Tools for Trustworthy Autonomy: Robust Predictions, Intuitive Control, and Optimized Interaction

Katherine Driggs Campbell

EECS Department
University of California, Berkeley
Technical Report No. UCB/EECS-2017-41
May 9, 2017

http://www2.eecs.berkeley.edu/Pubs/TechRpts/2017/EECS-2017-41.pdf

In the near future, robotics will impact nearly every aspect of life. Yet for technology to smoothly integrate into society, we need interactive systems to be well modeled and predictable; have robust decision making and control; and be trustworthy to improve cooperation and interaction. To achieve these goals, we propose taking a human-centered approach to ease the transition into human-dominated fields. In this work, our modeling methods and control schemes are validated through user studies in a realistic motion simulator and demonstrate improved interaction, predictability, and trustworthiness. Autonomous vehicles are a great motivating example, due to the wealth of interesting problems that arise with human-in-the-loop control and multi-agent interaction and cooperation. While autonomous vehicles will likely be publicly available soon, it can be assumed that the transition will not be instantaneous, suggesting that: (1) levels of autonomy will be introduced incrementally, and (2) autonomous vehicles will have to be capable of driving in a mixed environment, with both humans and autonomy. In both cases, the human drivers must be modeled in an accurate and precise manner that easily integrates into control frameworks.

We present a data-driven approach to hybrid system tools, that approximates the forward reachable set of a coupled human-robot system. This empirical reachable set is an alternative look at a classic control theoretic safety metric and allows us to predict driver behavior over long time horizons in a robust, yet informative, manner. This method is compared to an extension of traditional reachability, in which the optimal disturbances are uncovered using empirical metrics with probabilistic guarantees. Applications of this work include the design of minimally invasive intervention schemes for semi-autonomous vehicles and of planning nuanced interactions between humans and autonomy in interactive maneuvers. We also consider concerns that arise with shared control. Given a fixed semi-autonomous framework, we model the communication between the human and automation using information theory metrics. By controlling information flow, we observe an information/performance trade-off that follows a strongly concave relationship. This is formulated as an optimization paradigm, giving a model-based approach to interface design for optimizing interaction.

Advisor: Ruzena Bajcsy


BibTeX citation:

@phdthesis{Driggs Campbell:EECS-2017-41,
    Author = {Driggs Campbell, Katherine},
    Title = {Tools for Trustworthy Autonomy: Robust Predictions, Intuitive Control, and Optimized Interaction},
    School = {EECS Department, University of California, Berkeley},
    Year = {2017},
    Month = {May},
    URL = {http://www2.eecs.berkeley.edu/Pubs/TechRpts/2017/EECS-2017-41.html},
    Number = {UCB/EECS-2017-41},
    Abstract = {In the near future, robotics will impact nearly every aspect of life.  Yet for technology to smoothly integrate into society, we need interactive systems to be well modeled and predictable; have robust decision making and control; and be trustworthy to improve cooperation and interaction.  To achieve these goals, we propose taking a human-centered approach to ease the transition into human-dominated fields.  In this work, our modeling methods and control schemes are validated through user studies in a realistic motion simulator and demonstrate improved interaction, predictability, and trustworthiness.  Autonomous vehicles are a great motivating example, due to the wealth of interesting problems that arise with human-in-the-loop control and multi-agent interaction and cooperation.   While autonomous vehicles will likely be publicly available soon, it can be assumed that the transition will not be instantaneous, suggesting that: (1) levels of autonomy will be introduced incrementally, and (2) autonomous vehicles will have to be capable of driving in a mixed environment, with both humans and autonomy.  In both cases, the human drivers must be modeled in an accurate and precise manner that easily integrates into control frameworks. 

We present a data-driven approach to hybrid system tools, that approximates the forward reachable set of a coupled human-robot system.  This empirical reachable set is an alternative look at a classic control theoretic safety metric and allows us to predict driver behavior over long time horizons in a robust, yet informative, manner.  This method is compared to an extension of traditional reachability, in which the optimal disturbances are uncovered using empirical metrics with probabilistic guarantees.  Applications of this work include the design of minimally invasive intervention schemes for semi-autonomous vehicles and of planning nuanced interactions between humans and autonomy in interactive maneuvers.  We also consider concerns that arise with shared control.  Given a fixed semi-autonomous framework, we model the communication between the human and automation using information theory metrics.  By controlling information flow, we observe an information/performance trade-off that follows a strongly concave relationship.  This is formulated as an optimization paradigm, giving a model-based approach to interface design for optimizing interaction.}
}

EndNote citation:

%0 Thesis
%A Driggs Campbell, Katherine
%T Tools for Trustworthy Autonomy: Robust Predictions, Intuitive Control, and Optimized Interaction
%I EECS Department, University of California, Berkeley
%D 2017
%8 May 9
%@ UCB/EECS-2017-41
%U http://www2.eecs.berkeley.edu/Pubs/TechRpts/2017/EECS-2017-41.html
%F Driggs Campbell:EECS-2017-41