Active Intent Disambiguation, Control Interpretation, and Arbitration for Assistive Robotics

Edakkattil Gopinath, Deepak

doi:https://doi.org/10.21985/n2-mgta-q804

Work

Active Intent Disambiguation, Control Interpretation, and Arbitration for Assistive Robotics

Public

Download PDF

Download All Files (.zip)

The goal of this dissertation is to develop models, algorithms, and interaction protocols to improve the efficacy and quality of Human-Autonomy Interaction (HAI) in the domain of assistive robotics. In this domain, the most common control paradigm is that of manual teleoperation using control interfaces such as joysticks, switch-based head arrays, and sip-and-puff. However, manual teleoperation can become physically and cognitively burdensome to the human due to the limitations of the control interface, inherent complexities of the assistive machine, and motor impairments. Although introducing full robotics autonomy could be a viable approach to ameliorate these challenges, a more attractive control paradigm is that of shared autonomy in which the human and the autonomous agent share control responsibilities, thereby ensuring that the human still has agency. This dissertation focuses on two important aspects of a shared-autonomy assistive system, namely, intent inference and control arbitration. In a shared autonomy human-robot team, the autonomous agent's ability to infer human intent accurately is critical for providing correct and timely assistance to the human. However, due to the sparsity, low information content, and imperfections in the control signals, accurate intent inference is rather difficult. To improve the autonomous agent's ability to infer user intent this dissertation introduces the idea of intent disambiguation. Algorithms and protocols for intent disambiguation are designed to alter or nudge the human's decision-making context in a principled manner so that the signals generated by the human contain more information regarding underlying intent. Intent disambiguation algorithms endow the autonomous agent with active learning capabilities. More meaningful and informative teleoperation signals from the human enables the autonomous agent to perform accurate intent inference which in turn improves the assistance provided to the human. In Chapter 5, a heuristic approach for intent disambiguation that reasons over the space of control modes is introduced. Building upon this idea, in Chapter 6, a more rigorous formalism of intent disambiguation that is grounded in information-theoretic principles is presented. Another important topic that this dissertation addresses is the question of how to share robot control between the human and the autonomous agent. Upon successful user intent inference the autonomous agent can rely on different types of autonomous controllers to generate appropriate assistance towards the user's intended goal. Yet, in a shared control setting the autonomous control signal needs to be combined with the human's control signal in some fashion. Control sharing in a shared control assistive system lies on a continuum with full manual teleoperation on one end of the spectrum to fully autonomous solution on the other. How exactly should control be arbitrated between the human and the autonomous agent is a critical question that impacts the overall performance of the human-autonomy team. Implicitly, the question could be framed as a constrained optimization problem in which the optimization objective is to balance both task-related metrics (such as successful task completion, minimizing effort, and minimizing energy expenditure) and subjective metrics (such as satisfaction, sense of agency, and user's assistive preferences). To this end, in Chapter 7 we propose a human-in-the-loop solution to this constrained optimization problem in which the human uses an easy-to-understand, interpretable protocol to optimize the arbitration parameters according to their own optimality criteria to achieve their desired outcomes. In the assistive setting, robot teleoperation is facilitated using physical control interfaces. Autonomous agents benefit a great deal if they can make a distinction between conceptual and physical aspects of interface operation. To address this problem, this thesis introduces the notion of interface awareness into probabilistic models of interface-mediated robot teleoperation. In Chapter 4, the autonomous agent relies on an interface-aware robot teleoperation model to reason about user intent at the level of interface signals and then provide appropriate types of modifications and corrections to faulty or unintended interface operation that arise due to lack of motor skill, motor impairments or inherent noise in the physical interface. As a last contribution, Chapter 8 presents a software tool for conducting web-based crowdsourced human-robot interaction experiments. Data collection in the domain of assistive robotics can be an arduous task and the goal of this tool is to facilitate rapid prototyping and testing of novel algorithms developed for assistive autonomous agents. To summarize, the work presented in this dissertation tackles different challenges that arise in human-autonomy interaction in the context of interface-mediated assistive robot teleoperation and proposes algorithms and protocols to improve the overall human experience of interacting with an assistive robot.

Creator