IEEE Robotics & Automation Magazine - September 2018 - 99

search for samples. These action sequences were then passed
down to the execution layer to carry out the detailed robot
control. The hierarchical separation made the autonomy
software easier to develop and manage and provided the
modularity, flexibility, and robustness needed to make Cataglyphis functional.
Another major autonomy challenge was decision making
under high degrees of uncertainty posed by the interactions
between Cataglyphis and the natural environment. Anticipating behaviors to perform in all of the conditions that a robot
may encounter may be possible in a constrained, structured
environment, but this was not feasible for the SRR Challenge.
Cataglyphis needed to be able to autonomously handle situations
that would not exactly match any that were anticipated during
software development or experienced during previous testing.
If only incomplete or conflicting information was available,
Cataglyphis must make an educated guess at the next behavior
it should perform. Ideally, this guessed behavior would be targeted to gain more information and eliminate the confusion so
that a better-informed decision can be made next.
Cataglyphis had a total of 17 unique behaviors, such as
choosing new ROIs to search, searching an ROI, approaching
and collecting samples, and returning home to drop off samples. A set of primary decision heuristics were implemented
as condition flags to provide an overview of the robot's state
and its local environment and to enable the robot to respond
with an appropriate behavior. To reduce the chance that the
robot software would get stuck in an undefined state caused
by a condition that was not accounted for during development, a set of fallback heuristics were added to put Cataglyphis into an assumed set of safe behaviors. If the primary
heuristics could not produce a valid decision, the fallback
heuristics would tell Cataglyphis to do one of the following: 1)
if it was not lost and not possessing a sample, go to another
ROI; 2) if it was lost, perform maneuvers to search for home;
and 3) if it was possessing a sample and not lost, go home.
This allowed Cataglyphis to move forward with its mission.
With this approach, Cataglyphis was able to deal with more
complex situations in the final year of the challenge, as compared to the state machine approach used previously [2].
Cataglyphis spent most of its time during the SRR Challenge traversing across the field; thus, correctly choosing where
to go then driving there safely and efficiently were crucial for
success. Cataglyphis chose foraging locations dynamically
through a cost-benefit analysis, as opposed to following a predefined plan. An ROI with a high probability of containing a
sample provided a high reward. The distance to get to the location and the hazards along the way all represented costs. The
initial sample probability and hazard information was estimated by team members based on the provided satellite images.
After an ROI was searched, its probability of containing a sample was lowered so that, on the next planning step, other ROIs
would more likely be chosen. Once Cataglyphis decided where
it needed to go, it utilized the fast marching square (FM 2)
method [7] to decide on a route to safely traverse the field,
avoiding known obstacle locations. The method used a 1-m

resolution grid map containing traversability values based on
the provided satellite imagery and topographic map.
Additional prior knowledge was added to the path planner by
magnetic paths, which were corridors in the map known to be
safe, even under moderate robot navigation uncertainties (e.g.,
where the robot should be drawn toward as it crossed the field).
Magnetic paths provided sensible corridors through narrow
areas with cluttered obstacles that may have otherwise resulted in
meandering paths. Once an idealized path was planned with the
FM 2 planner, an additional step discretized the path into a small
number of straight line segments to make driving the path simpler. Figure 8 shows a path planned by Cataglyphis to traverse
from one end of the competition field to the other.
While driving toward these waypoints, Cataglyphis generated a 2-D occupancy grid (named the hazard map) of detected obstacles relative to the robot using lidar data. The local
hazard map was generated using a similar method as the creation of the 2-D local grid map in the SLAM approach but
with different filtering parameters prioritizing the detection of
traverse hazards instead of SLAM features. Additional obstacle
detection was performed using random sample consensus to
find holes or small protrusions in the ground [2]. If any obstacles were detected inside a 2-m-wide corridor within 5 m of
the robot, Cataglyphis slowed down in anticipation of an
avoidance maneuver. If an obstacle was detected within 3.5 m
of the robot, Cataglyphis attempted to maneuver around the
obstacle before resuming course to the original destination. In
this case, it selected a sequence of actions to perform from a
library of predesigned motion primitives [8]. The set of
maneuvers with the lowest cost not intersecting obstacles was
chosen to resume course toward the original destination. If
frequent obstacle avoidance requests were triggered within a
short driving distance, Cataglyphis gave up the likely unreachable waypoint and choose a new destination.
Field Tests and the 2016 Competition
Field tests were performed in various settings to validate system
performance, expose issues, and fine-tune the robot design.
One of the key objectives was to increase the mean time
between human interventions as a metric for evaluating the
robot autonomy. As Cataglyphis interacted with new environments, issues that it could not yet handle were encountered and
subsequently addressed. Another benefit of the thorough robot
testing was for human designers to achieve a better understanding of the robot's behavior, as Cataglyphis was fully autonomous and made its own judgment in each situation. It was
noted that Cataglyphis had a tendency to revisit nearby ROIs
that it had searched but was unable to find a sample in these
areas previously, as opposed to selecting unvisited ROIs at the
far ends of the field. This allowed the development of a comprehensive set of plans that could be used during communication
update periods to correct undesirable robot behaviors.
During the final SRR Challenge in 2016, Cataglyphis performed mostly as expected. It was allowed to run fully autonomously and uninterrupted for the first 80 min. During this
period, Cataglyphis collected four samples, worth a total of
september 2018

IEEE ROBOTICS & AUTOMATION MAGAZINE

IEEE Robotics & Automation Magazine - September 2018

Table of Contents for the Digital Edition of IEEE Robotics & Automation Magazine - September 2018