Date on Master's Thesis/Doctoral Dissertation

8-2021

Document Type

Doctoral Dissertation

Degree Name

Ph. D.

Department

Electrical and Computer Engineering

Degree Program

Electrical Engineering, PhD

Committee Chair

Popa, Dan O.

Committee Co-Chair (if applicable)

Inanc, Tamer

Committee Member

Inanc, Tamer

Committee Member

Nasraoui, Olfa

Committee Member

Welch, Karla

Author's Keywords

robotics; human-robot interaction; variable autonomy; planning algorithms

Abstract

As robotic agents become increasingly present in human environments, task completion rates during human-robot interaction has grown into an increasingly important topic of research. Safe collaborative robots executing tasks under human supervision often augment their perception and planning capabilities through traded or shared control schemes. However, such systems are often proscribed only at the most abstract level, with the meticulous details of implementation left to the designer's prerogative. Without a rigorous structure for implementing controls, the work of design is frequently left to ad hoc mechanism with only bespoke guarantees of systematic efficacy, if any such proof is forthcoming at all. Herein, I present two quantitatively defined models for implementing sliding-scale variable autonomy, in which levels of autonomy are determined by the relative efficacy of autonomous subroutines. I experimentally test the resulting Variable Autonomy Planning (VAP) algorithm and against a traditional traded control scheme in a pick-and-place task, and apply the Variable Autonomy Tasking algorithm to the implementation of a robot performing a complex sanitation task in real-world environs. Results show that prioritizing autonomy levels with higher success rates, as encoded into VAP, allows users to effectively and intuitively select optimal autonomy levels for efficient task completion. Further, the Pareto optimal design structure of the VAP+ algorithm allows for significant performance improvements to be made through intervention planning based on systematic input determining failure probabilities through sensorized measurements. This thesis describes the design, analysis, and implementation of these two algorithms, with a particular focus on the VAP+ algorithm. The core conceit is that they are methods for rigorously defining locally optimal plans for traded control being shared between a human and one or more autonomous processes. It is derived from an earlier algorithmic model, the VAP algorithm, developed to address the issue of rigorous, repeatable assignment of autonomy levels based on system data which provides guarantees on basis of the failure-rate sorting of paired autonomous and manual subtask achievement systems. Using only probability ranking to define levels of autonomy, the VAP algorithm is able to sort modules into optimizable ordered sets, but is limited to only solving sequential task assignments. By constructing a joint cost metric for the entire plan, and by implementing a back-to-front calculation scheme for this metric, it is possible for the VAP+ algorithm to generate optimal planning solutions which minimize the expected cost, as amortized over time, funds, accuracy, or any metric combination thereof. The algorithm is additionally very efficient, and able to perform on-line assessments of environmental changes to the conditional probabilities associated with plan choices, should a suitable model for determining these probabilities be present. This system, as a paired set of two algorithms and a design augmentation, form the VAP+ algorithm in full.

Share

COinS