environments. These challenges include mechanisms
to support autonomous operations over extended
periods of time, techniques that facilitate the use of
human assistance in learning and decision making,
learning to reduce the reliance on humans over time,
addressing the practical scalability of existing methods,
relaxing unrealistic assumptions, and alleviating safety
concerns about deploying these systems.
The AAAI fall symposium on Reasoning and
Learning in Real-World Systems for Long-Term Autonomy consisted of 18 paper presentations and 3
invited talks, concluding in a lively and interactive
panel discussion moderated by Joydeep Biswas. In
total, there were 12 long papers and 6 short papers,
ranging in topic from planning and learning to architectures and real-world systems. The three invited
talks by Nick Hawes, Maarten Sierhuis, and Peter
Wurman presented work with fully operational deployments of assistant service robots, semiautonomous vehicles, and large-scale multiagent warehouse
robots, respectively. The techniques leveraged across
the papers and talks included hierarchical and mul-tiobjective (PO)MDP models; reinforcement learning
with general value functions; deep learning for
grasping, environment understanding, and risk-aware
planning; multiagent models for system robustness;
and robotic architectures with a focus on tight component integration. Applications included autonomous vehicles, delivery robots, activity recognition
in smart homes, mobile warehouse robots, air traffic
surveillance, dual-arm grasping robots, and mobile
Throughout the symposium, four key themes
emerged as topics for long-term autonomy research:
( 1) integration of multiple AI components beyond
traditional architectural, hierarchical, and multi-objective approaches; ( 2) methods to proactively leverage humans to overcome any exceptional issues
encountered, diminishing this reliance over time; ( 3)
standard metrics and verification methods to properly
measure the effectiveness of long-term autonomous
agents, such as by their exceptional issues encountered, number of human help requests, effect of system improvements made, and degree of learning
performed; and ( 4) a focus on the robustness of the
system to enable these long-term deployments.
Conclusions drawn during the panel discussion at
the end of the symposium suggested long-term autonomous systems benefit greatly from a symbiotic
collaboration with other connected agents, humans,
and a cloud-based AI central support system. To be
sufficiently robust also requires a much tighter development of the theoretical frameworks with the
implementation itself. Finally, objectively evaluating
the system continuously over time and across many
metrics is crucial, both as it is developed and as it
autonomously learns. Evaluation can be done using
general metrics, such as how many tasks were completed, which tasks were completed and their completion
times, how many failures occurred, and what kinds of
failures occurred and their failure times.
In addition, specific domain-related metrics can
improve this measurement, such as how far an autonomous vehicle has driven or how many objects per
day a robot grasped. Such an array of metrics allow us
to confirm that the holistic AI system is robust and
capable of long-term autonomy.
The symposium was organized by Kyle Hollins
Wray (chair), Julie A. Shah, Peter Stone, Stefan J.
Witwicki, and Shlomo Zilberstein. The papers from
the symposium were published by the University of
Massachusetts Amherst as Technical Report UM-CS-
Aaron Adler is a senior scientist at BBN Technologies in
Prithviraj Dasgupta is a professor in the Computer Science
Department at the University of Nebraska, Omaha.
Nick DePalma is a research engineer at Samsung Research of
Mohammed Eslami is the chief data scientist at Netrias in
Richard G. Freedman is a researcher at Smart Information
Flow Technologies (SIFT) and a PhD candidate at the University of Massachusetts Amherst.
John E. Laird is the John Tishman Professor of Engineering
in the Division of Computer Science and Engineering at the
University of Michigan.
Christian Lebiere is a research faculty member in the Psychology Department at Carnegie Mellon University.
Katrin Lohan is an associate professor of computer science at
Ross Mead is the founder and CEO of Semio AI.
Mark Roberts is a researcher at the US Naval Research
Paul S. Rosenbloom is a professor in the Department of
Computer Science and director for cognitive architecture
research at the Institute for Creative Technologies at the
University of Southern California.
Emmanuel Senft is a research fellow at the University of
Plymouth, United Kingdom.
Frank Stein is the director of the A3 Center at IBM.
Tom Williams is an assistant professor of computer science
at the Colorado School of Mines.
Kyle Hollins Wray is a graduate student in the College of
Information and Computer Sciences at the University of
Fusun Yaman is a senior scientist at BBN Technologies in
Shlomo Zilberstein is a professor and associate dean of research and engagement in the College of Information and
Computer Sciences at the University of Massachusetts