The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part Bl. Beijing 2008
provides vital information on what the obstacle is doing. Third,
and most importantly, what is the predicted path of the obstacle.
All of these functions depend heavily on highly accurate pose
estimation.
Estimation of state and feature extraction all depend on accurate
pose estimation. As mentioned previously, errors in pitch and
roll of only 0.5 degrees can result in false characterization of
terrain and obstacles. This is more critical in sensing obstacles
far away rather than close to the vehicle. As the course
presented obstacles in rapid succession the robots required
accurate pose estimation to avoid colliding with them. However,
errors in roll and pitch are more pronounced over longer
distances and higher speeds. The absolute vertical error
increases as the pitch error angle expands over the range of the
sensor.
One of the key elements which determined success in this
Urban Challenge was real time situational awareness and data
fusion. Such a capability required two levels of characterization,
that of the robotic vehicle in relation to the road and the
dynamic obstacles on it. The challenge is illustrated in figure
10 where the robot is sensing the way to a waypoint, but
encounters traffic around it. The vehicle must not only track
and predict where it will go, but it must do this while tracking
within its lane, sensing the terrain (road radius of curvature,
grade / cross fall) to ensure any maneuvers are within the
performance envelope and actually predict where the obstacle
will move to. In the previous Grand Challenge robots had a
choice of path candidates (in the Red Team example given
previously, an onboard computer generates ‘s’ splines or
multiple path candidates immediately adjacent to the intended
path of travel, all which are viable alternate routes taking into
consideration the vehicle’s dynamic state). Here, the path
candidates around an obstacle need to be able to change rapidly
and the vehicle will do most of the thinking.
6. URBAN GRAND CHALLENGE LESSONS
APPLIED TO REAL WORLD SCENARIOS
The goal of the Urban Grand Challenge is to apply the various
technologies employed to successfully navigate the course to
real world problems. Looking at this competition at its most
fundamental level, these autonomous vehicles are mobile
mapping platforms. The advances made here have significant
implications for how mobile mapping data is used. Consider
the automotive industry for example. Currently, GPS is utilized
as a convenience feature utilizing GPS, map matching and
odometer data to route a driver (albeit not very accurately)
through GPS outages. When looking at position and orientation
data in terms of driver assistance / active safety systems, the
accuracy required changes dramatically. Data needs to be
thought of in a layered approach for this application much like
the data fusion discussed above. Base maps utilized by onboard
computers need to be very accurate for sensors to determine
dynamics in relation to a vehicle’s current and predicted path so
the vehicle can determine if a driver is making turns at unsafe
speeds or passing through an intersection without stopping. By
having detailed maps along with accurate position and
orientation data, vehicles will be able to actively ensure the
safety of passengers.
Military applications present another example of how vehicle
automation saves lives. The Pentagon is aiming to have one
third of its forces automated by 2015. This applies to combat
forces as well as re-supply elements. Mobile mapping will
become particularly automated in this field and employ several
layers of data from different sources to achieve a particular
mission. For example, UAVs employing LIDAR and other
sensors will provide up to date intelligence for automated
ground convoys traveling through hostile terrain. Ground
vehicles utilizing their own LIDAR and optical sensors will
map their way to an objective relying on accurate base maps
and accurate position and orientation data.
5.
RESULTS OF THE RACE
7.
SUMMARY
The DARPA Urban Grand Challenge took place in Victorville
California at George AFB. The National Qualifying Event
(NQE) saw thirty six Teams participate in a number of rounds
designed to illustrate the requisite skills required to successfully
complete the three DARPA missions. Of the thirty six Teams,
eleven were qualified to participate in the final race on
November 3 rd , 2007. Of the eleven Teams, only six managed to
finish all three missions. Applanix Corporation partnered with
Tartan Racing, Stanford Racing and MIT to secure first, second
and fourth place finishes.
Team
Name
ID#
Vehicle
Type
Time
Taken
(h:m:s)
Result
Tartan
Racing
19
Boss
2007 Chevy
Tahoe
4:10:20
1st Place; averaged approximately
14 mph (22.53 km/h) throughout
the course
Stanford
Racing
03
Junior
2006
Volkswagen
Passat Wagon
4:29:28
2nd Place; averaged about 13.7
mph (22.05 km/h) throughout the
course
VictorTango 32
Odin
2005 Ford
Hybrid Escape
4:36:38
3rd Place: averaged 13 mph (20.92
km/h) throughout the course
MIT
79
Talos
Land Rover LR3
6:00:00
4th Place.
The Ben
Franklin
Racing
Team
74
Little Ben
2006 Toyota
Prius
No
official
time.
Finished
Cornell
26
Skynet
2007 Chevy
Tahoe
No
official
time.
Finished
Figure 11: DARPA Urban Grand Challenge Results
Accurate and reliable position and orientation data is a
fundamental part of autonomous vehicle guidance and control.
What we have shown is that even small errors in pose
estimation can lead to erroneous terrain characterization which
impacts vehicle performance. The significance of accuracy was
highlighted in the Urban Grand Challenge where dynamic
obstacles and terrain characterization in adverse GPS
environments were key skills that robots demonstrated in order
to successfully navigate the course and complete the three
DARPA missions. Position and orientation data accuracy was
essential to win the race which required sensor fusion and
precise vehicle dynamic control to interact with a constant
changing environment. These core elements will revolutionize
how we think about mobile mapping in general. The precise
location of roads, their geometry and roadside features will be
essential elements for vehicle guidance and control, not just
basic navigation. Accurate geospatial information and the real
time interpretation of that information are essential elements for
autonomous vehicles to demonstrate before such technology
becomes mainstream.
1241