Mechanics Based Design of Structures and Machines

An International Journal

ISSN: 1539-7734 (Print) 1539-7742 (Online) Journal homepage: https://www.tandfonline.com/loi/lmbd20

Obstacle Avoidance for Autonomous Mowing

Ben C. Creed , Aaron Arsenault , Steven A. Velinsky & Ty A. Lasky

To cite this article: Ben C. Creed , Aaron Arsenault , Steven A. Velinsky & Ty A. Lasky (2012)
Obstacle Avoidance for Autonomous Mowing, Mechanics Based Design of Structures and
Machines, 40:3, 334-348, DOI: 10.1080/15397734.2012.685441

To link to this article: https://doi.org/10.1080/15397734.2012.685441

Published online: 05 Jul 2012.

Submit your article to this journal

Article views: 307

View related articles

Citing articles: 2 View citing articles

Full Terms & Conditions of access and use can be found at

https://www.tandfonline.com/action/journalInformation?journalCode=lmbd20
Mechanics Based Design of Structures and Machines, 40: 334–348, 2012
Copyright © Taylor & Francis Group, LLC
ISSN: 1539-7734 print/1539-7742 online
DOI: 10.1080/15397734.2012.685441

OBSTACLE AVOIDANCE FOR AUTONOMOUS MOWING#

Ben C. Creed, Aaron Arsenault, Steven A. Velinsky,

and Ty A. Lasky
Department of Mechanical and Aerospace Engineering,
University of California, Davis, California, USA

This paper discusses obstacle avoidance for autonomous mowing for application in
highway vegetation control. In recent work, a unique autonomous mower testbed
was developed to prove the efficacy of such an approach. The testbed has been
augmented with laser range finding to support obstacle detection. This paper presents
experimental results for reactive obstacle avoidance using the pure-pursuit geometric
control algorithm. The results show that the laser-based obstacle detection and
avoidance is effective for the mowing application.

Keywords: Automation; Highway maintenance; Landscaping; Obstacle avoidance; Robotics.

INTRODUCTION
The California Department of Transportation (Caltrans) is responsible for the
sustained maintenance of over 24,000 km (15,000 miles) of highway and more than
93,000 ha (230,000 acres) of right-of-way (Caltrans, 1997). This includes vegetation
management, which makes up approximately 40% of Caltrans maintenance budget
(Teeter-Balin, 2004). Primary vegetation control needs include maintenance of
roadbed integrity, visibility, drainage, noxious weed control, aesthetic roadside
appearance, and fire protection (Caltrans, 1997). Mowing of medians and rights-of-
way is an important vegetation management practice for Caltrans, but it is labor-
intensive and requires expensive and specialized equipment.
Previous work presented the application environment and explored scenarios
and their effectiveness for autonomous mowing (Arsenault et al., 2010). A separate
paper investigated the performance, particularly mowing width consistency, of
the autonomous mower (Arsenault et al., 2011). The current work extends these
efforts by adding laser-based obstacle detection and avoidance. The performance
evaluation of a pure-pursuit geometric controller for reactive obstacle avoidance in
the context of mowing is presented.

Received January 30, 2012; Accepted April 12, 2012

#
Communicated by B. Ravani.
Correspondence: Steven A. Velinsky, Department of Mechanical and Aerospace Engineering,
University of California, One Shields Avenue, Davis, CA 95616, USA; E-mail: [email protected]

334
 OBSTACLE AVOIDANCE FOR AUTONOMOUS MOWING 335

THE PROTOTYPE AUTONOMOUS MOWER TEST PLATFORM

A proof-of-concept, fully functional testbed was designed and built. This
section concisely summarizes the features of the testbed. The interested reader
is referred to Arsenault et al. (2011) and Arsenault (2007) for details of the
system including modeling, sensor fusion, hardware implementation, and control
architecture.
Although the testbed was developed as a research tool, its size, configuration,
and layout have been designed to be consistent with an eventual commercial unit.
Also, the testbed was designed so that it could be easily outfitted with an array of
sensor and hardware technology. Original thoughts were to modify a pre-existing
platform. The Robomower from Friendly Robotics was considered as a candidate.
However, this autonomous mowing unit was designed for homeowner lawns, and
was unsuitable for the rough terrain and harsh vegetation present in the roadside
environment. Chandler et al. (1999) modified a Toro rechargeable electric push
mower to serve as an autonomous mowing platform, but the platform had serious
mobility limitations. Most other studies aimed at developing autonomous mowing
technology have made use of riding lawn mowers with more success (see Aono et al.,
1998; Kiriy, 2002). Yet, these platforms tend to be large, heavy, costly, and unsafe,
often with limited maneuverability.
For the reasons listed above, a fully customized testbed/platform was
designed and built to avoid the inherent mechanical, cost, safety, and size
limitations of modifying a pre-designed platform. The platform has been designed to
succeed in the harsh roadside environment. Modular design techniques were used,
incorporating commercially available hardware such as motors, wheels, tubing,
fasteners, and other parts.
An autonomous mobile robot needs some mode of locomotion, such as wheels,
tracks, legs, etc. Although tank-like tracks were considered as a possible means for
locomotion, standard wheels were more practical and a differentially steered device
was selected as the configuration (Dudek, 2000). Differential drive systems are easy
to implement and yield highly maneuverable platforms. Also, such configurations
have been used with much success for existing commercial autonomous mowers, and
for a variety of riding lawnmowers; see, for example, Kiriy (2002). Finally, many
successful tracking control algorithms have been developed for differential drive
systems (Hong et al., 1999; Zhang et al., 1997).
Differential drives are made up of two independent drives, one for each side
of the vehicle. These drives typically consist of an actuator/motor (which may be
linked to a gearbox) and some type of wheel. The two wheels most often lie on
one common axis. Proper control of each independent motor yields all steering
and traversing motions. An electric power source was selected based on safety,
size, and cost limitations. DC permanent magnet motors were selected due to their
comparatively small, light, and efficient characteristics for a given power rating
(Rizzoni, 2005). Overall design specifications of the platform are summarized in
Table 1. Figure 1 shows the testbed.

KINEMATICS AND TRACKING CONTROL

A kinematic model is appropriate for low-speed, low-acceleration, and lightly-
loaded applications; a dynamic model is more appropriate for heavily-loaded,
336 CREED ET AL.

Table 1 Low-cost mowing platform requirements, configuration, and components

Requirements

Maximum speed 3.0 mph (4.8 km/h)

Operating speed 1.5 mph (2.4 km/h)
Acceleration To 1.5 mph (2.4 km/h) in less than 1.0 seconds
Torque Sufficient to overcome friction, gravity, and acceleration loads in the worst
case scenario
Energy Motors should operate at near optimum efficiency
Weight Less than 250 lb (1100 N)
Performance Drive up a 30-degree (0.52 radians) incline
Tool/cutting load Maximum of 25 lb (110 N)
Size Moderately sized mowing unit for easy transportation to the mowing site

Platform configuration, and commercial components

Vehicle type Differential drive

Drive wheel Pneumatic, Outside diameter 14 in (36 cm)
Caster wheel Pneumatic, Outside diameter 10 in (25 cm)
Overall length 40 in (101 cm)
Overall width 38.3 in (97 cm)
Wheelbase 22 in (55.9 cm)
Track 34.5 in (87.6 cm)
Weight 231.5 lb (105 kg)
Longitudinal distance rear axle to center of gravity, 4.3 in (11 cm)
Longitudinal distance front axle to center of gravity, 17.7 in (45 cm)
GPS Antennae spacing center to center, 31.5 in (80 cm)
Power source (2) 12 V Diehard rechargeable automotive batteries
Drive motors (2) NPC Robotics NPC-T64 DC permanent magnet with 20:1 reduction
gearbox
Controller RoboteQ AX2850 dual-channel forward/reverse digital controller
Encoders (2) Encoder Products #15T
GPS Crescent Vector OEM Board
Antenna (2) CSI Wireless CDA-3RTK

highly-dynamic situations (Boyden and Velinsky, 1994). A purely kinematic tracking

approach is susceptible to several error sources, resulting in accumulating drift.
Here, shaft-mounted encoders measure small variations in wheel rotation to
determine traveled wheel distance based on effective wheel radii, rRe and rLe . The
experimentally calibrated effective wheel radii are strongly surface-dependent—
sensor fusion, as described later, compensates for this. Odometry is particularly
vulnerable to uncertainties due to wheel slippage, predominantly on orientation,
which is used to derive position.
Tracking control for a differentially steered vehicle is challenging due to
natural nonholonomic constraints. The configuration on a 2D planar surface is
uniquely described by three generalized coordinates x =
x y T , but the vehicle
has only two independent control inputs: the right and left wheel actuation.
Consequently, position and orientation are inherently coupled in the path tracking
process and independent control in all three generalized coordinates x y  is not
possible. The approach and notation of Zhang et al. (1997) is used herein for
 OBSTACLE AVOIDANCE FOR AUTONOMOUS MOWING 337

Figure 1 The autonomous mower testbed.

kinematic tracking control. The performance of the controller has been proven in
Zhang et al. (1997).
Tracking point velocity and platform angular velocity are

1 b 1
u=  r + L rLe   v = R rRe − L rLe    = R rRe − L rLe  (1)
2 R Re d d
Here, R and L are the right and left wheel speeds, d is track width, and b is
distance from baseline to body frame.
The control law is defined as

 


R t −1
ẋB t xB t − xA t

d = = Gp  + Kp  (2)
L t ẏB t yB t − yA t
d

where Kp = diagkx  ky , kx > 0 ky > 0 is the gain matrix which defines the
exponential rate of convergence, and
 
− d 2br
sin t
+ cosr t d cos t
+ sinr t
Gp−1  =  d sin t 
Re Re 2br Re Re
(3)
2br
+ cos t
r
− d cos t
2br
+ sin t
r
Le Le Le Le

From (2), control input
d =
R t L tTd is determined (at 5 Hz) from the
reference path velocity in the X Y global coordinate frame and the x y error
offset of the testbed tracking point from the desired reference point, with desired
speed
d then input to the motor control. This control law yields exponentially
convergent tracking with rate minkx  ky , neglecting the motor control dynamics.
338 CREED ET AL.

As such, a vehicle can exactly track the x y position of a predefined reference path.
This solution does not yield independent control of the heading angle, and exact
orientation tracking occurs only when the reference path is a straight line (Zhang
et al., 1997). For automated mowing with long sections of straight cutting swaths,
independent orientation tracking is not necessary.

THE LOW-COST SENSOR ARRAY

Performance Measures
A robot that can mow the roadside quickly, efficiently, and effectively is
desired. The concepts summarized here are discussed in detail in Arsenault et al.
(2010). These metrics capture the ultimate performance, and guide the experimental
evaluation below.
A quick mowing rate is essential. The overall mowing rate should be
comparable to existing practices. A single automated unit with a baseline of only
60–101 cm (24–40 in) cannot achieve the rate of existing high-capacity mowers.
However, automated technology can work continuously (even overnight) with little
human intervention, and multiple units can be deployed, yielding comparable
overall rate.
Efficiency is also important—ideally the mower should never overlap
previously covered paths and only target uncut vegetation, reducing power
consumption, and supporting quicker mowing rates. Effectiveness minimizes the
amount of vegetation left behind. Missed patches of vegetation are costly to target
by hand. Effective mowing is achieved by partially overlapping swaths, reducing
efficiency; since efficiency and effectiveness act against one another, a suitable
compromise is necessary to maximize overall performance.

Sensor Selection
For an automated roadside mowing unit to supplement existing practices,
a cost-effective solution will yield a significant reduction in labor, material, and
equipment cost in comparison to the total investment, without accounting for the
gains from improved safety. Hardware and sensor technology usually dominate
overall system cost; thus, this paper’s goal is an effective but low-cost sensor and
hardware array.
An approach similar to that taken in Aono et al. (1998) was adopted.
Odometry from relative sensors (wheel shaft encoders) is fused with absolute GPS
(Global Positioning System) information to produce an accurate localization system.
Aono et al. (1998) found that dead reckoning and GPS measurements could be
fused to result in accuracy of about 0.2 m with a GPS accuracy of only 1.0 m. For
efficient mowing, final system accuracy of about 1 ft (0.3 m) or better is acceptable,
suggesting, based on Aono’s study (Aono et al., 1998), a GPS unit with accuracy
of roughly 1.5 m (60 in) or better. A low-cost (∼$1000) Crescent Vector OEM
board from Hemisphere GPS was selected. This unit provides reasonable accurate
global positioning data and excellent absolute heading information, using two
antennas separated with a known baseline, and uses WAAS corrections to yield
horizontal accuracy of less than 0.6 m (24 in), 95% of the time. This configuration
 OBSTACLE AVOIDANCE FOR AUTONOMOUS MOWING 339

was convenient, as freely available WAAS correction provides practical code-based

solution accuracy without a differential station. The Crescent board is also aided
by an on-board single-axis gyro that provides heading accuracy better than 1 deg
(0.02 rad) for upto 3 min during GPS signal loss.
For accurate heading information, the Crescent Vector OEM board uses two
antennas, separated by a known baseline. Two CSI Wireless CDA-3RTK antennas
were mounted at a separation of 0.8 m (31 in). The performance of the absolute
measured heading is directly correlated to the antenna separation. Experiments
showed that a separation of 0.8 m (31 in) produced accurate heading measurements
at an acceptable acquisition rate. The further apart the antennas, the better the
heading precision, but the slower the acquisition.
Tracking based only on a 2D model neglects variations in the terrain,
e.g., the bumps, small hills, and inclines common throughout roadside medians
and shoulders. Kiriy (2002) suggested that the errors that arise from a 2D
planar approximation can be compensated for by incorporating absolute sensor
information. In addition, Fuke and Krotkov (1996) determined that for most
low-dynamic situations a 3D dead reckoning model (using accelerometers and
gyroscopes) provided insignificant improvement over a 2D-only model. Here, a 2D
planar approximation that uses absolute GPS positioning information was selected,
i.e., absolute GPS information is used to continually correct for the inaccuracies that
result from the 2D tracking model.

LOCALIZATION AND PATH PLANNING

For localization, an Extended Kalman Filter (EKF) (at 5 Hz update) fuses
stochastic information from the encoders and differential GPS (DGPS) (Aono et al.,
1998; Rezaei and Sengupta, 2007). For the 2D planar coordinate system selected
for localization, three state variables are important: the absolute x y positioning
and the heading orientation . As such, a standard 3-state EKF for stochastic
state estimation is adapted from Kiriy (2002). Since odometry is a poor orientation
source but a good displacement measure, the fusion process relies heavily on the
GPS heading measurements, but not so heavily on the position readings.
The ideal choice for GPS-aided mowing is parallel consecutive mowing swaths
along the longer axis, such as the zigzag path configuration displayed in Fig. 2, also
known as a boustrophedral or raster pattern, realistic since the differential drive
platform’s turn radius is zero. Here, simple rectangular sections of vegetation are
studied; extension to irregular sections is achievable, including sections with random
obstacles, by using simple cell decomposition techniques (Schworer, 2005).
A simplified online path-planning algorithm builds the reference path, which
is stored as a sequence of line and arc segments (Fig. 2). Each segment is correlated
to a parameter array: start and end position, time, speed, and heading. These
parameters define the reference for real-time control.
The current path-planning algorithm can use a predefined rectangular mowing
section (length, width, and direction) to derive the mowing path, or absolute
coordinates (waypoints). The path planning is performed only once during startup.
Waypoints can be defined using aerial images (from Google Earth, or higher-
accuracy and higher-resolution imagery), as discussed in Shair et al. (2008); higher
accuracy will be required for the current application, but the approach applies. In
340 CREED ET AL.

Figure 2 Planned reference path (color figure available online).

more complex environments, such as highway interchanges, optimal aerial imagery-

based path-planning approaches may be appropriate, for example, Jan et al. (2008);
these techniques must be carefully applied to assure effective coverage. Simultaneous
Location And Mapping (SLAM) and obstacle detection and avoidance will be
important in a dynamic roadside environment; effective approaches include Kalman
filtering to fuse inertial, GPS, laser scanning, and odometry data for SLAM (Asadi
and Bozorg, 2009).

OBSTACLE DETECTION AND AVOIDANCE

The task of efficiently processing sensor data is dependent on the particular
obstacle avoidance algorithm employed. Performing obstacle avoidance in real time
can be a large challenge and the time between receiving sensor data and acting
on it is critical. As a result, there is a tradeoff between completeness of detail and
response time (Payton, 1986). This dilemma leads to a class of algorithms commonly
referred to as “reflexive” or “reactive.” This type of algorithm provides a timely
response without reliance on detailed world models, and generates actions triggered
by current perceptions (Ravela, 1992). The reactive nature makes it possible to
efficiently avoid both static and moving obstacles. For example, in Martinez et al.
(1998), a simple algorithm is introduced that can follow walls, follow moving
targets, and avoid obstacles. Reactive navigation methods only need to extract the
essential information for a particular path from sensor data. An algorithm proposed
in Martinez et al. (1998) makes use of a single data point corresponding to the
shortest detected range from a full LIDAR (Light Detection and Ranging) scan.
A qualitative evaluation of several obstacle avoidance navigation techniques in the
context of autonomous roadside mowing was conducted in Chavez (2007).
The testbed control architecture employed is shown in Fig. 3, wherein the
main focus of the current paper is contained in the dotted line, i.e., the reactive
 OBSTACLE AVOIDANCE FOR AUTONOMOUS MOWING 341

Figure 3 Control architecture.

obstacle avoidance driven by the LIDAR sensor and feeding the pure-pursuit
tracker. Another key component is the EKF, which takes input from the robot’s
wheel sensors as well as a dual-antenna GPS and a gyro to generate the current
state estimate, as discussed in detail in Arsenault (2007) and Arsenault et al. (2011).
To provide obstacle sensing on the testbed, a SICK LMS291 2D laser scanner
has been implemented. The LMS291 is a Class 1 (eye-safe) 2D laser scanner with a
maximum scanning swath of 180? ( rad). The data gathered by the laser contains a
distance reading to the nearest object for each angular increment as small as (1/4)
(0.004 rad). Under ideal conditions, it has a range of approximately 80 m (262 ft).

Path Tracker and Obstacle Avoidance Algorithms

The combination of the path tracking evaluation results in Creed (2011) and
obstacle avoidance algorithm requirements motivated the move to a geometric path-
tracking algorithm. Pure-pursuit is one of the simplest geometric path-tracking
algorithms, and is widely used (Howard et al., 2006). The algorithm works by
repeatedly fitting circular arcs to goal points on the path as the vehicle moves
forward. Unfortunately, the traditional forms of pure-pursuit found in most
literature can only track a point on the center of the robot’s baseline. Additionally,
it does not attempt to minimize path tracking orientation error.
There exist many variations and improved versions of the traditional pure-
pursuit algorithm. One such variation (Kuwata et al., 2009) tracks a point off
the baseline of a bicycle, tricycle, or Ackerman Steered vehicle, and is modified
for a differentially steered vehicle herein. Also, in Stentz et al. (2002), a simple
modification to the basic pure-pursuit algorithm is presented to add orientation
correction to an autonomous tractor that is designed to traverse similar paths as
our mower.
Accordingly, the pure-pursuit tracker is modified to track a point anywhere
on the centerline of the differentially-steered vehicle and to correct orientation error
with the use of a heading error term multiplied by an experimentally determined
gain Kh . The result is expressed as

L2A + 2yA b
RB = (4)
2xA + Kh p − 
342 CREED ET AL.

where RB represents the radius of the circle traversed by baseline center point B,
LA is the look ahead distance to the goal point from the tracking point, xA and yA
represent the distance to the goal point in a coordinate frame fixed at tracking point
A with the y-axis aligned with the vehicle centerline, b is the distance between points
A and B, and p −  denotes the heading error.
A large variety of collision avoidance algorithms exist, varying in method,
computational intensity, and sensor and vehicle requirements. The emphasis of
this work is on the application, and thus two existing reactive algorithms were
investigated and evaluated. These are summarized below for completeness.
Purely Reactive Collision Avoidance I: In the purely reactive collision avoidance
algorithm (Martinez et al., 1998), the equations that govern the behavior of the
vehicle are quite simple. Both the look-ahead distance and lateral error must be
known to utilize the pure-pursuit tracker. With reactive obstacle avoidance, the
look-ahead distance, L, is a dynamic value determined by the distance to the nearest
obstacle. The algorithm is composed of two layers. The first layer is responsible for
the basic pure-pursuit path tracking. A path is provided to the vehicle, and pure-
pursuit is repeatedly used to determine and track goal points that are L ahead of
the vehicle’s current position. Simultaneously, the laser rangefinder (LRF) performs
180 ( rad) sweeps in front of the vehicle. The shortest range and corresponding
orientation is returned. If the minimum distance is outside of a specified threshold,
no action is taken. If inside, the algorithm selects a new goal point by rotating the
detected obstacle point’s coordinates by 90 ( /2 rad) away from the obstacle.
Singapore Harbor Reactive Collision Avoidance: The second collision avoidance
algorithm is based on a reactive algorithm presented in Bandyophadyay et al.
(2010). This algorithm utilizes the entire data set returned by the LRF to
intelligently plan a path.
When a threatening object is detected by the LRF, a vehicle-bounding radius
is drawn around the points that define the left and right boundaries of the detected
obstacle. Next, tangent points are calculated that connect a line from the vehicle’s
current position to the endpoint circles. The distance from each tangent point to the
specified “goal point” is calculated, and the point with the shortest distance becomes
a sub-goal, and is used to determine the new heading of the vehicle. Once the vehicle
arrives at the sub-goal, it reorients back towards the original goal.

Evaluation of Collision Avoidance Algorithms

Though the primary objective of a collision avoidance algorithm is to
successfully guide a robot around an obstacle, there are a variety of other
important considerations. For mowing, an optimized collision avoidance algorithm
encompasses the ability to continually guide the vehicle near the obstacle without
collision, and return it to the reference trajectory as soon as possible. Other
important factors such as path efficiency, smoothness, and risks taken by the vehicle,
make this a multiple objective optimization problem.
To achieve such an optimization or evaluation, a variety of measures may be
needed. One class of measures presented in Lampe and Chatila (2006) are referred
to as performance measures. Among these measures is total traveled distance. In
the context of autonomous roadside mowing, the optimal path traces the perimeter
of the obstacle, and may not always be the shortest possible path. An additional
 OBSTACLE AVOIDANCE FOR AUTONOMOUS MOWING 343

measure is called the “Mean Distance to Goal”, or “MGD” (Munoz et al., 2007).
This metric is useful for robots that follow a reference trajectory, and quantifies the
average error between the traversed path and the reference path. Small MGD values
correspond to a traversed path that is close to the desired trajectory.
“Security Metrics” address the vehicle’s distance to potential obstacles (Munoz
et al., 2007). The first security metric (SM1) presented in Munoz et al. (2007)
calculates the average of the distances between the vehicle and obstacles throughout
the entire mission. With autonomous mowing, minimizing this value helps ensure
that the vehicle remains near the obstacle. The minimum distance that the vehicle
maintains from the obstacle is also important and it can be extracted when
calculating SM1.
In addition to performance and security measures, smoothness measures
should be examined to help minimize energy and time requirements. The “Control
Effort” (Martinez et al., 1998) is defined as the average absolute value of the
curvature increments for all the control intervals. The control effort can be used as
a relative measurement between algorithms to help determine which path required
more energy for the vehicle to traverse.
The above measures, combined as a normalized weighted sum, provide a good
measure of the quality of the whole system. The weights were determined as listed
in Table 1, based on the importance of each measure for the specific mowing
application. Each algorithm was evaluated on the robot as it traversed around
obstacles in a test course. The score was then used to make the final selection for
the most effective algorithm.
The performance of each algorithm was evaluated based on two simple test
courses. The first course consisted of a single obstacle placed in the center of the
desired trajectory. This simple course was used to identify the basic behavior of
each algorithm, for clear comparison. The second course consisted of two obstacles
separated by a distance d. This course was utilized for a set of tests, each with a
different initial posture and various obstacle spacing distances. It was designed to
determine how the robot handled the presence of multiple obstacles.
For brevity, the quantitative analysis of only the single obstacle tests are
included herein; results for the multiple obstacle tests are in Creed (2011). The data
presented are based on an average of three trial runs.
For each algorithm, the relative quantitative measures are tabulated in
Table 2. Two figures are provided for the algorithms. The first illustrates both
desired and actual trajectory traversed as well as relevant LRF data points (Fig. 4).

Table 2 Algorithm score calculation

Reactive Harbor

Measure Weight Value Normalized Score Value Normalized Score

MGD 35% 0.016 1 0.35 0.018 0.89 0.312

SM1 35% 1.641 0.874 0.306 1.435 1 0.35
Actual length 15% 8.800 1 0.15 9.428 0.93 0.140
Control effort 15% 0.285 1 0.15 0.356 0.8 0.12
Total score 0.956 0.921
344 CREED ET AL.

Figure 4 Reactive path and trajectory history (color figure available online).

The second plot illustrates path error as a function of time (Fig. 5), where the error
is quantified as the distance to the closest path point.
Both obstacle avoidance algorithms successfully guided the vehicle around the
circular obstacle. Each of them worked differently however, and the paths taken

Figure 5 Reactive algorithm error vs. time (color figure available online).
 OBSTACLE AVOIDANCE FOR AUTONOMOUS MOWING 345

Figure 6 Test in field (color figure available online).

differ significantly. From Fig. 4, the Singapore Harbor algorithm guided the vehicle
to the left of the obstacle, and the reactive algorithm resulted in a much more
circular route. Traversal of a noncircular path would result in a larger control effort
(due to curvature variations) than the circular case. Neither algorithm resulted in
a collision with the obstacle, but both algorithms left and returned to the path
at different points. When returning to the path, the Singapore Harbor algorithm
produced considerably higher overshoot. Since both algorithms utilized the pure-
pursuit path-tracking algorithm with the same look-ahead distance, the overshoot
is likely an artifact of computational delays.
The error plots presented in Fig. 5 reveal that the harbor algorithm
consistently guided the vehicle further from the obstacle, even though the reaction
threshold distance was the same for both algorithms. The overshoot discussed from
Fig. 4 is also shown in the Fig. 5, peaking at approximately t = 45 sec. Based on
these results, the reactive algorithm was selected for final implementation.

PERFORMANCE EVALUATION IN A REAL WORLD ENVIRONMENT

As in Arsenault et al. (2010, 2011), in addition to the onboard Crescent
Vector GPS unit, a relatively expensive onboard NovAtel GPS board was used
to provide ground truth (GT) information to assess the mowing performance. To
improve the performance of the onboard NovAtel GT positioning information,
a second stationary NovAtel GPS differential station was set up to collect GPS
information. After testing was completed, information from the onboard NovAtel
and stationary NovAtel unit were postprocessed using daily web-based corrections
to provide absolute positioning accuracies (1-sigma) on the order of 2 cm (0.8 in) or
less. This provided GT in the final test cases.
Final system validation was performed in a rough grassy field representative
of the anticipated mowing environment. The results of this test are shown in Fig. 6.
Two obstacles were placed in the direct line of anticipated robot travel. Obstacle 1
was located midway on swath 1, with obstacles 2 located between swaths 4 and 5.
346 CREED ET AL.

The results showed that the robot successfully navigated autonomously around each
obstacle. In addition, the robot maintained path-tracking consistency similar to
results in Arsenault et al. (2011), which is essential for efficient and effective mowing.
While the state estimation and GT do show some variance, the essential point is
that consistent mowing width is maintained. Thus, the robot maintains its ability to
perform its primary task but has the added capability of obstacle avoidance.

CONCLUSION
This paper presented obstacle avoidance for an autonomous mower for
highway vegetation control. Reactive algorithms were evaluated based upon
quantitative performance measures. The selected algorithm was combined with
the pure-pursuit-tracking controller, which was successfully validated in a harsh
outdoor environment including obstacles. The performance measures, the control
architecture, and implementation approach were all tailored specifically for this
robotic application.
While highly promising, additional research is necessary before such a system
could be fully deployed into highway operations. First, due to the close proximity
of traffic, it is essential that a detailed failure mode analysis be performed. Second,
it is well known that highway maintenance and construction work is a distraction to
passing motorists. While the autonomous mowers will be moving at a relatively slow
speed, it is still necessary to assess and mitigate the impacts on driver distraction.

ACKNOWLEDGMENTS
The authors gratefully acknowledge the Division of Research and Innovation
of the California Department of Transportation for the support of this work
through the Advanced Highway Maintenance and Construction Technology
(AHMCT) Research Center at the University of California, Davis.

REFERENCES
Aono, T., Fujii, K., Hatsumoto, S., Kamiya, T. (1998). Positioning of Vehicle
on Undulating Ground Using GPS and Dead Reckoning. Presented at the IEEE
International Conference on Robotics and Automation, Leuven, Belgium, May
16–20.
Arsenault, A. (2007). Implementation and Validation of a Low Cost Sensor Array for
Autonomous Roadside Mowing. MS thesis, University of California, Davis.
Arsenault, A., Velinsky, S. A., Lasky, T. A. (2010). Autonomous mowing –
Improving efficiency and safety of roadside vegetation control. ASCE Journal of
Infrastructure Systems 16(3):206–215.
Arsenault, A., Velinsky, S. A., Lasky, T. A. (2011). A low-cost sensor array
and test platform for autonomous roadside mowing. IEEE/ASME Transactions
Mechatronics 16(3):592–597.
Asadi, E., Bozorg, M. (2009). A decentralized architecture for simultaneous
localization and mapping. IEEE/ASME Transactions Mechatronics 14(1):64–71.
 OBSTACLE AVOIDANCE FOR AUTONOMOUS MOWING 347

Bandyophadyay, T., Sarcione, L., Hover, F. (2010). A simple reactive obstacle

avoidance algorithm and its application in Singapore Harbor. Field and Service
Robotics 62:455–465.
Boyden, F. D., Velinsky, S. A. (1994). Dynamic Modeling of Wheeled Mobile Robots
for High Load Applications. Presented at the IEEE International Conference on
Robotics and Automation, San Diego, CA, May 8–13.
Caltrans (1997). California Roadsides: A New Perspective. Sacramento, CA: Jones &
Stokes Associates, Inc.
Chandler, R. C., K. A. Meiszer, and A. A. Arroyo (1999). LawnShark: A New
Platform for Autonomous Mowing and Navigation. Presented at the Florida
Conference on Recent Advances in Robotics, Gainesville, Florida.
Chavez, R. C. (2007). A Study of Object Detection and Avoidance for an Autonomous
Roadside Mowing Application. MS thesis, University of California, Davis.
Creed, B. (2011). Implementation and Validation of Obstacle Avoidance for Autonomous
Roadside Mowing. MS thesis, University of California, Davis.
Dudek, G. (2000). Computational Principles of Mobile Robotics. Cambridge:
Cambridge University Press.
Fuke, Y., Krotkov, E. (1996). Dead Reckoning for a Lunar Rover on Uneven Terrain.
Presented at the IEEE International Conference on Robotics & Automation,
Minneapolis, Minnesota, April 22–28.
Hong, D., Velinsky, S. A., Feng, X. (1999). Verification of a wheeled mobile robot
dynamic model and control ramifications. Transactions ASME, Journal of Dynamic
Systems, Measurement, and Control 121(1):58–63.
Howard, T., Knepper, R. A., Kelly, A. (2006). Constrained Optimization Path
Following of Wheeled Robots in Natural Terrain. Presented at the 10th International
Symposium on Experimental Robotics, Rio de Janeiro, Brazil, July 6–10.
Jan, G. E., Chang, K. Y., Parberry, I. (2008). Optimal path planning for mobile
robot navigation. IEEE/ASME Transactions Mechatronics 13(4):451–460.
Kiriy, E. (2002). A Localization System for Autonomous Golf Course Mowers. MS
thesis, McGill University, Montreal, Quebec, Canada.
Kuwata, Y., Karaman, S., Teo, J., Frazzoli, E., How, J. P., Fiore, G. (2009).
Real-time motion planning with applications to autonomous urban driving. IEEE
Transactions on Control Systems Technology 17(5):1105–1118.
Lampe, A., Chatila, R. (2006). Performance Measure for the Evaluation of Mobile
Robot Autonomy. Presented at the IEEE International Conference on Robotics and
Automation, Orlando, Florida, May 15–19.
Martinez, J. L., Pozo-Ruz, A., Pedraza, S., Fernandez, R. (1998). Object Following
and Obstacle Avoidance Using a Laser Scanner in the Outdoor Mobile Robot Auriga-
. Presented at the IEEE International Converence on Intelligent Robots and
Systems, Victoria, B.C., Canada, October 13–17.
Munoz, N. D., Valencia, J. A., Londono, N. (2007). Evaluation of navigation of
an autonomous mobile robot. Workshop on Performance Metrics for Intelligent
Systems, Washington, D.C.
Payton, D. (1986). An architecture for reflexive autonomous vehicle control. IEEE
International Conference on Robotics and Automation.
Ravela, S. S. (1992). A Survey of Reactivity. Amherst, MA: University of
Massachusetts.
348 CREED ET AL.

Rezaei, S., Sengupta, R. (2007). Kalman filter based integration of DGPS and
vehicle sensors for localization. IEEE Transactions on Control System Technology
15(6):1080–1088.
Rizzoni, G. (2005). Principles and Applications of Electrical Engineering. New York:
McGraw-Hill.
Schworer, I. J. (2005). Navigation and control of an autonomous vehicle. MS thesis,
Virginia Polytechnic Institute and State University, Virginia.
Shair, S., Chandler, J. H., Gonzalez-Villela, V. J., Parkin, R. M., Jackson, M. R.
(2008). The use of aerial images and GPS for mobile robot waypoint navigation.
IEEE/ASME Transactions Mechatronics 13(6):692–699.
Stentz, A. T., Dima, C., Wellington, C., Herman, H., Stager, D. (2002). A system
for semi-autonomous tractor operations. Autonomous Robots 13(1):87–103.
Teeter-Balin, J. (2004). Feasibility of Alternative Vegetation Maintenance Control
Technologies. MS thesis, University of California, Davis.
Zhang, Y., Velinsky, S. A., Feng, X. (1997). On the tracking control of differentially
steered wheeled mobile robots. Transactions ASME, Journal of Dynamic Systems,
Measurements, and Control 119(3):455–461.