Advanced robot hand myoelectric control strategies for improved human-robot interaction

Human hands -and human beings, in general- are the gold standards for researchers in the domain of robot hand design and control. Consequently, electromyographic signals represent one of the primary tools adopted in human-in-the-loop robot hand control applications. Indeed, myoelectric control has the potential to enable a natural and seamless interaction between humans and robots, thus promoting embodiment and sense of agency over the robotic device. However, the longstanding obstacles faced by myoelectric control solutions are still in place. Therefore, this thesis presents the research work carried out to address some of these challenges. The main contributions regard the development of advanced myoelectric control strategies for improved, natural, and intuitive human-robot-environment interaction. The proposed solutions suggest improvements to the myoelectric control loop from two different perspectives. Well-known challenges of intent estimation approaches (e.g., the need for point-by-point dataset labeling and the difficulty of performing nonlinear fitting) are addressed through the use of new tools, like the dynamic time warping algorithm, to build a minimally supervised regression paradigm, or through the adaptation of old tools, like non-negative matrix factorization, in a self-supervised fashion. Moreover, feedback information is exploited to design a shared autonomy framework that leverages a probabilistic approach to handle the uncertainties and variability of humans, robots, and the environment during grip strength regulation applications. Finally, a small parenthesis addresses the problem of human-robot-environment interaction in household settings from a complementary standpoint, outlining end-effector design requirements and relative quantitative analysis.