The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part Bl. Beijing 2008
1240
own viable path candidates using the interpolated path as a
reference. Utilizing this approach, control points, and spline
angle vectors that adjust to specify the location and orientation
of a path.
4. THE 2007 DARPA URBAN GRAND
CHALLENGE - THE CHALLENGE OF A DYNAMIC
ENVIRONMENT
The 2005 Grand Challenge pushed participants to develop
solutions for terrain perception and obstacle avoidance which
required obstacle detection and avoidance at ranges of 40m
directly in front of the vehicle. This next iteration of the Grand
Challenge in an urban environment pushed the state of the art
not only in perception, but object prediction and autonomous
vehicle interaction in dynamic environments. Teams needed to
contend with spurious GPS conditions in some areas of the
course, ensuring the robots followed the rules of the road and
interaction with 50 manned vehicles simulating traffic along
with the other robots attempting to finish DARPA’s three
mandated mission tests. Each of these missions tested different
core skills (parking, traffic collision avoidance, driving
precision) and needed to be completed in the shortest amount of
time. The intention of each of the missions was to test how the
vehicles interacted with situations human drivers encounter on a
regular basis. For example, there were several intersections
where four vehicles were waiting at a stop light. The
autonomous vehicle needed to know not only when it was safe
to proceed, but deal with other vehicles which may malfunction
and need to get out of the way while taking into consideration
other traffic and the rules of the road. This required a new
breed of sensors which provided the vehicle with a 360
representation of its surroundings. Such capability is shown in
Figure 9.
Figure 9: 360 degree scanning LIDAR for situational awareness
The Teams demanded very high performance from their
positioning and orientation systems. Pose estimation was
critical to perception, planning, control and providing key data
to the drive-by-wire systems of autonomous vehicles. As
demonstrated in the following architecture from Tartan Racing,
accurate position and orientation estimation was essential to the
perception and world modeling routines constructed by the
robot. As discussed previously, data fusion is a key factor in
determining how successful the robot is in characterizing and
interacting within its environment to achieve its mission. Pose
estimation is provided to the behavior generation and motion
planning routines which are bounded by the mission planning
parameters programmed into the vehicle. In each phase of the
mission, the robot needed to integrate the composite
representation of the world and understand what were safe and
unsafe maneuvers given the changing targets around it impeding
its route.
Given the missions and skills needed to be demonstrated by the
robots, the benefits of pre-planning were not as profound as
with the 2005 Challenge. Teams did require a substantial
amount of data up-sampling from the sparse points provided in
the DARPA RNDF file. However, instead of having 2 hours to
prepare the vehicle, Teams only had 15 minutes between
missions to prepare for the next portion of the race. Route
planning was absolutely critical to finish the missions in the
least amount of time, however the missions required much more
processing of real time obstacle avoidance rather than following
exact waypoints. This required high bandwidth, low latency
data to be constantly available to the system especially for
dynamic data fusion. Detecting a static obstacle is a simple
process of determining where the target is located, what the lane
corridor as defined by the Robot’s sensors as compared to the
RNDF is and what is the safest speed and steering angle around
the obstacle to avoid it.
Figure 10: Tartan Racing Architecture [9]
Figure 10: Vehicle Tracking and Prediction
When the obstacle is dynamic, there are three fundamental
challenges. The first is reliable position tracking relative to
where the vehicle is and where it needs to go (in the local
coordinate). Second, with accurate range and target bearing the
robot can determine what lane the obstacle is in from the route
network definition file (RNDF) or if it is off the road. This