exciting and encompassing theme for research in
motion planning, control, machine learning, perception, and human-robot interaction. Furthermore, its
applications extend human capabilities, to deep-sea
exploration, disaster recovery, rehabilitation, and
assistive care, enabling us to explore new possibilities
safely and efficiently. This symposium provided a
deep dive into the dual themes of fundamental theoretical algorithms for shared autonomy, as well as
their real-world practical application through a
unique partnership with the FDA. The goal of this
symposium was to build bridges between research
domains and also between researchers and practitioners, and to set the foundations for the future of
Through keynote and contributed talks, a variety
of different goals for shared autonomy were revealed.
One use of shared autonomy was as a tool to teach
human operators how to better perform a control
task. In rehabilitation robotics, this could be seen
through object avoidance assistance in powered
wheelchairs that was reduced as the operators
became more familiar with the wheelchair’s capabilities. In another application, operators were taught to
control a robot performing the classic inverted pendulum problem by rejecting operator inputs that
were not consistent with the automated system’s
optimal policy. Another use of shared autonomy was
to provide intelligent safety constraints in applications such as aircraft autopilots and surgical robotics.
In these situations, there may not be time for the AI
to communicate why a particular override is occurring, and it opened the floor for a discussion on creating “explainable” AI systems, and how transparent
the AI should be with the operator about its intentions.
Another major theme was how to evaluate shared
autonomy systems. Because shared autonomy often
uses learned customization to each operator, who in
turn has changing abilities over time, it is difficult to
assess effectiveness of the overall system in a standardized way. The most common assessment tool
being used is to lower bound performance by either
the human or system operating individually, and use
this baseline to compare improvements of the joint
system. Ensuring safety with a constantly evolving
and learning system was brought up during the panel discussion on commercialization of shared autonomy technologies. Evaluation is necessary for regulatory approval, as well as insurance coverage in the
case of medical devices. Safety evaluation is an open
challenge among all new AI technologies, from
autonomous driving to assistive wheelchairs and surgical robots. Developing metrics to evaluate how
closely operator intentions are being followed and
bounding the probabilities on undesirable outcomes
are important steps toward enabling shared autonomy systems to be broadly utilized.
The symposium participants discussed challenges
inherent in all shared autonomy systems where a
human and AI are interacting. These challenges
included managing operator expectations, maintain-
ing operator engagement at increasing levels of
autonomy, identifying the right moments to have
assertive AI, providing the AI with enough input to
determine relevant context, and the impact of auton-
omy on the operator’s trust of the system. This sym-
posium brought together a wide range of participants
from industry, regulatory bodies, and academe and
sparked discussions and collaborations that may nev-
er have formed otherwise due to the diverse nature of
shared autonomy applications.
Laura Herlant served as chair of this symposium.
The papers of the symposium were published by
AAAI Press as Technical Report FS-16-05 within the
compilation titled The 2016 AAAI Fall Symposium
Series: Technical Reports FS-16-01 – FS-16-05.
Patrícia Alves-Oliveira is a PhD student in psychology
applied to human-robot interaction at Instituto Univer-sitário de Lisboa and INESC-ID.
Richard G. Freedman is a PhD candidate in the College of
Information and Computer Sciences at the University of
Dan Grollman is the Robot Brain Architect of Sphero, Inc.
Laura Herlant is a PhD candidate at the Robotics Institute
of Carnegie Mellon University.
Laura Humphrey is a research engineer at the Aerospace
Systems Directorate of the Air Force Research Laboratory.
Fei Liu is an assistant professor in the Department of Computer Science at the University of Central Florida.
Ross Mead is the founder and CEO of Semio.
Frank Stein is the director of the Analytics Solution Center
Tom Williams is a PhD candidate in the Human-Robot
Interaction Lab at Tufts University.
Shomir Wilson is an assistant professor in the Electrical
Engineering and Computer Science Department of the University of Cincinnati.