robots, then
focuses on lessons learned from each deployment. Finally,
this paper describes how this entire project came to a close,
offering a cautionary tale for those who wish to embark on
such an effort in the future.
Introduction


The history of autonomous mobile robotics research has
largely been a story of closely supervised, isolated
experiments on platforms which do not last long beyond
the end of the experiment. In January 1998, we and
others began work on Chips, an autonomous robot
intended to be more than a short experiment. Our goal
would be to install Chips as a permanent member of the
museum staff at the Carnegie Museum of Natural History
in Pittsburgh, Pennsylvania [1].
Shortly thereafter, Mobot, Inc. was incorporated with a
charter to improve and extend the Chips technology in a
series of robot installations. Following Chips, three more
robots were developed in succession; three of the four
operated every day until February 2002. Of the four
robots, three are museum robot installations, offering
visitors of various exhibit spaces augmented detail
regarding the exhibits at hand. Together, these three
museum robots have logged more than 2,500 days of
operation in separate, real-world public spaces.
In striving to deploy autonomous robots in a social
niche, we had two high-level goals. First, the robots must
be autonomous to the greatest extent possible. Human
supervision of a full-time robot is unacceptable. At most,
the robots should require only occasional human help, and
should request such help explicitly. Even the routine trip
to a battery charger should be performed by the robot
autonomously.
Secondly, since the robots would be deployed in public,
they must have sufficiently rich personalities to achieve
compelling and fruitful interaction with humans in their



environments. Note that we care not just about
compelling but about fruitful as well. These robots
have an educational charter and are therefore justified
only if they demonstrate real educational efficacy.
In the end of the day the robots did achieve some
measure of educational efficacy as well as long-term
robustness, but the social mission of the robots as well as
the commercial justification for such robotic endeavours
proved to be a challenge beyond our reach. These
autonomous robots are now off-line, as of April 2002, and
so, as the penultimate section of this paper explains, this
long-term experiment is now complete.
Robot Overview
The three robots described in this paper all same the same
basic motion platform (the holonomic Nomadic
Technologies XR4000 base); the same operating system
(RedHat Linux); and the same programming environment
(Gnu C++). The first robot of the series, Chips, began
work at the Carnegie Museum of Natural History on May
22, 1998. Chips operated exclusively in Dinosaur Hall,
which contains the large bone collections of T. Rex and
other massive dinosaurs as well as ancillary exhibits
focusing on topics such as paleogeology and ancient
aquatic life. Chips charter was to provide tours in
Dinosaur Hall, presenting audiovisual information
regarding both the large bone collections as well as the
less frequented, smaller exhibits. Until it was taken off-
line, Chips operated for just under 4 years, covering a
total travel distance exceeding 500 km.
The second robot, Sweetlips, conducted tours in the
Hall of North American Wildlife, also at the Carnegie
Museum of Natural History. This space is comprised of
dioramas, where preserved wildlife specimens are shown
in naturalistic settings. This portion of the museum has
extremely low visitor traffic, so Sweetlips charter was to
attract additional visitors and to then bring the static
dioramas to life using high-quality video footage of the
wildlife in their natural habitat.

Figure 1: Chips the Dinosaur Hall Robot

Sweetlips operated beginning May 19 1999 and covered a
total distance exceeding 185 km autonomously.
Figure 2: The Sweetlips robot near a diorama

The third robot, Joe Historybot, operated in the Atrium of
the Heinz History Center. Its mission was to welcome
visitors to this historical museum and to provide both
information and a tour of various permanent exhibits

Figure 3: Joe Historybot in the Heinz History Center

placed throughout the atrium. Joe provided historical
context in an entertaining multimedia format. The robot
also provided tutorials on topics such as speaking English
with a Pittsburgh accent. Joe began operations on July 8,
1999 and covered a total distance exceeding 162 km
during its total period on-line.
Lessons Learned
The underlying goals of compelling interaction and
maximal autonomy have remained constant throughout
the creation of all three robots. However, each
succeeding robot was the product of a complete re-design
phase based on lessons learned from previous robots.
Although some technical aspects remained unchanged,
such as the programming language and robot mobility
chassis, virtually all else evolved in an effort to improve
both the autonomy and interactivity efficacy of each
robot.
We are in the unique position of having an established
trajectory of real-world interactive social robots.

Studying the evolving lessons learned from these
installations may prove useful in uncovering information
that is valuable to future robot installations. In the
following two sections we present such lessons learned,
discriminating between the two top-level goals of
providing maximal autonomy and producing maximally
effective robot-human interaction.
On Robot Autonomy

The first requirement of a public robot is safety, both for
the general public and for the robot itself. At the heart of
the matter is the robots method for avoiding collisions,
which must be especially robust, since the robots operate
Figure 4: The fiducial docking landmark for Chips

without supervision. It is notable that the collision
avoidance code on these robots is the least changed code
over the course of their creation and installation. The
robots use 48 ultrasonic range-finding sensors to detect
both static and dynamic obstacles in the robots path.
This obstacle detection problem is simplified by both the
density of sonar sensors on the XR4000 and by the
structure of the navigable spaces specified to each robot.
Navigation is performed within polygons that serve as
highways. The robot plans to achieve position subgoals
that lie at the intersection of adjacent highways.

Furthermore the highways are specified as conservative,
open spaces within the museums hallways and open



areas. Given this particular abstraction of navigable
space, any short sonar reading can be treated as an
unexpected obstacle, greatly simplifying the detection
component of obstacle avoidance.
The obstacle avoidance algorithm is further simplified
by several factors. First, the XR4000 robot is truly
holonomic, constraining the avoidance algorithm to
changes in 2D trajectory alone, fully decoupling these
decisions from the robots angle which is independently
governed by the vision system to maintain visual sighting
of the fiducials at the end of hallways. Second, the speed
at which we chose to move these robots in such crowded
environments was sufficiently slow that dynamics would
not need to be considered for safe stopping and turning.
With these simplifications in mind, the obstacle
avoidance system was a simple case-based chain, taking
into account only the desired direction of travel and the
instantaneous sonar data. The case-based obstacle
avoidance technique achieves very high reliability due to
its simplicity and diagnostic transparency [2].
Because of the limited accuracy of sonar at close range,
the robots will occasionally become stuck when they
approach a wall too closely. Given the infrequency of
this failure mode (less than once per week), we feel the
increased peace of mind due to conservative motion
primitives was worth the price.
A second critical aspect of autonomy in our
unsupervised application is the ability to detect failure and
signal humans for help. Early in the development of the
Chips platform, we began using pagers, which the robot
was able to signal via electronic mail messages. Both
active requests for help were implemented as well as
regular, twice daily pages indicating healthy robot status,
enabling us to verify liveness throughout the week. Basic
failure detection is straightforward on Chips due to its
sensors, which provide almost full observability into its
own physical state in the environment. Encoders enable
direct measurement of actual distances traversed; the
vision system verifies continuously that the optical
fiducials remain in the field of view; and analog inputs
verify that the batteries charge/discharge performance
has not yet degraded.
Initially, Chips sent for help quickly, giving up as soon
as a failure was detected. Soon we began adding
diagnostic methods to reset subsystems that were not
functioning correctly. This evolved into a generic retry
method for diagnostics within our software architecture:
each time a task is performed, check the result for
validity. If the command