Helvis - a Small-scale Agricultural Mobile Robot Prototype for Precision Agriculture

Abstract

The use of agricultural robots is emerging in a complex scenario where it is necessary to produce more food to feed a crescent population, decrease production costs, fight plagues and diseases, and preserve nature. Around the world, there are many research institutes and companies trying to apply mobile robotics techniques in agricultural fields. Mostly, large prototypes are being used and their shapes and dimensions are very similar to tractors and trucks. In the present study, a small-scale prototype was designed, aiming to facilitate the controller development phase and the execution of experiments in the university using a farm-like scenario (before validating the controllers in a real scenario). It is important to highlight that all control parameters were parameterized to allow the control portability to other prototypes. Helvis is an electric small-scale car-like platform whose traction and steering systems are powered by Maxon motors and driven by EPOS2 boards. Its navigation system uses 2 LiDAR sensors (UTM-30LX) to scan the environment (one in the front and the other in the back) and an Inertial Navigation System (IG500N) to estimate its orientation. In this paper we present experiments carried out in rows of a corn crop field. As previously mentioned, before making the experiments in a real farm, a farm-like scenario was constructed in the lab to calibrate the controller parameters. Since each cornrow constitutes itself a discontinuous wall, a filter based on the LiDAR data was developed in order to create virtual continuous walls. So, these virtual walls were used as references for a wall-follower control system. When it is possible to create walls in both sides of the robot, the navigation problem can be simplified to moving the robot in a virtual aisle. The filter calculates the distances between robot and virtual walls, which are used as input data for the fuzzy controller responsible to keep the robot in the path between corn rows. Its output signal acts in Helvis’ steering system. Real environment experiments allowed adjusting fuzzy rule set and improving robot performance.

Full text

HELVIS - A SMALL-SCALE AGRICULTURAL MOBILE
ROBOT PROTOTYPE FOR PRECISION AGRICULTURE
Velasquez A. E. B+, Higuti V. A. H+, Guerrero H. B+, Milori D. M*, Magalhães D. V+,
Becker M+.
+ São Carlos Engineering School, São Paulo University, São Paulo, Brazil.
*Embrapa Instrumentation, São Carlos, São Paulo, Brazil
A paper from the Proceedings of the
13th International Conference on Precision Agriculture
July 31 – August 4, 2016
St. Louis, Missouri, USA
Abstract. The use of agricultural robots is emerging in a complex scenario where it is necessary to
produce more food to feed a crescent population, decrease production costs, fight plagues and
diseases, and preserve nature. Around the world, there are many research institutes and companies
trying to apply mobile robotics techniques in agricultural fields. Mostly, large prototypes are being
used and their shapes and dimensions are very similar to tractors and trucks. In the present study, a
small-scale prototype was designed, aiming to facilitate the controller development phase and the
execution of experiments in the university using a farm-like scenario (before validating the controllers
in a real scenario). It is important to highlight that all control parameters were parameterized to allow
the control portability to other prototypes. Helvis is an electric small-scale car-like platform whose
traction and steering systems are powered by Maxon motors and driven by EPOS2 boards. Its
navigation system uses 2 LiDAR sensors (UTM-30LX) to scan the environment (one in the front and
the other in the back) and an Inertial Navigation System (IG500N) to estimate its orientation. In this
paper we present experiments carried out in rows of a corn crop field. As previously mentioned,
before making the experiments in a real farm, a farm-like scenario was constructed in the lab to
calibrate the controller parameters. Since each cornrow constitutes itself a discontinuous wall, a filter
based on the LiDAR data was developed in order to create virtual continuous walls. So, these virtual
walls were used as references for a wall-follower control system. When it is possible to create walls
in both sides of the robot, the navigation problem can be simplified to moving the robot in a virtual
aisle. The filter calculates the distances between robot and virtual walls, which are used as input data
for the fuzzy controller responsible to keep the robot in the path between corn rows. Its output signal
acts in Helvis’ steering system. Real environment experiments allowed adjusting fuzzy rule set and
improving robot performance.
Keywords. Mobile Robot, Precision Agriculture, Perception, Navigation.

Proceedings of the 13th International Conference on Precision Agriculture
July 31 – August 3, 2016, St. Louis, Missouri, USA Page 2
The authors are solely responsible for the content of this
paper, which is not a refereed publication.. Citation of this wor
k should state that it
is from the Proceedings of the 13
th International Conference on Precision Agriculture. EXAMPLE:
Lastname, A. B. & Coauthor, C. D.
(2016). Title of paper. In Proceedings of the 13
th
International Conference on Precision Agriculture (unpaginated, online). Monticello, IL:
International Society of Precision Agriculture.

Proceedings of the 13th International Conference on Precision Agriculture
July 31 – August 3, 2016, St. Louis, Missouri, USA Page 3
INTRODUCTION
Nowadays agriculture is facing unprecedented demands. United Nations estimate that the world
population will achieve 10 billion people by 2050 (UNFPA, 2015) and that there will be globally about
795 million people undernourished in 2014-16 (FAO, 2015). In this scenario, agricultural production
must grow substantially while confronting the need of reducing its environmental footprint (Foley et
al., 2011). In this context, it was possible to observe in the last few years that robotics technology is
becoming more present in many agriculture areas (Reina et al., 2015).
Cheein and Carelli (2013) summarize the most important abilities of automatic agricultural vehicles
into four categories: guidance, detection, action, and mapping. For these activities, devices like
camera modules, Real Time Kinetics Global Positioning System (RTK-GPS) and Light Detection and
Ranging (LiDAR) are needed to retrieve information from environment. An example of vision based
system can be found in (English et al., 2014) which describes a crop rows tracking method using
vision based texture model. An automatic and accurate guidance of a four-wheel-steering mobile
robot using RTK-GPS is described in (Cariou et al., 2009). Weiss and Biber (2011) used a 3D LiDAR
sensor to detect single plants in crop rows in real time.
Auto-guided agricultural vehicles with GPS-based navigation systems started appearing around
1997. They were able to follow a predefined map, bringing benefits like accuracy increase and
driver’s fatigue reduction (Heraud and Lange, 2009). But when dealing with unmanned vehicles,
which do not have an operator to manually interfere in the presence of environmental alterations, a
GPS-based system is not enough to guarantee that the vehicle will not ride over the crop and
unmapped elements (like other vehicles, animals, humans or recent changes) thus raising a major
safety issue (Reina et al., 2015). Besides that, high-accuracy GPS systems, such as differential GPS
and RTK-GPS, are expensive but even so, subject to interruptions or poor availability of the signal
which leads to great position errors and possibly, navigation failure (Hiremath et al., 2009). Although
great improvement can be seen in (English et al., 2014), vision based methods are sensitive to
variations in environment light and atmospheric effects. Outdoor environments, such as the
agricultural ones, would require frequent calibration procedures (Hiremath et al., 2009).
Regarding LiDAR-based navigation systems, Hiremath et al. (2009) proposed and showed the
robustness of a LiDAR-based autonomous navigation based on a particle filter. Barawid et al. (2007)
highlighted two advantages of using such systems: as it takes into account real-time local
information, a pre-surveying task to generate a map can be omitted (thus saving time); and they can
be used when GPS becomes non-functional. In addition, Weiss and Biber (2011) state that in
contrast to most stereo vision cameras, the ranging information is calculated on the LiDAR sensor
itself, thus no further time-consuming calculation has to be carried out on the computer, and the
distance values’ precision is also independent from the measured distances.
For the aforementioned reasons, this study is focused on the development of a LiDAR-based
autonomous navigation system of a car-like mobile robot through cornfield without relying on a
planned path. For this purpose, a fuzzy controller is proposed and experimentally tuned for the crop
row following. Appropriate row exit and entrance maneuvering routines based on LiDAR and robot’s
heading readings were also implemented. Although high-level decision making would require
additional resources, such as landmarks, Radio-Frequency IDentification (RFID) tags or even a GPS-
based map, the developed navigation system is able to work in standalone mode ensuring crop’s
integrity.
HELVIS PROTOTYPE
Helvis (Fig 1) is an electric small-scale car-like platform (Length | Width | Height: 0.65m x 0.23m
0.45m) whose traction set consists in an EPOS2 50/5 board, an automotive-like differential
mechanism and an EC-4Pole 30 Maxon Motor (reference 309756) which has two shafts (rear and
front). The rear shaft is coupled to an incremental encoder (reference 255778) and the front shaft is

Proceedings of the 13th International Conference on Precision Agriculture
July 31 – August 3, 2016, St. Louis, Missouri, USA Page 4
coupled to a planetary gearhead GP 32 (reference 326661) for a 23:1 reduction. Its steering set has
an EPOS2 24/5 board, an Ackerman mechanism, a power screw and a DC Maxon motor REmax29
(reference 226802). The rear shaft of DC Maxon Motor is coupled to an incremental encoder
(reference 225805) and its front shaft is coupled to a planetary gearhead GP 32 (reference 166938)
for a 33:1 reduction. In order to scan the environment, it has two embedded LiDAR sensors (UTM-
30LX) manufactured by Hokuyo (Hokuyo Automatic, 2012): the first one is located in the front part of
the robot and the other one is in the rear part. Also, the robot has an Inertial Navigation System
(IG500N) manufactured by SBG Systems (SBG Systems, 2013) in order to estimate its orientation
and position (it will be used in future works). In this work, orientation refers to yaw orientation angle.
Its main control unit is a Raspberry Pi 2 running Raspbian operation system. For monitoring some
variables during tests, Helvis has a router to allow remote communication between Raspberry Pi 2
and an external computer. The interaction between its embedded devices is presented in Fig 2.
Finally, characterization of sensors and description of each part of Helvis structure are more detailed
in (Velasquez, 2015).
Fig 1. Helvis in middle of cornrow crop.
The steering system described in (Velasquez, 2015) was changed in order to increase its robustness
and steering angle ranges. Its new steering and traction systems are showed in Fig 2.
Fig 2. Steering and traction sets of Helvis.
CONTROL SYSTEM OF HELVIS
Helvis’ control system is divided in two levels: High and Low control level systems (Fig 3). High-level
system (also called wall-follower control system) consists of a fuzzy controller that keeps the robot on
the path between the cornrows. Its input is the difference between distances from the robot to each

Proceedings of the 13th International Conference on Precision Agriculture
July 31 – August 3, 2016, St. Louis, Missouri, USA Page 5
wall of corn rows and its output is the desired angle of the robot’s front wheels. As previously
mentioned in abstract, each corn row constitutes itself a discontinuous wall, then a filter was
developed to generate two continuous virtual walls based on the sensor readings (vehicle’s
orientation measured by IG500N and distances measured by UTM-30LX). The controllers and filters
are programmed with C++ language.
Fig 3. Interaction of Helvis’ embedded devices.
The high-level control system output is front wheels’ desired steering. This output and vehicle’s
desired speed, which maximum value is fixed at 0.2 m/s, are sent to low-level control system through
serial communication. Low-level control system consists of two EPOS2 boards in charge of
controlling traction and steering motors’ speed and position. For this purpose, the board EPOS2 has
embedded Proportional and Integral (PI) controllers (their description is available in Maxon Motors,
2013). While EPOS2 24/5 gets desired steering angle and makes DC Maxon Motor’s position control,
EPOS2 50/5 gets desired robot speed and makes EC Maxon Motor’s velocity control.
Filter to create the virtual walls
As previously mentioned in abstract, all experiments presented in this paper were carried out in a
farm-like scenario representing rows of corn crop field. Each row has about 40 corn plants organized
in a straight line (Fig 4). Separation between each corn plant is 0.2m and distance between each row
crop is approximately 1m.
Fig 4. Farm-like scenario created in the lab.

Proceedings of the 13th International Conference on Precision Agriculture
July 31 – August 3, 2016, St. Louis, Missouri, USA Page 6
The robot uses a LiDAR sensor to scan the environment. LiDAR’s scan range is 270° and its angular
resolution is 0.25° (270°/1080 steps). Every 0.1s, the sensor updates a vector with 1080 positions
with distances (in mm) between the sensor and any detected object for each step of its scan range.
When it comes to the filter, three situations are defined: the first situation is when robot is moving
without orientation error (eψ) (Fig 5a); the second one is when it is moving with a positive orientation
error (Fig 5b); and the last one is when it is moving with a negative orientation error (Fig 5c).
Orientation error (eψ) is the angular difference between path orientation (red arrow) and actual robot
orientation (blue arrow). Path orientation is the obtained orientation when the robot enters the crop
row.
Fig 5. a- The robot is running without orientation error. b- The robot is running with positive orientation error. c- The robot is
running with negative orientation error.
The filter’s main goal is to generate virtual walls based on information of two interest regions. The
first interest region, further referred as right-side interest region, comprehends the obtained readings
of steps located between the two orange arrows showed in Fig 5a, 5b, and 5c. The second region,
further referred as left-side interest region, contains the obtained readings of steps located between
the two black arrows showed in Fig 5a, 5b, and 5c. Angular separation between black arrows and
between orange arrows are both 60° (240 steps). For all cases presented in Fig 5, definition of an
initial and a final step for each interest region is needed. For the case A (right-side interest region -
Fig 5a), initial step is 180 (0°) and final step is given by equation 1.
__ =__ +󰇡
.󰇢 (1)
For the case of left-side interest region (Fig 5a), initial and final steps are defined by equations 3 and
2, respectively.
__ =__ +󰇡
.󰇢 (2)
__ =__  󰇡 
.󰇢 (3)
For the cases B and C (Fig 5b and Fig 5c), initial and final steps of right-side interest region are
defined by equations 4 and 5, respectively, while initial and final steps of the left-side interest region
are defined by equations 6 and 7, respectively.

Proceedings of the 13th International Conference on Precision Agriculture
July 31 – August 3, 2016, St. Louis, Missouri, USA Page 7
__ =180 + 󰇡 
.󰇢 (4)
__ =__ +󰇡
.󰇢 (5)
__ =__ +󰇡
.󰇢 (6)
__ =__  󰇡 
.󰇢 (7)
Each region has readings of 240 steps. Then, two vectors (each one with 240 positions) were
created in order to save information about interest regions. The first vector is called
points_right_region and contains information about right-side interest region. The second one is
called points_left_region and contains information about left-side interest region.
Fig 6. Global and Local Coordinate frames.
The main issue with information saved in points_right_region and points_left_region vectors is getting
their x and y coordinates (in meters) relative to coordinate frame {2}, shown by X2 and Y2 axes in Fig
6. This coordinate frame is a translation of global coordinate frame {G} (XG,YG axes in Fig 6) to the
origin of robot’s local coordinate frame {R} (XR,YR in Fig 6). While {G}’s origin is fixed at one corner of
the corn field, {R} is attached to the robot, with its origin in the center of front LiDAR sensor and XR
axis coincident with robot’s longitudinal axis.
As it can be seen in Fig 6, orientation error (eψ) is the angle between X2 and XR. Thus in order to
obtain projections in frame {2} from points_right_region and points_left_region vectors, equations 8,
9, 10 and 11 uses eψ to compensate robot’s orientation deviations. Using these equations, four
vectors are generated, and they contain x and y coordinates of points_right_region and
points_left_region vectors. Their names are X_right_region, Y_right_region, X_left_region and
Y_left_region.
__()=__()sin( 0.) +/1000 (8)

Proceedings of the 13th International Conference on Precision Agriculture
July 31 – August 3, 2016, St. Louis, Missouri, USA Page 8
__()=__()cos( 0.25) +/1000 (9)
__()=__()sin( 0.) +/1000 (10)
__()=__()cos( 0.25) +/1000 (11)
where steprw is an incremental variable (with increments equal to 1) which starts at initial_stepright-
side_region and ends at final_stepright-side_region. The steplw is another incremental variable (also with
increments equal to 1) which starts at initial_stepleft-side_region and ends at final_stepleft-side_region. Finally,
posrw and poslw are described by equations 12 and 13, respectively.
 =  __ (12)
 = __ (13)
Those projection vectors are filtered to generate virtual walls. The filter for virtual left wall is described
by the routine presented in equation 14.
if ((Y_left_region (stepfilter) > -1.2) && ((Y_left_region (stepfilter) < -0.2)
{
Yvirtual_left_wall (stepfilter) = Y_left_region (stepfilter);
Xvirtual_left_wall (stepfilter) = X_left_region (stepfilter);
}
(14)
And the filter for right wall is presented in equation 15.
if ((Y_right_region (stepfilter) < 1.2) && ((Y_right_region (stepfilter) > 0.2)
{
Yvirtual_right_wall (stepfilter) = Y_right_region (stepfilter);
Xvirtual_right_wall (stepfilter) = X_right_region (stepfilter);
}
(15)
Where stepfilter is an incremental variable (with increments equal to 1) that starts at 0 and ends at
239. Filtered vectors have X and Y coordinates of virtual walls.
Fig 7. a- Map obtained without filter. b- Map with the virtual walls obtained with the filter.
In order to calculate distances between robot and virtual walls, mean values of Yvirtual_left_wall and
Yvirtual_right_wall are obtained. Yvirtual_left_wall‘s mean value is the distance of the robot relative

Proceedings of the 13th International Conference on Precision Agriculture
July 31 – August 3, 2016, St. Louis, Missouri, USA Page 9
to left virtual wall (dlw) while Yvirtual_right_wall’s mean value is the distance of the robot relative to
right wall (drw). Fig 7a shows the map obtained with LiDAR readings without the filter and Fig 7b
shows the map with virtual walls obtained with filter.
Wall-follower system control
As mentioned before, the wall-follower system control has a fuzzy controller whose main objective is
to keep the robot in the path between the cornrows. Block diagram of wall follower system control is
shown in Fig 8.
Fig 8. Block diagram of the wall-follower system control.
For this subsection, distance represents the extension of space between Helvis longitudinal axis and
middle between crop rows. Wall-follower control system’s input is desired distance (dd) and its output
is the actual distance (da). This latter value is constantly obtained by using LiDAR sensor and the
filter mentioned in the previous section, and it is negative feedback to the fuzzy controller through
variable ed (distance error) which is the difference between dd and da. Value of desired distance is
always zero and value of real distance is the sum of drw and dlw (described in the filter section).
Fuzzy controller’s output is steering angle (u) for Helvis. Implementation of fuzzy controller was
based on the fuzzy controllers developed in (Guerrero et al., 2014; Higuti et al., 2015). It has an input
fuzzy set with five triangular pertinence functions (Fig 9a) and an output fuzzy set with three
triangular pertinence functions (Fig 9b). The five pertinence functions of input fuzzy set are named:
VFLW (Very_Far_Left_Wall), FLW (Far_Left_Wall), MP (Middle Path), FRW (Far_Right_Wall) and
VFRW (Very_Far_Right_Wall). And the three pertinence functions of the output fuzzy set are named:
LS (Left_Steering), NS (No_Steering) and RS (Right_steering).
Fig 9. a- Input fuzzy sets. b- Output fuzzy sets.
A value of distance error (ed) generates two pertinence values (µ1 and µ2) in two different input fuzzy
sets at the same time. These pertinence values are used to cut the pertinence function of the output
fuzzy set according to the fuzzy rules described in equation 16.
If ed = VFLW or if ed = FLW, then µ = LS

Proceedings of the 13th International Conference on Precision Agriculture
July 31 – August 3, 2016, St. Louis, Missouri, USA Page 10
If ed = MP, then µ = NS (16)
If ed = FRW or if ed = VFRW, then µ = RS
When pertinence values cut output fuzzy set’s pertinence functions, some output fuzzy values (Z) are
obtained. These values are used to find the defuzzified output value (Z*) which is obtained with the
centroid method (Guerrero et al., 2014). The Z* is fuzzy controller’s output (u) and it is a steering
angle for the robot.
SWITCHING BETWEEN TWO CROP ROWS
This section describes the maneuver process used to change crop rows. There are two maneuvers.
The first maneuver is used when the robot changes to another row located on its right side (Fig 10).
And the second one is a mirrored version as it is used when the following row is located to the left
side. The maneuver process is divided into four steps (Fig 10).
1. End of first crop row.
2. Finding next crop row
3. Entrance to next crop row.
4. On second crop row.
First step comprehends: detection of row end, measurement of actual robot’s orientation (ψfirst_step),
calculation and subsequent movement to desired orientation for second step (equation 17).
_ =_  (17)
In order to find row end, information of Yvirtual_left_wall and Yvirtual_right_wall vectors is used.
When the sum of number of elements of those two vectors is less than 220 elements, it indicates row
end was found. At this point, a mean value of 100 consecutive readings of orientation angle is taken
and set as desired orientation for first step. Then, ψsecond_step is calculated by equation 17 and the
robot continues moving forward but with front wheels totally steered towards right side. About
orientation signal, when robot steers to the right, the measured orientation decreases. And for the
opposite movement, it increases. When actual robot’s orientation is near (±10°) ψsecond_step, the robot
motion stops and front wheels are steered back to 0° (end of first step).
Fig 10. Maneuver to change the crop row.
In the beginning of second step, desired orientation (ψsecond_step) is updated. Reverse motion, crop
rows’ detection filter and fuzzy controller for reverse motion are activated. The condition to finish
second step is the counting of two crop rows by the robot. When this condition is satisfied, robot
motion, fuzzy controller and filter are stopped.
Third step consists in the entrance to the next row, which involves a 90° right or left turn. Thus,
desired orientation for third step (ψthird_step) is calculated by equation 18.

Proceedings of the 13th International Conference on Precision Agriculture
July 31 – August 3, 2016, St. Louis, Missouri, USA Page 11
_ =_  (18)
Subsequently, robot motion is activated and its front wheels are steered to the maximum steering
value towards its right side. When the robot orientation is near (±10°) ψthird_step, robot motion is
stopped and front wheels return to 0° (end of third step). In the final step, ψfourth_step is given by
equation 19 as the best estimative for the new crop row is that its path orientation is parallel but in
the opposite direction of the previous row. Then, robot motion is activated. From this point until next
turning maneuver, ψfourth_step is used in the filter (used to create the virtual walls) as the path
orientation. Finally, the wall-follower control system is activated.
_ =_  (19)
The second maneuver is used when the next crop row is located in the left side. This maneuver is
divided in the same steps as first one. Orientation for second step (calculated in the first step) and
fourth step (calculated in the fourth step) are different and they are defined by equations 20 and 21,
respectively. In the first and third steps of this maneuver, the robot steers its front wheels to its left
side. Finally, the second and fourth steps are the same.
_ =_ + (20)
_ =_ + (21)
Fuzzy Controller for the Reverse Motion
In order to reduce deviations while going backwards (second step of maneuver process), another
appropriate version of fuzzy controller was implemented, also following (Guerrero et al., 2014; Higuti
et al., 2015). Similar to wall-follower control system, a block diagram representing this control system
can be seen in Fig 11.
Fig 11. Block diagram of Reverse Motion Control System
Reverse motion control system’s input is desired orientation (ψd) and its output is the actual
orientation (ψa). This latter value is constantly obtained by using IG500N unit sensor, and it is
negative feedback to the fuzzy controller through variable eψ (orientation error) which is the difference
between ψd and ψa. As mentioned, the value of desired orientation is updated in the beginning of the
second step and should be around 90 degrees off the orientation Helvis left the crop row. Fuzzy
controller’s output is steering angle (u) for Helvis. This fuzzy controller also has an input fuzzy set
with five triangular pertinence functions (Fig 12a) and the same output fuzzy (Fig 12b) as it handles
the steering like in the wall-follower. Here, the five pertinence functions of input fuzzy set are named:
VNOE (Very_Negative_Orientation_Error), NOE (Negative_Orientation_Error), COE
(Central_Orientation_Error), POE (Positive_Orientation_Error) and VPOE
(Very_Positive_Orientation_Error).

Proceedings of the 13th International Conference on Precision Agriculture
July 31 – August 3, 2016, St. Louis, Missouri, USA Page 12
Fig 12. A- Input fuzzy sets. B- Output fuzzy sets.
Analogous to the fuzzy controller of wall-follower control system, a value of orientation error (eψ)
generates two pertinence values (µ1 and µ2) in two different input fuzzy sets at the same time. These
pertinence values are used to cut the output fuzzy set’s pertinence function according to the fuzzy
rules described in equation 22.
If eψ = VNOE or if eψ = NOE, then µ = LS
If eψ = COE, then µ = NS (22)
If eψ = POE or if eψ = VPOE, then µ = RS
A better understanding of equation 22 can be achieved analyzing Fig 6: in reverse motion, negative
eψ requires steering to the left while positive eψ needs steering to the right in order to correct robot’s
orientation. With the output fuzzy values obtained when pertinence values cut output fuzzy set’s
pertinence functions, the defuzzified output value (Z*) can be found using the centroid method
(Guerrero et al., 2014). Again, the Z* is fuzzy controller’s output (u) and it is a steering angle for the
robot.
Detection of Crop Rows in Reverse Motion
This subsection describes how crop rows are detected in reverse motion. For this purpose, rear
LiDAR readings are used. As the filter used to create virtual walls, an interest region of LiDAR
readings is defined. Location of the interest region depends on the maneuver used to change the
crop row. For the first maneuver, interest region comprehends the obtained readings of steps located
between the two red arrows shown in Fig 13a.
Fig 13. Interest region used to detect the crop rows.

Proceedings of the 13th International Conference on Precision Agriculture
July 31 – August 3, 2016, St. Louis, Missouri, USA Page 13
Fig 13b shows the interest region (obtained readings of steps located between the two blue arrows)
for the second maneuver. For both cases, angular separation between arrows is 5° (20 steps) and
black arrow indicates the sense of robot’s reverse motion. Initial and final steps of the interest region
for first maneuver are defined by equations 23 and 24, respectively.
_ =900 +󰇡
.󰇢 (23)
_ =_  󰇡 
.󰇢 (24)
Initial and final steps of the interest region for second maneuver are defined by equations 25 and 26,
respectively.
_ =180 +󰇡
.󰇢 (25)
_ =_ +󰇡
.󰇢 (26)
In order to obtain vectors with projections of LiDAR readings for the first maneuver (X_left_region
and Y_left_region), equations 10, 11, 23 and 24 are used. For the second maneuver, vectors with
projections of LiDAR readings (X_right_region and Y_right_region) are obtained using equations 8, 9,
25 and 26. To detect a crop row for the first maneuver, a filter (equation 27) is applied to
X_left_region and Y_left_region vectors.
if ((X_left_region (stepfilter) > 0) && (X_left_region (stepfilter) < 0.02))
{
if ((Y_left_region (stepfilter) > -1.7) && (Y_left_region (stepfilter) < -0.6))
{
interest_points = interest_points + 1;
}
}
(27)
Where interest_points is the count of filtered points. If the value of interest_points is greater than 5
then a possible crop row was detected. In order to consider that a real crop row was detected, three
consecutives possible crop rows might be detected. For the second maneuver, the used filter is
described by equation 28.
if ((X_right_region (stepfilter) > 0) && (X_right_region (stepfilter) < 0.02))
{
if ((Y_right_region (stepfilter) > 0.6) && (Y_right_region (stepfilter) < 1.7))
{
interest_points = interest_points + 1;
}
}
(28)
Fig 14. shows the moment when a crop row was detected in a test performed. The green lines
represent the interest region of LiDAR readings, blue points are Cartesian projections of rear LiDAR
and red crosses are the filtered points used to determine if a crop row was detected. Finally, the
position of rear LiDAR is represented with a filled orange square.

Proceedings of the 13th International Conference on Precision Agriculture
July 31 – August 3, 2016, St. Louis, Missouri, USA Page 14
Fig 14. A crop row is detected.
EXPERIMENTAL RESULTS
In the mentioned farm-like scenario, several tests were carried out to improve overall system
algorithms, define parameters, and evaluate filters and controller systems described in previous
sections. Except for rainy days, occasions when tests are cancelled as the robot is not waterproof,
environment factors such as temperature and luminosity have not interfered in Helvis performance in
a perceptible way.
A successful test is one in which the robot drives through five crop rows with bounded distance error
and response to disturbances quick enough to avoid crashes with the surroundings. When the robot
arrives in the end of crop row, it needs to turn, alternately to the right and to the left, and enter the
following crop row.
Both Fig 15. and Fig 16 represent a successful complete course around the five crop rows in
scenario and show the fuzzy controller output for the steering system. The first depicts distance error
when the robot is moving along the crop rows (sections ○;A○;C○;E○;G○;I), thus trying to follow the
virtual walls. And the second shows the orientation error in the maneuver process (sections ○;B○;D
○;F○;H).
○;A ○;B ○;C ○;D ○;E ○;F ○;G ○;H ○;I
○;I

Proceedings of the 13th International Conference on Precision Agriculture
July 31 – August 3, 2016, St. Louis, Missouri, USA Page 15
Fig 15. Distance error (blue plot) and Fuzzy controller output (red plot) through time
Fig 16. Orientation error (blue plot) and Fuzzy controller output (red plot) through time
As one may observe in Fig 15, the output is proportional to distance error when this value is smaller
than 0.1m. There is a substantial increase in output when distance error is greater than 0.1m, which
can be observed around time 281s in Fig 17a, a zoomed view of section ○;I. This behavior is
expected as a result of the chosen pertinence functions of input fuzzy set for wall-follower, which
already accounts an absolute distance error of 0.1m either as Far (to) Right Wall or Far (to) Left Wall
and begins to consider it as Very Far (to) Right Wall or Very Far (to) Left Wall.
The beginning of sections ○;C, ○;E, ○;G, and ○;I presents a high value of distance error because
maneuvering process does not guarantee that the robot will start row in the exact middle between
crop rows. Rather, because of maneuver, it is expected that the robot starts parallel to rows. And in
the end of these sections, increased oscillatory behavior is expected as less information is available
to create virtual walls.
Sections ○;B, ○;D, ○;F, and ○;H in Fig 16. represent the steps 1, 2, and 3 of the maneuvering
process. A sawtooth shape represents the step 1 and 3 as the robot turns from its actual orientation
until desired one (90° off the initial one). Between sawtooth shapes, reverse motion (step 2) is
adjusted using fuzzy controller with actual orientation error as input. This part can be seen in more
details in Fig 17b.
Fig 17. Zoomed views of: a – Fig 15. b – Fig 16.
Initial experiments were carried out without fuzzy controller for reverse motion. It was set that the
robot should go backwards with zero steering until it counts two rows. Because of uneven surface,
○;A ○;B ○;C ○;D ○;E ○;F ○;G ○;H ○;I

Proceedings of the 13th International Conference on Precision Agriculture
July 31 – August 3, 2016, St. Louis, Missouri, USA Page 16
there was always deviation and the robot either diverged, being too far to consider LiDAR readings of
rows as valid ones, or converged, riding over the row. Although not optimized to quickly diminish the
orientation error, the controller is able to counteract soil disturbances in order to keep the robot more
or less orthogonal to row crops when going backwards, greatly improving chances of successful
maneuver. For this experiment, orientation error was always less than 5°.
CONCLUSIONS AND FUTURE WORKS
This study focused on developing controllers and filters for an application of Helvis in agricultural
context. It is a local navigation system that can be further used with a global one based on GPS data.
Its purpose is, therefore, guaranteeing crop safety at all times, as GPS-based systems can suffer
from signal loss. The wall-follower fuzzy controller presented an overall great response to soil
disturbances, but in the borders of the rows, improvements can be made. First, a more reliable
orientation estimation can be achieved using LiDAR readings and Simultaneous Localization and
Mapping (SLAM) algorithms. As it will take into account surroundings in order to define path
orientation, the possibility of wrong estimations because robot suffered disturbances on its orientation
is mitigated. In addition, for a range of few meters, UTM30LX LiDAR sensor proved to be reliable
under different luminosity conditions. Second, the presented maneuver process is a starting point for
the switching row task. As one of the first maneuvers implemented in Helvis, it relies on few sensor
measurements and simple calculations. However, it has the major drawback of not being able to
overcome environmental changes, such as uneven soil, leading to orientation deviations, and dense
foliage, complicating the correct counting of rows.
Acknowledgements
The authors would like to thank to FINEP, CNPq, CAPES, EMBRAPA and EESC-USP, for their
support in this work.
References
Barawid Jr, O. C., Mizushima, A., Ishii, K., & Noguchi, N. (2007). Development of an Autonomous Navigation System using a
Two-dimensional Laser Scanner in an Orchard Application. Biosystems Engineering, doi:
10.1016/j.biosystemseng.2006.10.012.
Cariou, C., Lenain, R., Thuilot, B., & Berducat, M. (2009). Automatic guidance of a four-wheel-steering mobile robot for
accurate field operations. Journal of Field Robotics, doi: 10.1002/rob.20282.
Cheein, F. A., & Carelli, R. (2013). Agricultural robotics: Unmanned robotic service units in agricultural tasks. IEEE Industrial
Electronics Magazine, doi: 10.1109/MIE.2013.2252957.
English, A., Ross, P., Ball, D., & Corke, P. (2014). Vision based guidance for robot navigation in agriculture. In Proceedings of
EEE International Conference on Robotics and Automation (ICRA) (website). Hong Kong, Korea: IEEE. doi:
10.1109/ICRA.2014.6907079.
FAO (2015). The State of Food Insecurity in the World. Online Document. FAO. http://www.fao.org/3/a4ef2d16-70a7-460a-
a9ac-2a65a533269a/i4646e.pdf. Accessed 21 April 2016.
Foley, J. A., Ramankutty, N., Brauman, K. A., Cassidy, E. S., Gerber, J. S., Johnston, M., Mueller, N. D., O’connell, C., Ray, D.
K., West, P. C., Balzer, C., Bennett, E. M., Carpenter, S. R., Hill, J., Monfreda, C., Palasky, S., Rockstrom, J., Sheehan, J.,
Siebert, S., Tilman, D., &Zaks, D. P. M. (2011). Solutions for a cultivated planet. Nature, doi:10.1038/nature10452.
Guerrero, H. B., Velasquez, A. E. B., Barrero, J. F., Côco, D. Z., Risardi, J. C., Magalhães, D. V., & Becker, M. (2014).
Orientation (Yaw) Fuzzy Controller Applied to a Car-Like Mobile Robot Prototype. In Proceedings of the IEEE 5th
Colombian Workshop on Circuits and Systems - CWCAS2014 (website). Bogota, Colombia: IEEE. doi:
10.1109/CWCAS.2014.6994603.
Heraud, J. A., & Lange, A. F. (2009). Agricultural Automatic Vehicle Guidance from Horses to GPS: How We Got Here, and
Where We Are Going. Michigan: American Society of Agricultural and Biological Engineers.
Higuti, V. A. H., Guerrero, H. B., Velasquez, A. E. B., Pinto, R. M., Tinelli, L. M., Magalhães, D. V., Becker, M., Barrero, J. F., &
Milori, D. M. B. P. (2015). Low-cost embedded computer for mobile robot platform based on raspberry board. In
Proceedings of 23rd ABCM International Congress of Mechanical Engineering - Cobem2015 (website). Rio de Janeiro,
Brazil: ABMC. doi: 10.20906/CPS/COB-2015-0336.
Hiremath, S. A., van der Heijden, G. W. A. M., van Evert, F. K., Stein, A., & ter Braak, C. J. F. (2014). Laser range finder model

Proceedings of the 13th International Conference on Precision Agriculture
July 31 – August 3, 2016, St. Louis, Missouri, USA Page 17
for autonomous navigation of a robot in a maize field using a particle filter. Computers and Electronics in Agriculture, doi:
10.1016/j.compag.2013.10.005.
Hokuyo Automatic (2012). Scanning Laser Range Finder UTM-30LX/LN Specifications. Product Specifications. Hokuyo
Automatic co. http://www.hokuyo-aut.jp/02sensor/07scanner/download/pdf/UTM-30LX_spec_en.pdf. Accessed 20 April
2016.
Maxon Motors ag (2013).EPOS 2 Positioning controllers. Application Notes Collection. Maxon Motors ag.
http://www.maxonmotor.com/medias/sys_master/8811457937438/EPOS2-Application-Notes-Collection-En.pdf. Accessed
21 April 2016.
Reina, G., Milella, A., Rouveure, R., Nielsen, M., Worst, R., & Blas, M. R. (2015). Ambient awareness for agricultural robotic
vehicles. Biosystems Engineering. doi:10.1016/j.
SBG Systems (2013).IG-500N. Product Specifications. SBG Systems. http://www.sbg-systems.com/docs/IG-500N-Leaflet.pdf.
Accessed 20 April 2016.
UNFPA (2015). State of world population. Online Document. ONU. http://www.unfpa.org/sites/default/files/pub-
pdf/State_of_World_Population_2015_EN.pdf. Accessed 20 April 2016.
Velasquez, A. E. B. (2015). elvis III – Desenvolvimento e Caracterização da Plataforma Robótica. Masters dissertation.
University of Sao Paulo. http://www.teses.usp.br/teses/disponiveis/18/18149/tde-05052015-101452/es.php. Accessed 21
April 2016.
Weiss, U., & Biber, P. (2011). Plant detection and mapping for agricultural robots using a 3D LiDAR sensor. Robotics and
Autonomous Systems, doi: 10.1016/j.robot.2011.02.011.