DTIC ADA522603: Autonomous Driving in Traffic: Boss and the Urban Challenge

1 MB·English

by  Defense Technical Information Center

Articles
Autonomous Driving in Traffic:
Boss and the Urban Challenge
Chris Urmson, Chris Baker, John Dolan, Paul Rybski, Bryan Salesky,
William “Red” Whittaker, Dave Ferguson, and Michael Darms
I
n The DARPA Urban Challenge was a compe-
tition to develop autonomous vehicles capable
of safely, reliably, and robustly driving in traf- n 2003 the Defense Advanced Research Projects Agency
fic. In this article we introduce Boss, the (DARPA) announced the first Grand Challenge with the goal of
autonomous vehicle that won the challenge. developing vehicles capable of autonomously navigating desert
Boss is a complex artificially intelligent soft-
trails and roads at high speeds. The competition was generated
ware system embodied in a 2007 Chevy Tahoe.
as a response to a mandate from the United States Congress that
To navigate safely, the vehicle builds a model of
a third of U.S. military ground vehicles be unmanned by 2015.
the world around it in real time. This model is
used to generate safe routes and motion plans To achieve this goal DARPA drew from inspirational successes
both on roads and in unstructured zones. An such as Charles Lindbergh’s May 1927 solo transatlantic flight
essential part of Boss’s success stems from its aboard the Spirit of St. Louis and the May 1997 IBM Deep Blue
ability to safely handle both abnormal situa- chess victory over reigning world chess champion Garry Kas-
tions and system glitches. parov. Instead of conventional funded research, these successes
were organized as challenges with monetary prizes for achieving
specific goals.
DARPA’s intent was not only to develop technology but also
to rally the field of robotics and ignite interest in science and
engineering on a scale comparable to the Apollo program. Thus,
the Grand Challenges were born. Organized as competitions
with cash prizes awarded to the winners ($1 million in 2004; $2
million in 2005; and $2 million for first, $1 million for second,
and $500,000 for third in 2007). For the first two years the gov-
ernment provided no research funding to teams, forcing them
to be resourceful in funding their endeavours. For the 2007
Urban Challenge, some research funding was provided to teams
through research contracts. By providing limited funding, the
competition encouraged teams to find corporate sponsorship,
planting the seeds for strong future relationships between aca-
demia (the source of many of the teams) and industry.
The Grand Challenge1 was framed as a cross-country race
where vehicles had to drive a prescribed route in the minimum
time. There would be no moving obstacles on the course, and
no one was allowed to predrive the course. To complete the
Copyright © 2009, Association for the Advancement of Artificial Intelligence. All rights reserved. ISSN 0738-4602 SUMMER 2009 17
Report Documentation Page Form Approved
OMB No. 0704-0188
Public reporting burden for the collection of information is estimated to average 1 hour per response, including the time for reviewing instructions, searching existing data sources, gathering and
maintaining the data needed, and completing and reviewing the collection of information. Send comments regarding this burden estimate or any other aspect of this collection of information,
including suggestions for reducing this burden, to Washington Headquarters Services, Directorate for Information Operations and Reports, 1215 Jefferson Davis Highway, Suite 1204, Arlington
VA 22202-4302. Respondents should be aware that notwithstanding any other provision of law, no person shall be subject to a penalty for failing to comply with a collection of information if it

does not display a currently valid OMB control number.
1. REPORT DATE 3. DATES COVERED
2009 2. REPORT TYPE 00-00-2009 to 00-00-2009
4. TITLE AND SUBTITLE 5a. CONTRACT NUMBER
Autonomous Driving in Traffic: Boss and the Urban Challenge
5b. GRANT NUMBER
5c. PROGRAM ELEMENT NUMBER
6. AUTHOR(S) 5d. PROJECT NUMBER
5e. TASK NUMBER
5f. WORK UNIT NUMBER
7. PERFORMING ORGANIZATION NAME(S) AND ADDRESS(ES) 8. PERFORMING ORGANIZATION
Carnegie Mellon University,Robotics Institute,Pittsburgh,PA,15213 REPORT NUMBER
9. SPONSORING/MONITORING AGENCY NAME(S) AND ADDRESS(ES) 10. SPONSOR/MONITOR’S ACRONYM(S)
11. SPONSOR/MONITOR’S REPORT
NUMBER(S)
12. DISTRIBUTION/AVAILABILITY STATEMENT
Approved for public release; distribution unlimited
13. SUPPLEMENTARY NOTES
14. ABSTRACT

15. SUBJECT TERMS
16. SECURITY CLASSIFICATION OF: 17. LIMITATION OF 18. NUMBER 19a. NAME OF
ABSTRACT OF PAGES RESPONSIBLE PERSON
a. REPORT b. ABSTRACT c. THIS PAGE Same as 12
unclassified unclassified unclassified Report (SAR)
Standard Form 298 (Rev. 8-98)
Prescribed by ANSI Std Z39-18
Articles
Figure 1. The Urban Challenge Final Event.
Eleven teams qualified for the Urban Challenge final event. Eighty-nine teams from the United States and countries as far away as China
registered for the competition.
challenge, the autonomous vehicles would need to world to advance intelligent, autonomous vehicle
drive the route in less than 10 hours. The compe- technology (figure 1). Competitors were required
tition was scoped to require driving both on and to develop full-size autonomous vehicles cable of
off road, through underpasses, and even through driving with moving traffic in an urban environ-
standing water. These diverse challenges lead to a ment at speeds up to 30 miles per hour (mph). The
field with a wide variety of vehicles, ranging from vehicles were required to demonstrate capabilities
military high-mobility multipurpose wheeled that were undeniably intelligent, including obey-
vehicles (HMMWVs) and sport utility vehicles ing traffic rules while driving along roads, negoti-
(SUVs) to motorcycles and custom-built dune bug- ating intersections, merging with moving traffic,
gies. parking, and independently handling abnormal
The first running of the Grand Challenge situations.
occurred on March 3, 2004. The 142-mile course
The Urban Challenge was fundamentally about
was too difficult for the field, with the best vehicle
intelligent software. The most visible artifacts of
traveling 7.4 miles. Though described as a failure
the research and development are the numerous
in the popular press, the competition was a tech-
autonomous vehicles that were entered in the
nical success; though the vehicles drove a small
competition, but the magic is really in the algo-
fraction of the course, the top competitors did so at
rithms that make these vehicles drive. The compe-
high, sustained speeds that had not been previ-
tition was a 60-mile race through urban traffic; a
ously demonstrated. Given the excitement and
drive that many people face daily. Thus the chal-
promise generated by this first challenge, the com-
lenge was not in developing a mechanism that
petition was rerun in 2005.
could survive these rigors (most automotive man-
The 2005 Grand Challenge began at 6:30 AMon
ufacturers do this already), but in developing the
October 8. Vehicles launched at intervals from
algorithms and software that would robustly han-
three start chutes in the Mojave desert. Approxi-
dle driving, even during unexpected situations. To
mately 7 hours later, the first of five finishers com-
complete the challenge, vehicles had to consis-
pleted the challenge. Stanley (Thrun et al. 2006),
tently display safe and correct driving behavior;
from Stanford University, finished first, in 6 hours
any unsafe or illegal maneuver could result in elim-
and 53 minutes, followed soon after by the two
ination from the competition.
Carnegie Mellon robots, Sandstorm and
In this article we present an overview of Boss,
H1ghlander (Urmson et al. 2006). Kat-5, funded by
the Gray Insurance Company, and TerraMax, a the vehicle that won the Urban Challenge (figure
30,000-pound behemoth developed by the 2). Boss was developed by a team of researchers
OshKosh Truck Corporation, rounded out the fin- and students from Carnegie Mellon University,
ishers. With five vehicles completing the course, General Motors, Caterpillar, Continental, and
the 2005 Grand Challenge redefined the percep- Intel. Boss drives at speeds up to 30 mph by mon-
tion of the capability of autonomous mobile itoring its location and environment and con-
robots. stantly updating its motion plan and the model of
The DARPA 2007 Urban Challenge2 continued the world around it. Throughout this article we
the excitement generated by the previous chal- provide pointers to other publications that delve
lenges, engaging researchers from around the into each area in greater depth.
18 AI MAGAZINE
Articles
Figure 2. Boss.
Boss uses a combination of 17 sensors to navigate safely in traffic at speeds up to 30 mph.
The Vehicle driving controls, and pressing a button reverts it to
a human-driven vehicle.
Boss is a 2007 Chevrolet Tahoe modified for Boss integrates a combination of 17 different
autonomous driving. While it was built from the sensors (a mixture of light detection and ranging
chassis up for the Urban Challenge, it incorporat- [lidar], radio detection and ranging [radar], and
ed many lessons learned from the previous chal- global positioning system [GPS]). These sensors are
lenges. Like most of the vehicles that competed in mounted to provide a full 360 degree field of view
the Urban, we built Boss from a standard chassis, around the vehicle. In many cases, sensor field of
to reduce mechanical complexity and benefit from views overlap, providing redundancy and in some
the reliability of a highly engineered platform. The cases complementary information. The primary
Tahoe provides ample room to mount sensors, sensor is a Velodyne HDL-64 lidar, mounted on the
computers, and other components while main- top of the vehicle. This lidar provides dense range
taining seating for four, which is important for safe data around the vehicle. Three Sick laser range
and efficient testing and development. Computer finders are used to provide a virtual bumper
control of the vehicle is made possible by using around the vehicle while two other long-range
electric motors to actuate the brake pedal, gear lidars are used to track vehicles at ranges of up to
selector, and steering column and through con- 100 m. Three Continental ACC radars are mount-
troller area network (CAN) bus communication ed in the front (two) and rear (one) bumpers to
with the engine control module. Despite these detect and track moving vehicles at long ranges. In
changes, the Tahoe maintains its standard human addition, a pair of Continental sensors (lidar and
SUMMER 2009 19
Articles
Mission Planning
n
o
e on ati
hicl epti Behavioral Executive bstr
e c A
V Per al
o
G
Motion Planning
Sensor Data Abstraction
Figure 3. The Boss Software Architecture.
The software architecture consists of a perception layer that provides a model of the world to a classical three-tier con-
trol architecture.
radar) are mounted on pointable bases, enabling be reliable and robust and able to run in real time.
Boss to “turn its head” to look long distances down Boss’s software architecture is fundamentally a dis-
cross streets and merging roads. To determine its tributed system, enabling decoupled development
location, Boss uses a combination of GPS and iner- and testing. In all, 60 processes combine to imple-
tial data fused by an Applanix POS-LV system. In ment Boss’s driving behavior. Architecturally these
addition to this system, Boss combines data from a are clustered into four major components: percep-
pair of down looking SICK laser range finders to tion, mission planning, behavioral execution, and
localize itself relative to lane markings described in motion planning (figure 3).
its on-board map. The number of sensors incorpo- The perception component interfaces to Boss’s
rated into Boss provides a great deal of information numerous sensors and fuses the data into a coher-
to the perception system. ent world model. The world model includes the
For computation, Boss uses a CompactPCI chas- current state of the vehicle, the current and future
sis with ten 2.16 gigahertz Core2Duo processors, positions of other vehicles, static obstacles, and the
each with 2 gigabytes of memory and a pair of location of the road. The perception system uses a
gigabit Ethernet ports. Each computer boots off of variety of probabilistic techniques in generating
a 4 gigabyte flash drive, reducing the likelihood of these estimates.
a disk failure. Two of the machines also mount 500 The mission planning component computes the
gigabyte hard drives for data logging, and all are fastest route through the road network to reach the
time synchronized through a custom pulse-per- next checkpoint in the mission, based on knowl-
second adapter board. The computer cluster pro- edge of road blockages, speed limits, and the nom-
vides ample processing power, so we were able to inal time required to make special maneuvers such
focus our efforts on algorithm development rather as lane changes or U-turns. The route is actually a
than software optimizations. policy generated by performing gradient descent
on a value function calculated over the road net-
Software Architecture work. Thus the current best route can be extracted
at any time.
To drive safely in traffic, the software system must The behavioral executive executes the mission
20 AI MAGAZINE
Articles
plan, accounting for traffic and other features in its motion; it is an interpretation of the current
the world (Ferguson et al. 2008). It encodes the state of the world around the vehicle.
rules of the roads and how and when these rules Each dynamic obstacle is represented by a series
should be violated. It also implements an error of oriented rectangles and velocities. Each rectan-
recovery system that is responsible for handling gle represents the state, or predicted state, of the
off-nominal occurrences, such as crowded inter- obstacle at an instant in time over the next few sec-
sections or disabled vehicles. The executive is onds. Additionally, each obstacle can be flagged as
implemented as a family of state machines that either moving or not moving and as either observed
translate the mission plan into a sequence of moving or not observed moving. An obstacle is
motion planning goals appropriate to the current flagged as moving if its instantaneous velocity esti-
driving context and conditions. mate is above some minimum noise threshold.
The motion planning component takes the The obstacle is not flagged as observed moving
motion goal from the behavioral executive and until the instantaneous velocity remains above a
generates and executes a trajectory that will safely noise floor for some threshold time. Similarly, a
drive Boss toward this goal, as described in the fol- moving obstacle does not transition from observed
lowing section. Two broad contexts for motion moving to not observed moving until its instanta-
planning exist: on-road driving and unstructured neous velocity remains below threshold for some
driving. large amount of time. These flags provide a hys-
teretic semantic state that is used in the behavioral
executive to improve behavior while maintaining
Software Components
the availability of full fidelity state information for
In this section we describe several of the key soft- motion planning and other modules.
ware modules (world model, moving obstacle Static obstacles are represented in a grid of regu-
detection and tracking, motion planning, intersec- larly spaced cells. The advantages of this represen-
tion handling, and error recovery) that enable Boss tation are that it provides sufficient information
to drive intelligently in traffic. for motion planning and is relatively inexpensive
to compute and query. The disadvantages are that
World Model
this simple representation makes no attempt to
Boss’s world model is capable of representing envi- classify the type of obstacle and uses a relatively
ronments containing both static and moving large amount of memory. To generate the map,
obstacles. While this representation was developed range data is first segmented into ground and non-
for the Urban Challenge, it can be also used more ground returns and then accumulated in a map
broadly for general autonomous driving. The using an occupancy gridlike algorithm. To main-
world model includes the road network, static tain a stable world representation, dynamic obsta-
obstacles, tracked vehicles, and the position and cles that are flagged as not observed moving and
velocity of Boss. The road network defines where not moving are painted into the map as obstacles.
and how vehicles are allowed to drive, traffic rules Once a dynamic obstacle achieves the state of
(such as speed limits), and the intersections observed moving, the cells in the map that it over-
between roads. The road representation captures laps are cleared. In the process of building the stat-
sufficient information to describe the shape (the ic obstacle map, an instantaneous obstacle map is
number of lanes, lane widths, and geometry) and generated. This map contains all static and dynam-
the rules governing driving on that road (direction ic obstacles (that is, data associated to dynamic
of travel, lane markings, stop lines, and so on). obstacle hypotheses with the observed moving flag
Static obstacles are those that do not move dur- set are not removed). This map is much noisier
ing some observation period. This includes obsta- than the static obstacle map and is used only for
cles off and on road. Examples of static obstacles validating potential moving obstacle tracks within
include buildings, traffic cones, and parked cars. In the dynamic obstacle tracking algorithms.
contrast, dynamic obstacles are defined as objects
Moving Obstacle Detection and Tracking
that may move during an observation period. With
this definition all vehicles participating actively in The moving obstacle detection and tracking algo-
traffic are dynamic obstacles even if they are tem- rithms fuse sensor data to generate a list of any
porarily stopped at an intersection. By these defi- dynamic obstacles in a scene (Darms, Rybski, and
nitions an object is either static or dynamic; it Urmson 2008). The tracking system is divided into
should not be represented in both classes. An two layers: a sensor layer and a fusion layer (see fig-
object can, however, change from static to dynam- ure 4). The sensor layer includes modules that are
ic and vice versa: a person may, for example, get specialized for each class of sensor and runs a mod-
into a parked car and actively begin participating ule for each physical sensor on the vehicle. The
in traffic. The information about whether an sensor layer modules perform sensor-specific tasks
object is static or dynamic is not tied exclusively to such as feature extraction, validation, association,
SUMMER 2009 21
Articles
Object Hypothesis Set
Road World
Fusion Layer
Model and
Object Management Global Target Validation Instantaneous
Estimation and Prediction RWM Checking Map
Model Selection
Global Classification
Measurement
Validated
(Observations, Proposals, Features
Features
Movement Observation)
Sensor Layer
Local Classification and Proposal Generation
Association
Local Target Validation
Feature Extraction
Figure 4. The Moving Obstacle Tracking System Is Decomposed into Fusion and Sensor Layers.
and proposal and observation generation. Each of using the road network and driving context (for
these tasks relies on sensor-specific knowledge, long time intervals).
generally encoded as heuristics; these are the
Motion Planning
important details that enable the machinery of the
Kalman filter to work correctly. For example, radar Motion planning is responsible for moving the
sensors may generate a specific class of false echoes vehicle safely through the environment. To do
that can be heuristically filtered, but the filtering this, it first creates a path toward the goal, then
approach would be irrelevant for a lidar or even a tracks it by generating a set of candidate trajecto-
different class of radar. By encapsulating these sen- ries following the path to varying degrees and
sor-specific details within sensor modules we were selecting the best collision-free trajectory from this
able to modularize the algorithms and maintain a set. During on-road navigation, the motion plan-
clean, general implementation in the fusion layer. ner generates its nominal path along the center
The fusion layer combines data from each of the line of the desired lane (Ferguson, Howard, and
sensor modules into the dynamic obstacle list, Likhachev 2008a). A set of local goals are selected
which is communicated to other components. The at a fixed longitudinal distance down this center
fusion layer performs a global validation of sensor line path, varying in lateral offset to provide sever-
observations, selects an appropriate tracking mod- al options for the planner. Then, a model-based
el, incorporates measurements into the model and trajectory generation algorithm (Howard et al.
predicts future states of each obstacle. The global 2008) is used to compute dynamically feasible tra-
validation step uses the road network and instan- jectories to these local goals. The resulting trajec-
taneous obstacle map to validate measurements tories are evaluated against their proximity to stat-
generated by the sensor layer. Any observation ic and dynamic obstacles, their distance from the
either too far from a road, or without supporting center line path, their smoothness, and various
data in the instantaneous obstacle map, is discard- other metrics to determine the best trajectory for
ed. For each remaining observation, an appropriate execution by the vehicle. This technique is suffi-
tracking model is selected, and the state estimate ciently fast and produces smooth, high-speed tra-
for each obstacle is updated. The future state of jectories for the vehicle, but it can fail to produce
each moving obstacle is generated by either extrap- feasible trajectories when confronted with aberrant
olating its track (over short time intervals) or by scenarios, such as blocked lanes or extremely tight
22 AI MAGAZINE
Articles
turns. In these cases, the error recovery system is added to the pool of front bumpers and is treated
called on to rescue the system. in the same manner to ensure equitable estimation
When driving in unstructured areas the motion of arrival times. Figure 5 shows a typical occupan-
planner’s goal is to achieve a specific pose (Fergu- cy polygon extending three meters back along the
son, Howard, and Likhachev 2008b). To achieve a lane from the stop line. Once the polygon becomes
pose, a lattice planner searches over vehicle posi- occupied the time is recorded and used to estimate
tion, orientation, and velocity to generate a precedence. To improve robustness of the dynam-
sequence of feasible maneuvers that are collision ic obstacle list, the occupancy state is not cleared
free with respect to static and dynamic obstacles. instantly when there is no obstacle in the polygon.
This planner is much more powerful and flexible Instead, the polygon remains occupied for a small
than the on-road planner, but it also requires more amount of time. Should an obstacle reappear in
computational resources and is limited to speeds the polygon in this time, the occupancy state of
below 15 mph. The recovery system leverages this the polygon is maintained. This allows moving
flexibility to navigate congested intersections, per- obstacles to flicker out of existence for short inter-
form difficult U-turns, and circumvent or over- vals without disturbing the precedence ordering.
come obstacles that block the progress of the vehi- In cases where Boss stops at an intersection and
cle. In these situations, the lattice planner is none of the other vehicles move, Boss will wait 10
typically biased to avoid unsafe areas, such as seconds before breaking the deadlock. In this case
oncoming traffic lanes, but this bias can be the precedence estimator assumes that the other
removed as the situation requires. traffic at the intersection is not taking precedence
and asserts that Boss has precedence. In these situ-
Intersection Handling
ations other elements of the behavioral executive
Correctly handling precedence at intersections is limit the robot’s maximum speed to 5 mph, caus-
one function of the behavioral executive (Baker ing it to proceed cautiously through the intersec-
and Dolan 2008). Within the executive, the prece- tion. Once the robot has begun moving through
dence estimator has primary responsibility for the intersection, there is no facility for aborting the
ensuring that Boss is able to safely and robustly maneuver, and the system instead relies on the
handle intersections. The precedence estimator motion planner’s collision-avoidance algorithms
encapsulates all of the algorithms for determining and, ultimately, the intersection recovery algo-
when it is safe to proceed through an intersection. rithms to guarantee safe, forward progress should
Before entering the intersection, the intersection another robot enter the intersection at the same
must be clear (both instantaneously and through- time.
out the time it predicts Boss will take to travel
Error Recovery
through the intersection) and Boss must have
precedence. The precedence order is determined Despite best efforts to implement a correct and reli-
using typical driving rules (that is, the first to arrive able system, there will always be bugs in the imple-
has precedence through the intersection and in mentation and unforeseen and possibly illegal
cases of simultaneous arrival, yield to the right). behavior by other vehicles in the environment.
The precedence estimator uses knowledge of the The error recovery system was designed to ensure
vehicle’s route, to determine which intersection that Boss handles these situations and “never gives
should be monitored, and the world model, which up.” To this end, a good recovery algorithm should
includes dynamic obstacles around the robot, to generate a sequence of nonrepeating motion goals
calculate the precedence ordering at the upcoming that gradually accept greater risk; it should adapt
intersection. To ensure robustness to temporary behavior to the driving context, since each context
noise in the dynamic obstacle list, the algorithm includes road rules that should be adhered to if
for computing precedence does not make use of possible and disregarded only if necessary in a
any information beyond the geometric properties recovery situation; and finally, the recovery system
of the intersection and the instantaneous position should also be kept as simple as possible to reduce
of each dynamic obstacle. Instead of assigning the likelihood of introducing further software
precedence to specific vehicles, precedence is bugs.
assigned to individual stop lines at an intersection, Our approach is to use a small family of error
using a geometric notion of “occupancy” to deter- recovery modes and maintain a recovery level.
mine arrival times. Polygons are computed around Each recovery mode corresponds to a nominal
each stop line, extending backward along the driving context (driving down a lane, handling an
incoming lane by a configurable distance and intersection or maneuvering in an unstructured
exceeding the lane geometry by a configurable area). When in normal operation, the recovery lev-
margin. If the front bumper of a moving obstacle el is zero, but as failures occur, it is incremented.
is inside this polygon, it is considered an occupant The recovery level is used to encode the level of
of the associated stop line. Boss’s front bumper is risk that is acceptable. It is important to note that
SUMMER 2009 23
Articles
Stop Line Occupancy Polygon
m
1
m
1
1m 3m 1m
Intersection Lane
Figure 5. A Typical Occupancy Polygon for a Stopline.
the recovery goal-selection process does not explic- er-risk maneuvers. For example, if there is a lane
itly incorporate any surrounding obstacle data, available in the opposing direction, the road is
making it intrinsically robust to transient environ- eventually marked as blocked, and the system
mental effects or perception artifacts that may issues a U-turn goal (goal 5) and attempts to follow
cause spurious failures in other subsystems. an alternate path to the goal. The intersection
All recovery maneuvers are specified as a set of recovery algorithm involves attempting to navi-
poses to be achieved and utilize the movement gate alternate paths through an intersection and
zone lattice planner to generate paths through then alternative routes through the road network.
potentially complicated situations. The successful The desired path through an intersection is a
completion of any recovery goal returns the sys- smooth interpolation between the incoming and
tem to normal operation, resetting the recovery outgoing lanes that may be blocked by some obsta-
level. Should the same, or a very similar, normal cle in the intersection or may be infeasible to drive.
goal fail immediately after a seemingly successful Thus, the first recovery level (goal 1 in figure 6b) is
recovery sequence, the previous recovery level is to try to achieve a pose slightly beyond the origi-
reinstated instead of simply incrementing from nal goal using the lattice planner, with the whole
zero to one. This bypasses the recovery levels that intersection as its workspace. This quickly recovers
presumably failed to get the system out of trouble the system from small or spurious failures in inter-
and immediately selects a goal at a higher level. sections and compensates for intersections with
The on-road recovery algorithm involves gener- turns that are tighter than the vehicle can easily
ating goals at incrementally greater distances down make. If the first recovery goal fails, alternate
the current road lane to some maximum distance. routes out of the intersection (goals 2, 3, 4) are
These forward goals (goals 1, 2, 3 in figure 6a) are attempted until all alternate routes to the goal are
constrained to remain in the original lane of trav- exhausted. Thereafter, the system removes the con-
el. If none of the forward goals succeed, a goal is straint of staying in the intersection and selects
selected a short distance behind the vehicle (goal goals incrementally further out from the intersec-
4) with the intent of backing up to provide a new tion (goals 5, 6).
perspective. After backing up, the sequence of for- The unstructured zone recovery procedure varies
ward goals is allowed to repeat once more with an depending on the current objective. Initially, the
offset, after which continued failure triggers high- algorithm selects a sequence of goals in a regular,
24 AI MAGAZINE
Articles
a
5
4 1 2 3
b Original
6 5 1 Goal 2
4 3
c
6
Original
3 Goal
7 1
4
8 2
5
9
Figure 6. Boss Uses Context-Specific Recovery Goals to Provide Robustness
to Software Bugs and Abnormal Behavior of Other Vehicles.
triangular pattern, centered on the original goal as the algorithm tries first to use alternative exits, if
shown in figure 6(c). If, after attempting this com- those exits fail, then the algorithm issues goals to
plete pattern, the vehicle continues to fail to make locations on the road network outside of the park-
progress, a more specialized recovery process is ing lot, giving the lattice planner free reign in gen-
invoked. For parking spot goals, the pattern is con- erating a solution out of the parking lot.
tinuously repeated with small incremental angular An independent last-ditch recovery system
offsets, based on the assumption that the parking monitors the vehicle’s motion. If the vehicle does
spot will eventually be correctly perceived as emp- not move at least one meter every five minutes, it
ty, or the obstacle (temporarily) in the spot will overrides any other goals with a randomly selected
move. If the current goal is to exit the parking lot, local motion. The hope is that some small local
SUMMER 2009 25