Robots can be capable partners when interacting with humans, but their value is largely dependent on how information is communicated in that partnership. In physical human-robot interaction, information is communicated via motion---configurations, velocities, forces, and torques. The autonomy interprets these implicit signals using metrics, which ultimately drive the the autonomy. This thesis focus on how motion measures affect the performance of the closed loop controller and our ability to statistically characterize differences in motion due to deficit, assistance, and learning. Perhaps the most common way to implement a control solution is to include a feedback loop around the error with respect to a referent. In human motion, a single trajectory cannot capture all the possible solution strategies or variance in a single task. Therefore, I begin by describing a hybrid shared control that avoids specifying a time series of states by using methods from model predictive control to assess user action. The resulting controller improved training outcomes compared to unassisted practice and exhibited several features that are critical to learning in physical human-robot interaction (pHRI). Analysis of the study of the hybrid shared control showed that a measure of the information about the task encoded in the motion---ergodicity---was able to statistically capture the effect of assistance and training when error and task specific measures were not able to detect one or the other. I conclude this thesis by demonstrating how one could close the loop on this information measure, such that robot provides forceful feedback that supports the task goal rather than reduces local errors.

Creator

• Fitzsimons, Kathleen

DOI

• https://doi.org/10.21985/n2-tdgg-4m79

Subject

• Robotics
• Mechanical engineering

Language

• en

Alternate Identifier

• etdadmin_upload_781114
• http://dissertations.umi.com/northwestern:15423

Date created

• 2020-01-01

Resource type

• Dissertation

Rights statement

• In Copyright