My dissertation focuses on human motor learning and adaptation and centers on two questions: 1) Is an online platform conducive for conducting motor adaptation experiments? and 2) How successfully can human learning drive performance in a complex human-robot collaboration task? For the first question, we developed a web-based experimental platform to facilitate online data collection for motor learning studies. An online platform has several benefits including eliminating the need for expensive and finely calibrated hardware, reducing the time-intensive demand for on-site researchers, and allowing for the testing of a representative human population. However, there is a vital need to understand general feasibility and the considerations necessary to shift closely controlled human-subject experiments to an online setting. We conducted a validation study and a case study examining remotely collected data quality for an 80-minute experiment. We were able to replicate known motor learning phenomena and highlighted several specific challenges associated with online data collection. Most notably, we found that experiment durations > 60 minutes were susceptible to fatigue and compliance issues as there was a significant effect of time (trial number) on our results. However, we believe that targeted future development and careful execution will enable motor learning research to take advantage of online experimental capabilities. For the second question, we examined human performance in a human-robot collaboration task. The field is currently lacking in research on the human contribution to human-robot collaborations. Therefore, we developed a complex "trimanual" or 3-limb task where the human was solely responsible for driving performance. In addition to the two biological arms, a third robotic arm was controlled through electromyography (EMG) or muscle electrical signals. We evaluated task and coordination performance, providing a systematic evaluation of the differences between two muscle sites selected for EMG, the leg (tibialis anterior) and the head (temporalis muscle). Results demonstrated that subjects were able to show statistically significant improvement over all three primary metrics with a tolerable workload that was trending towards satisfactory by the conclusion of the experiment. Results also showed that ipsilateral limb coordination may be a limiter in three-limb coordination. Importantly, there was no statistical difference based on muscle site indicating promise for a variety of potential applications.

This dissertation explores how reported human trust in autonomous agents changes over time. Given a task with a shared goal, a human learns information about their collaborating partner through many factors. Autonomous systems can query this information through a measure of human trust. Discovering how much a human trusts an autonomous agent sheds light on how they may interact with them in future interactions, opening an avenue for anticipatory action planning.

We explore several questions in this work, such as: How does human trust evolve as a human interacts with an agent that changes strategy and capability? Can a pared-down anytime trust measurement survey capture trusting events? How may future autonomous systems enable decision-making in isolated space environments?

To answer the first question, we built a web browser game and infrastructure to administer a collaborative visual search task. Human subjects controlled a spotlight cursor to search a 2-dimensional grid of possible points of interest while an autonomous agent simultaneously searched the same area. The autonomous agent used three different search strategies and had three different capabilities. These combinations were presented to subjects with different ordering schemes. We found that subjects' trust in the autonomous agents evolved differently depending on the order.

Next, in a remote supervision task, subjects used a web browser to interact with a simulated robot arm on a simulated version of the International Space Station. Subjects were instructed to sort stowage to an assigned configuration using the robot and rate their trust whenever they felt their trust in the robot had changed. We found that subjects rated their trust shortly following a stowage placement, indicating they used information from the task outcome to rate their trust. This context enabled an anytime trust slider to identify a factor a subject used to rate their trust without a lengthy, repetitive survey.

The aim of my dissertation was to investigate the command and control of machines by humans. There have been many control methods to interface with machines, and this dissertation focused on surface electromyography (sEMG). Surface EMG is the measurement of electrical signals produced by muscle(s) and measured at the skin’s surface. Several EMG techniques rely on multiple sensors and machine learning for multi-DOF (degree of freedom) or multi-command control and are therefore more complex and may be sensitive to signal degradation. The first goal was to develop a robust, single-site sEMG control methodology that could communicate more than one command from a single muscle site or EMG sensor. A computer-based cursor-to-target task assessed the performance of the single-site sEMG for two levels of control: auto-rotate (automatic cursor rotation and user-controlled cursor forward motion) and manual rotate (user-controlled cursor rotation and forward motion). The experiment demonstrated that the auto-rotate method led to better throughput, or the rate of selecting targets adjusted for difficulty, but not significantly better path efficiency. However, the subjects were able to learn the manual-rotate method where a single-site sEMG control method communicated two commands. The manual rotate method appeared viable for expansion from two to four commands while maintaining a single sensor, but it was unclear how to best train subjects to use the system. The subsequent experiment investigated the effects of different training methodologies on performance, cognitive workload, and trust. The augmented feedback training techniques led to early and sustained performance gains with lower cognitive workload and higher trust. Even with extensive training to learn the commands, subjects did not perfectly perform the four sEMG commands. A subsequent analysis revealed that adjusting the sEMG command parameters may help personalize the command system to the individuals and improve command performance. The knowledge gained from these studies culminated in a three-handed coordination pilot study with a sEMG-controlled, collaborative robot serving as the third hand. Subjects learned how to improve their coordination of the three “hands.” The pilot study also served to inform future experiments that will continue to integrate the human with a collaborative robot for more complex tasks.

Understanding speech in noisy environments requires more than good hearing. Everyday acoustic scenes are complex and dynamic – far more than most laboratory or clinical settings – thus they extract a heavy cognitive toll. For instance, when following a conversation, we must exert sustained effort while continually switching attention among different talkers. However, despite the tremendous importance of these mechanisms for developing hearing interventions, we know relatively little about them. The following research has three primary aims: to design, build, and test a new class of attentional prosthetic platform as a potential aid to those with hearing loss; to characterize brain signals (EEG) that predict and track the locus of selective spatial attention during conversational turn taking; and to demonstrate the functional relationships between selective attention and listening effort that robustly signal listener fatigue.