. As a first contribution of this thesis, we present