428
•
Journal of Field Robotics—2008
Despite these mistakes and a very capable field of
competitors, Boss qualified for the final event and
won the Urban Challenge.
2. BOSS
Boss is a 2007 Chevrolet Tahoe modified for au-
tonomous driving. Modifications were driven by the
need to provide computer control and also to support
safe and efficient testing of algorithms. Thus, modi-
fications can be classified into two categories: those
for automating the vehicle and those that made test-
ing either safer or easier. A commercial off-the-shelf
drive-by-wire system was integrated into Boss with
electric motors to turn the steering column, depress
the brake pedal, and shift the transmission. The third-
row seats and cargo area were replaced with electron-
ics racks, the steering was modified to remove excess
compliance, and the brakes were replaced to allow
faster braking and reduce heating.
Boss maintains normal human driving controls
(steering wheel and brake and gas pedals) so that
a safety driver can quickly and easily take control
during testing. Boss has its original seats in addi-
tion to a custom center console with power and net-
work outlets, which enable developers to power lap-
tops and other accessories, supporting longer and
more productive testing. A welded tube roll cage
was also installed to protect human occupants in the
event of a collision or rollover during testing. For un-
manned operation a safety radio is used to engage
autonomous driving, pause, or disable the vehicle.
Boss has two independent power buses. The
stock Tahoe power bus remains intact with its 12-
V dc battery and harnesses but with an upgraded
high-output alternator. An auxiliary 24-V dc power
system provides power for the autonomy hardware.
The auxiliary system consists of a belt-driven alter-
nator that charges a 24-V dc battery pack that is in-
verted to supply a 120-V ac bus. Shore power, in the
form of battery chargers, enables Boss to remain fully
powered when in the shop with the engine off. Ther-
mal control is maintained using the stock vehicle air-
conditioning system.
For computation, Boss uses a CompactPCI chas-
sis with 10 2.16-GHz Core2Duo processors, each
with 2 GB of memory and a pair of gigabit Ether-
net ports. Each computer boots off of a 4-GB flash
drive, reducing the likelihood of a disk failure. Two
of the machines also mount 500-GB hard drives for
data logging. Each computer is also time synchro-
nized through a custom pulse-per-second adaptor
board.
Boss uses a combination of sensors to provide
the redundancy and coverage necessary to navigate
safely in an urban environment. Active sensing is
used predominantly, as can be seen in Table I. The de-
cision to emphasize active sensing was primarily due
to the team’s skills and the belief that in the Urban
Challenge direct measurement of range and target
velocity was more important than getting richer, but
more difficult to interpret, data from a vision system.
The configuration of sensors on Boss is illustrated in
Figure 2. One of the novel aspects of this sensor con-
figuration is the pair of pointable sensor pods located
above the driver and front passenger doors. Each pod
contains an ARS 300 radar and ISF 172 LIDAR. By
pointing these pods, Boss can adjust its field of re-
gard to cover crossroads that may not otherwise be
observed by a fixed-sensor configuration.
3. MOTION PLANNING
The motion planning layer is responsible for execut-
ing the current motion goal issued from the behav-
iors layer. This goal may be a location within a road
lane when performing nominal on-road driving, a lo-
cation within a zone when traversing through a zone,
or any location in the environment when performing
error recovery. The motion planner constrains itself
based on the context of the goal to abide by the rules
of the road.
In all cases, the motion planner creates a path
toward the desired goal and then tracks this path
by generating a set of candidate trajectories that fol-
low the path to various degrees and selecting from
this set the best trajectory according to an evaluation
function. This evaluation function differs depending
on the context but includes consideration of static
and dynamic obstacles, curbs, speed, curvature, and
deviation from the path. The selected trajectory can
then be directly executed by the vehicle. For more
details on all aspects of the motion planner, see
Ferguson, Howard, and Likhachev (2008, submitted).
3.1. Trajectory Generation
A model-predictive trajectory generator originally
presented in Howard and Kelly (2007) is responsible
for generating dynamically feasible actions from an
initial state x to a desired terminal state. In general,
this algorithm can be applied to solve the problem
of generating a set of parameterized controls u(p,x)
Journal of Field Robotics DOI 10.1002/rob
