You are on page 1of 7

2020 6th International Conference on Control, Automation and Robotics

Agri-Cost-Maps - Integration of Environmental Constraints into


Navigation Systems for Agricultural Robots
1st Vignesh Raja Ponnambalam 2nd Jaime Pulido Fentanes 3rd Gautham Das
Faculty of Science and Technology Lincoln Centre for Autonomous Systems Lincoln Centre for Autonomous Systems
Norwegian University of Life Sciences University of Lincoln University of Lincoln
Ås, Norway, 1430. Lincoln LN6 7TS, United Kingdom Lincoln LN6 7TS, United Kingdom
vignesh.raja.ponnambalam@nmbu.no jpulidofentanes@lincoln.ac.uk gdas@lincoln.ac.uk

4th Grzegorz Cielniak 5th Jon Glenn Omholt Gjevestad 6th Pål Johan From
Lincoln Centre for Autonomous Systems Faculty of Science and Technology Faculty of Science and Technology
University of Lincoln Norwegian University of Life Sciences Norwegian University of Life Sciences
Lincoln LN6 7TS, United Kingdom Ås, Norway, 1430. Ås, Norway, 1430.
gcielniak@lincoln.ac.uk jon.glenn.gjevestad@nmbu.no pal.johan.from@nmbu.no

Abstract—Robust navigation is a key ability for agricultural


robots. Such robots must operate safely minimizing their impact
on the soil and avoiding crop damage. This paper proposes a
method for unified incorporation of the application-specific con-
straints into the navigation system of robots deployed in different
agricultural environments. The constraints are incorporated as
an additional cost-map layer into the ROS navigation stack.
These so-called Agri-Cost-Maps facilitate the transition from the
tailored navigation systems typical for the current generation Fig. 1. Typical scenarios for agricultural robots demonstrating different
of agricultural robots to a more flexible ROS-based navigation constraints and characteristics: a strawberry polytunnel with tabletops (left),
framework that can be easily deployed for different agricultural an open field with crop rows (center) and wheel tracks (right).
applications. We demonstrate the applicability of this framework
in three different agricultural scenarios, evaluate its benefits in
simulation and demonstrate its validity in a real-world setting.
Keywords—Agricultural Automation, Constraints, Navigation The recent trend in designing agricultural robots, however, is
systems, Costmaps, Motion and Path Planning. to develop light and versatile agricultural robots designed to
adapt to different tasks in various agricultural environments
I. I NTRODUCTION ([4]). Such agricultural robots are designed to operate in
Autonomous robot navigation in agricultural environments various applications and environments (see Fig. 1), and hence
have traditionally been developed on a case-by-case basis require a much more flexible framework for navigation. The
resulting in many diverse approaches to tackle the problem. ROS navigation stack ([5]) is the most widely used framework
The main reason behind this fact is that agricultural robots for autonomous navigation which is highly flexible and can be
vary greatly in terms of their configuration, operational envi- deployed on variety of mobile platforms. In general, the ROS
ronment, and application. Currently, most agricultural robots navigation stack connects several independent components
rely on high-end Global Navigation Satellite Systems (GNSS), which deal with localization, motion planning, and obstacle
mostly because large open fields are the most common opera- avoidance. The trajectory and motion planning components
tional environment for this kind of system (e.g. [1]). Achieving of the navigation stack uses a unified occupancy grid type
high localization accuracy, in this case, requires Real-Time representation called cost-maps to maintain information about
Kinematic solutions (RTK-GNSS) which are expensive and the robot’s configuration space. In this work, we propose to
not always reliable. Besides, pure GNSS-based solutions are use this generic representation to incorporate, in a unified way,
useful only in situations where no changes to the global plan different application-specific navigation constraints character-
nor obstacles are expected. To deal with local motion planning istic to agricultural environments. We call this feature Agri-
and navigation, additional sensors are required such as lidars Cost-Maps (ACM) which allow for the rapid development of
or cameras ([2]) which are used to detect obstacles and follow robust navigation systems exploiting environmental constraints
a specific field characteristic such as tracks or crop rows ([3]). present in the specific applications.
The majority of agricultural robots and their navigation sys- In particular, the contributions of this paper are as follows:
tems developed so far were designed for a specific application. • a general framework for incorporating application-
*This work is part of the RASberry (Robotic and Autonomous Systems for specific constraints in agricultural surroundings into the
Berry Production) project, partly funded by IUK #103339. standard ROS navigation stack;

978-1-7281-6139-6/20/$31.00 ©2020 IEEE 214

Authorized licensed use limited to: ULAKBIM UASL - Sakarya Universitesi. Downloaded on February 02,2021 at 14:50:32 UTC from IEEE Xplore. Restrictions apply.
• experimental demonstration of benefits provided by the more efficient navigation. However, these methods are tailored
framework resulting in safer navigation and straightfor- for specific cases and applications.
ward constraint definition for three different simulated The usage of specialized cost-map layers to introduce
agricultural scenarios; soft navigation constrains has been proposed before in other
• validation of the approach in a real agricultural setting. applications such as human-robot interaction ([16]) where
The paper is structured as follows: a review of related a social cost-map layer was introduced to improve robot
work in navigation for agricultural robots is presented in navigation behavior around humans. In autonomous vehicles
Sec. II followed by the description of the proposed framework ([17]), the cost-maps have been used to enforce traffic rules
including application-specific lane detection components and such as the side of the road. [18] use cost-maps for outdoor
the general cost-map generator presented in Sec. III. The navigation to model traversability of the environment, while
proposed approach is experimentally evaluated in Sec. IV and other authors have proposed to use cost-map layers to improve
the paper concludes with general summary and discussions of navigation planning from sensory observations. For example,
the future work in Sec. VI. [19] proposes a ‘foresighted’ cost-map to avoiding potentially
cluttered areas. The methodology proposed in this work takes
II. R ELATED W ORK inspiration from the cost-map based approaches and enables
the incorporation of the application-specific constraints into an
The autonomous navigation systems for agricultural robots application-agnostic navigation framework such as the ROS
rely on various sensing and localization techniques ([6]). The navigation stack resulting in more robust navigation systems.
most common solutions are based on GNSS technology and
are often derived from auto-steering systems deployed in III. AGRI -C OST-M APS
precision farming machinery. The authors in [1] present a The ROS navigation stack is the most widely used naviga-
study on the performance of autonomous GNSS-based system tion framework which incorporates robot localization, sensor
for intra-row mechanical weed control which uses a real-time streams, and a goal pose and outputs safe velocity commands
kinematic solution (RTK-GNSS) allowing for achieving local- which are sent to a mobile robot. The planning component
ization errors lower than 10 cm. The work in [7] present RTK- of the navigation stack1 is divided into two different sub-
GNSS based guidance systems for different automated farming components: a global planner which provides a navigation
machinery that can deal with complex paths or terrains.Other plan for the robot based on the environment representation
approaches improve the robustness of GNSS-based systems (static map) from the robot’s current position to the specified
by using additional sources of information such as inertial goal, and a local planner which generates the necessary motion
measurement units (IMU) [8] or by using a combination of commands to follow the global plan as closely as possible
heterogeneous sources of information such as lidar, cameras taking into account unmapped navigation constraints such as
or topological constraints as in the approach presented in [9]. obstacles provided by the robot’s sensors.
The work presented in [10] combines GNSS and laser-based In this framework, navigation constraints are represented us-
localization specifically designed for vineyards. ing cost-maps and an occupancy grid-like representation which
Building upon these precise GNSS navigation methods, stores the costs of traversing different parts of the environment.
other authors have focused their efforts on developing planning Occupied parts of the environment are represented by high
methods that efficiently perform agricultural tasks taking into cost values whilst obstacle-free parts are represented by zero
account both the robot’s motion constraints and those of cost values. The standard Dynamic Window Approach (DWA)
the environment ([11]). Work of [12] demonstrates how the is used as local planner which samples the control action in
task constraints in scenarios with multiple and heterogeneous feasible velocity space and gives a trajectory for each sampled
systems, like automated combine harvesters and infield trans- actions. The local motion planner devises the desired trajectory
portation systems can be incorporated. In all these scenarios, taking the obstacle information in the form of cost-maps. With
however, the constraints are pre-defined by a priori information a set of fine-tuned parameters, the planner evaluates the score
such as geo-located obstacles, crop rows, or field limits. based on the default cost-maps layers (static, obstacle and
The need for a priori constraint definition and the lack inflation) and obtains the desired trajectory with the highest
of obstacle avoidance capabilities of these methods led to score. Therefore the cost function of the DWA local planner
the development of perception-based strategies which detect can be non-smooth which maintains the enough clearance from
environmental constraints using the robot’s sensory system static and dynamic obstacles in the environment.
and produce the desired behaviors. Most of these techniques In this work, we propose to incorporate the agricultural
focus on lane following behaviors using different row detection constraints such as lanes or tramlines into the planning frame-
techniques and sensing technologies as in work of [13]. The work by adding a specialized cost-map layer which we refer
main purpose of such systems is to drive the robot safely to as Agri-Cost-Maps (ACM) (see Fig. 2). There are two
around rows of crops but there were also attempts at using lane components in the proposed approach: a lane detector which
following to reduce traffic impact in the fields ([14]). Selected generates a general direction for the robot to be followed based
approaches combine the feature detection-based methods with
obstacle avoidance capabilities ([15]) to achieve safer and 1 http://wiki.ros.org/navigation

215

Authorized licensed use limited to: ULAKBIM UASL - Sakarya Universitesi. Downloaded on February 02,2021 at 14:50:32 UTC from IEEE Xplore. Restrictions apply.
where the weighted average mean ρN and θN based on
normalized weight wn is written as:
N N
X 1 X
wn = dl .
n=1
dlN n=1 n
N N   (2)
1 X 1 X π
ρN = ρn wn θN = θn + wn .
N n=1 N n=1 2


Hence the direction vector P representing the overall direc-
tion of crop rows or wheel tracks is obtained. The applied
Fig. 2. The standard ROS navigation stack (green components) with the
probabilistic hough transform is used from the OpenCV im-
proposed Agri-Cost-Maps (additional components indicated in yellow). plementation with the parameters maxLineGap, minLinLength,
threshold as h15|20, 275|350, 325|350i for the lane and wheel
track detector respectively.
on the sensory information, and an Agri-Cost-Map generator
which converts that direction into the ACM layer which is
then considered by the local planner when calculating new
motion commands. The lane detection method is application-
specific because it depends on the visual characteristics of the
environment (e.g. crop rows or tramlines). However, the Agri- (a) Real field crop rows
Cost-Map generator is universal regardless of the environment
as long as the direction vector is supplied.

A. Lane Detection
This section describes the components in charge of detecting (b) Simulated wheel tracks
the constrains to be considered for the robot’s navigation, in Fig. 3. Lane detector designed for row following.
different agricultural settings. We specifically focus on three
scenarios where there is a need to consider lane following 2) Polytunnel row detector: In the case of polytunnels, a
constrains for successful navigation. ACM for lane following 2D laser scanner is utilized for traversing the rows as described
can use any kind of lane or crop row detector, since the cost- in our earlier work by [20]. To move along a row, a RANSAC
map generation is done using one single descriptor in the form algorithm is used to filter laser scans and reliably detect two
of a normalized vector in the center of the detected row. parallel straight lines, which represent the plant rows on both
1) Lane and Wheel track detector: We utilize a sides of the robot. The poles are placed in a repetitive pattern
color camera to detect crop rows and wheel tracks and hence a bounding box is used to extract the parallel lines.
in the environment. The camera image is first con- Within the bounding box, two laser scan points are chosen
verted into the HSV space and a set of thresholds randomly to generate a line. Then all the remaining laser scan
applied to each channel h0|0|0,0|110|0,0|55.08|0i and points are examined, whether they fit the line model within
h90|125.66|180, 255|255|255, 160|171.11|145i as min-max the given distance d. If the scan points fit the model then
bounds which discriminate between the green crop rows and they are regarded as a set of inliers (Fig. 4). Each RANSAC
brown/grayish tracks (see Fig. 3). The Hough transform is fitting procedure runs for a total of k iterations (k is set to


applied to the segmented image resulting in several line 100) updating the best inlier set. Therefore, the vector P is
candidates which are then subject to some conditions including obtained as an average of two extracted lines.
a minimum number of inliers and minimum line length. The
N number of lines are represented using the Hesse’s normal
form: ln∈N = [ρn , θn ] in which ρ is the closest distance from
each line ln to the origin and θ ∈ [0, π] represents the angle to
the perpendicular line that passes through origin. The weighted
average line lN =[ρN , θN ] based on the line length dln for N
number of lines at the time t is formulated and the 2D vector


pose P which contains the position at an offset distance from (a) Real field polytunnels (b) Line detection methodology
the robot along the weighted average line lN . It represents the Fig. 4. Lane detector for in-row navigation in polytunnels.
overall direction of the crop rows and it is written in the form:

ρN cos(θN )
 B. Agricultural Cost-map Generation


P =  ρN sin(θN )  . (1) The cost-map generator uses the vector published by the
θN lane detector component to increase the navigation costs on

216

Authorized licensed use limited to: ULAKBIM UASL - Sakarya Universitesi. Downloaded on February 02,2021 at 14:50:32 UTC from IEEE Xplore. Restrictions apply.
both sides of the robot. To do so, it is necessary to define a set
of parameters as an input to the proposed cost-map generator.
The parameters cost-map height and cost propagation area
determine the length and width of the cost-map respectively.
These two parameters give the knowledge of the cost-map
edges, hence the parameter total segment points sets the total
number of line segments connecting those edges of the cost- (a) (b)
map on both sides. Each obstacle (segment) point is inflated in
Fig. 6. (a) The resulting Agri-Cost-Maps for Strawberry Polytunnels (b) and
the cost-map up to “n” times the cost-map radius parameter, Crop Rows environments.
within this inflation there are “n” different value steps or layers
that represent a different increase in the cost value ranging
from 252 for the layer closest to the obstacle to 1 for the where the robot navigates narrow passages formed by table
layer furthest apart from it. The output of the ACM generator tops. All three scenarios are implemented and evaluated in
and its parameters is shown in Fig. 5b. a realistic simulation using the ROS compatible Gazebo
simulator, while the validation of the approach is performed
with an actual robot in real open fields and strawberry
polytunnel. The robotic platform used for the experiments
is Thorvald ([4]), a versatile and lightweight agricultural
robot designed for various agricultural applications. The
costmap generator set the parameters like costmap height
h0.8|0.8|0.9|0.9i, cost propagation area h0.4|0.4|0.8|0.8i,
cost-map radius h0.125i, total segment points h25i and
total layers h4i according to their respective experiments. We
kept the same navigation parameters for all experiments such
as min/max velocities and accelerations. All the simulations
(a) (b) were tested over seven iterations for further evaluation.
Fig. 5. (a) Illustration of cost map values and layers for Agri-Cost-Maps
(cmr: costmap radius) (b) and The layout for the proposed Agri-Cost-Maps. A. Experimental Evaluation
All scenarios feature a topological map following the frame-
The parameters required for the cost-map generator are im-
work presented in [21], which has been defined to facilitate the
plemented using the dynamic reconfigure framework2 which
experimental framework. In each case, the robot’s task is to
aids adapting the agri-cost-map layer for various agricultural
navigate from one node of the map to another using the ROS
settings. The parameters can be tuned at run-time using
navigation stack. We consider the No ACM case, where the
reconfigure gui adapting to the dynamics of each agricultural
robot uses the standard configuration of the navigation stack
field. The resulting agricultural cost-map is used to update
to travel from one node to the other and the ACM case, where
the local grid map and it is influencing only the local base
the robot uses the proposed ACM in the navigation stack. The
planner (DWA in our case) in the navigation stack. Due to
robot performance in each test case scenario is analyzed using
the navigation costs around the robot, the local planner re-
the following metrics to evaluate the impact of ACM on the
plans its path accordingly. Fig. 5a illustrates the resulting cost-
robustness and safety of the navigation system:
maps generated by the ACM component which have exactly
• Path Error - the deviation of the actual robot trajectory
the same effect as the standard obstacle-based cost-maps on
the navigation system. Figures 6a and 6b present examples of from the centre-line of the row (optimal path).
• Trajectory Length - the length of the actual path taken by
the generated ACM for two different scenarios considered in
the experiments. Finally, the standard cost-map plugin features the robot.
• Total Time - the time taken for the robot to navigate from
the update frequency parameter which defines the frequency
in which the map is being updated. We set that value to 20 the starting node to the goal node.
Hz in all our experiments to assure the cost-maps being up to B. Testing Environments
date for their subsequent use by the local planner.
1) Crop Rows: The Crop Rows environment (see Fig. 7a)
IV. E XPERIMENTAL F RAMEWORK was constructed using real images of carrot crop taken from
To validate the ability of ACM to improve the robot’s a color camera mounted on a moving vehicle in the field
navigation abilities, its performance is tested in three different near South Scarle, Lincolnshire, UK. 69 images were stitched
agricultural scenarios: an open field with the robot navigating together into a 20 m long strip which was then replicated
along the crops rows, a grassland field where the robot into identical rows of crop and used as a texture added to
follows an existing wheel track and a strawberry polytunnel the ground plane in the simulator. The entire simulated site
covers 20×23 m area. The width of the robot was adjusted to
2 http : //wiki.ros.org/dynamic reconf igure fit a single row and the robot was equipped with a standard

217

Authorized licensed use limited to: ULAKBIM UASL - Sakarya Universitesi. Downloaded on February 02,2021 at 14:50:32 UTC from IEEE Xplore. Restrictions apply.
RGBD camera (Kinect One) facing the ground at 45◦ angle.
The topological nodes were positioned at the start and end
of every crop row so that ACM can be enabled when the
robot passes through the crop rows only. The global plan was
set such that the robot follows one of the rows (highlighted
in Fig. 7a). Once the camera starts seeing the crop rows, the
lane detector finds the crop rows as discussed in Section III-A.
From the detected crop row, the five layers of agricultural
(a) Wheel Tracks (b) Simulated Polytunnels
cost-maps are generated on either side of the robot at a fixed
distance. Moreover, a similar experiment is also demonstrated Fig. 8. The simulated environment scenarios with the global trajectory (red).
V. R ESULTS AND D ISCUSSION

The experimental results for the simulated and real test


scenarios are presented in Table I and Table II respectively.
The use of ACM significantly reduces the Path Error metrics
with no major differences in Trajectory Length and Total Time.
The standard ROS navigation stack without ACM does manage
to avoid the obstacles but it tends to run over the crop rows
(a) Simulated Crop Rows (b) Robot setup by taking larger rotations. The introduction of ACM enforces
Fig. 7. The robot is equipped with its sensors for the real field testing as tighter local trajectories around the obstacles (see Fig. 12a).
displayed in the highlighted section. From Table I, it can be noticed that when the agricultural
constraints are introduced, the deviation from the center of
on an actual strawberry farm in Norway. The strawberries are
the row is reduced significantly. This leads to smaller path
grown along the crop rows on the ground accordingly the
error values and shorter trajectory lengths although the total
autonomous robot has to follow the paths between the crop
time taken to finish the task were longer when ACMs are
rows for doing any agricultural application. Three crop rows
introduced. This is due to the fact that when the general
with two designated paths between each other were chosen
navigation costs are increased, the local planner takes more
as test tracks (Fig. 11a) covering a 112×2.25 m area. The
time to re-plan the path resulting in reduced crop damage (or
robot width was adjusted to 1.25 m so that it can navigates
more constrained navigation). During the tests in simulated
precisely over the crop rows. The robot is equipped with
crop rows (see Fig. 9a), the robot trajectories with ACM stay
an array of sensors (Fig. 7b) including an RGBD camera
close to the center-line whereas the trajectories obtained from
(Intel Realsense D435) for lane detection and the GNSS smart
the standard navigation system diverge from the crop rows.
antenna (Septentrio Altus NR3) along with Xsens MTi-G-710
IMU for robot localization and way-point navigation. A set of
way-points are given as a sequence of goals to the navigation TABLE I
P ERFORMANCE METRICS FROM THE SIMULATED SCENARIOS ( BASED ON 7
stack, consequently the motion planner plans to reach the given RUNS ).
current way-point until it reaches the end of the testing track.
The lane detector can detect the crop rows during tests so that Scenario Case
Path Trajectory Total
Error [m] Length [m] Time [s]
ACMs are generated online using the published vector.
Crop No 0.30 28.47 114.50
2) Wheel Tracks: The Wheel Tracks environment (see Rows ACM ±0.01 ±0.09 ±0.60
Fig. 8a) was constructed using a satellite image of a grass field Crop 0.25 27.94 119.20
near Nettleham, Lincolnshire, UK with clearly visible wheel ACM
Rows ±0.02 ±0.11 ±2.99
tracks. The entire simulated site covers roughly 200×100 m Wheel No 0.56 29.26 124.14
area. The width of the robot was adjusted to fit the tracks and Tracks ACM ±0.17 ±0.60 ±2.43
the robot was equipped with a standard RGBD camera (Kinect Wheel 0.41 28.94 148.40
ACM
Tracks ±0.24 ±0.07 ±4.10
One) facing the ground at 45◦ angle. The global plan was set Poly No 0.07 23.48 63.70
such that the robot follows profile of a single track (Fig. 8a). tunnels ACM ±0.01 ±1.18 ±9.00
3) Strawberry Polytunnels: The Strawberry Polytunnel en- Poly 0.06 23.53 61.30
ACM
vironment is a real mini strawberry production site located tunnels ±0.01 ±0.02 ±11.60
at the Riseholme campus of the University of Lincoln. The
site features two 30 m long tunnels with 5 rows of tabletops The experiments are repeated with the same setup in
mounted on steel poles forming 6 navigable rows (see Fig. 8b). the wheel tracks and polytunnels environment with results
The width of the robot was adjusted to fit the rows and the presented in Table I. The tests with the wheel tracks (see
robot was equipped with a 2D lidar mounted in front of the Fig. 9b) demonstrate that the robot needs to take smooth curves
robot for localization and mapping. The simulated version of instead of following straight lines like in the former testing
the environment was used in experimental evaluation while environments. As in Fig. 10b, the local planner is able to
the real case was used for validation of the approach. rectify its path thanks to ACM which prevents the robot from

218

Authorized licensed use limited to: ULAKBIM UASL - Sakarya Universitesi. Downloaded on February 02,2021 at 14:50:32 UTC from IEEE Xplore. Restrictions apply.
Simulated Test Case with Open Fields
8
Centerline
7.9 Robot Path without ACM
Robot Path with ACM
7.8
Deviation from Centerline [m]

7.7

7.6

7.5

7.4

7.3
(a) (b)
7.2
Fig. 10. (a) Robot Trajectories in curved sections and (b) Influence of ACMs
7.1 on Local trajectories (Green)
7
-20 -15 -10 -5 0 5 10 15 20
The real field validation (see Table II) in open fields gives the
Translation [m] understanding of limitations of the GNSS-based only way-
(a) Test runs on Simulated Crop Rows point navigation. Firstly, a set of way-points has to be manually
Simulated Test Case with Tramways
chosen as intermediate goals for the navigation system. The
0 Centerline GNSS-based navigation system will always plan to reach the
Robot Path without ACM
Robot Path with ACM upcoming way-point irrespective of the actual crop position
-1
along the row. For this reason, the robot might inflict more
damage to the plants along the crop rows. By using the
Deviation from Centerline [m]

-2

proposed ACM system with an additional colour camera, the


-3
explicit knowledge about the crop rows structure (Fig. 11b)
-4 is given to the navigation system. Therefore the crop damage
and negative impact on soil can be greatly reduced since the
-5
ACM restricts the robot moving away from the baseline of
-6 the crop row. The real crop row fields have rough terrains
with varying elevation levels. This rough terrain causes the
0 5 10 15
Translation [m]
20 25
robot’s orientation to vary causing deviations in GNSS antenna
placement leading to further localization errors and potential
(b) Test runs on Simulated Wheel tracks
misalignment with the crop rows can be corrected by ACM.
Simulated Test Case with Polytunnels
-3.26
Centerline
Robot Path without ACM
-3.28 Robot Path with ACM
TABLE II
P ERFORMANCE METRICS FROM THE REAL VALIDATION SCENARIOS
-3.3
( BASED ON 7 RUNS ).
Deviation from Centerline [m]

-3.32 Path Trajectory Total


Scenario Case
Error [m] Length [m] Time [s]
-3.34 Crop 0.57 112.57 738.00
No ACM
Rows ±0.01 ±0.19 ±4.30
-3.36 Crop 0.44 112.27 823.50
ACM
Rows ±0.04 ±0.17 ±3.47
-3.38 Poly 0.20 23.51 83.70
No ACM
tunnels ±0.03 ±0.10 ±2.30
-3.4 Poly 0.16 23.45 84.90
5 10 15 20 25 30 ACM
Translation [m]
tunnels ±0.03 ±0.19 ±1.50
Poly No ACM 0.27 23.57 92.60
(c) Test runs on Simulated Polytunnels tunnels (Obs) ±0.03 ±0.45 ±15.30
Fig. 9. Robot Trajectories are evaluated with and without introducing the Poly ACM 0.24 23.50 93.30
proposed ACM in all three simulations. It is prominent that the trajectory tunnels (Obs) ±0.02 ±0.26 ±6.40
with ACM is in close proximity with the ground truth path.

large deviations from the center-line. In these experiments, the For the last experiment in a real polytunnel, it is noticeable
cost-map parameter cost propagation area needs to change that the row chosen for field validations has similar trajectory
from 0.4 meters to 0.9 meters to accommodate the curved length compared to simulated polytunnel rows although the
lanes compared to the straight lanes. Therefore, the robot has path error and time taken to complete that similar row are
enough space to turn but also aligned with the center-line at a higher in field tests. In another validation, an obstacle was
same time. Similarly, the tests in polytunnels (see Fig. 9c) added randomly (see Fig. 12a) for further studies in the row
display the tighter navigation when ACMs are introduced. following process. The obstacle is kept in such a way that
In the constrained environments like polytunnels, the error the robot can still traverse through the row by avoiding the
tolerances are critical. Without ACM, the robot tends to get added obstacle. The main utility of the cost-maps in the second
closer to the obstacles therefore ACM assures safer navigation. setup is when the robot sees the obstacle, the robot tries to go

219

Authorized licensed use limited to: ULAKBIM UASL - Sakarya Universitesi. Downloaded on February 02,2021 at 14:50:32 UTC from IEEE Xplore. Restrictions apply.
R EFERENCES
[1] M. Nørremark, H. Griepentrog, J. Nielsen, and H. Søgaard, “The
development and assessment of the accuracy of an autonomous gps-
based system for intra-row mechanical weed control in row crops,”
Biosystems Engineering, vol. 101, no. 4, pp. 396 – 410, 2008.
[2] A. English, P. Ross, D. Ball, and P. Corke, “Vision based guidance for
robot navigation in agriculture,” in 2014 IEEE International Conference
on Robotics and Automation (ICRA). IEEE, 2014, pp. 1693–1698.
[3] F. Rovira-Más, Q. Zhang, J. Reid, and J. Will, “Hough-transform-based
(a) (b) vision algorithm for crop row detection of an automated agricultural
Fig. 11. (a) Open Fields environment in Norway with GNSS tagged test track vehicle,” Proceedings of the Institution of Mechanical Engineers, Part
and (b) Visualization of directional vector along with proposed ACM. D: Journal of Automobile Engineering, vol. 219, no. 8, pp. 999–1010,
2005.
towards the left-hand side for avoiding the obstacle. But due [4] L. Grimstad and P. J. From, “The Thorvald II Agricultural Robotic
to the presence of cost-maps, the system restricts the robot System,” Robotics, vol. 6, no. 4, 2017.
from having the larger rotation (Table II) which is moving [5] M. Quigley, K. Conley, B. Gerkey, J. Faust, T. Foote, J. Leibs,
R. Wheeler, and A. Y. Ng, “Ros: an open-source robot operating system,”
away from the center of the row, thereby providing seamless in ICRA workshop on open source software. Kobe, Japan, 2009, p. 5.
navigation even when there is an obstacle and reducing the [6] N. Shalal, T. Low, C. McCarthy, and N. Hancock, “A review of
chances of failures. A video demonstration for the experiments autonomous navigation systems in agricultural environments,” in SEAg
2013: Innovative Agricultural Technologies for a Sustainable Future,
can be seen here: https://youtu.be/huIuYXq9Pt4. September 2013.
[7] A. Stoll and H. D. Kutzbach, “Guidance of a forage harvester with gps,”
Second Test Case with Added Obstacle
25 Precision Agriculture, vol. 2, no. 3, pp. 281–291, 2000.
Centerline
Robot Path with ACM
Robot Path without ACM
[8] A. English, D. Ball, P. Ross, B. Upcroft, G. Wyeth, and P. Corke,
20 Obstacle “Low cost localisation for agricultural robotics,” in Proceedings of the
2013 Australasian Conference on Robotics & Automation. Australian
Robotics & Automation Association, 2013, pp. 1–8.
Translation [m]

15

[9] M. Imperoli, C. Potena, D. Nardi, G. Grisetti, and A. Pretto, “An


10
effective multi-cue positioning system for agricultural robotics,” IEEE
Robotics and Automation Letters, vol. 3, no. 4, pp. 3685–3692, Oct
5
2018.
[10] S. Marden and M. Whitty, “Gps-free localisation and navigation of an
0
-0.5 0 0.5 unmanned ground vehicle for yield forecasting in a vineyard,” in Recent
Deviation from centerline [m]
Advances in Agricultural Robotics, International workshop collocated
(a) (b) with the 13th International Conference on Intelligent Autonomous Sys-
Fig. 12. Second setup for experiments (a) and the real Polytunnels environ- tems (IAS-13), 2014.
ment depicting a table top rows, a robot, and a random obstacle. (b) and The [11] T. Bakker, K. van Asselt, J. Bontsema, J. Müller, and G. van Straten,
study of deviation from the center-line with/without ACM. “Autonomous navigation using a robot platform in a sugar beet field,”
Biosystems Engineering, vol. 109, no. 4, pp. 357–368, 2011.
[12] M. A. F. Jensen, D. Bochtis, C. G. Sørensen, M. R. Blas, and K. L.
Lykkegaard, “In-field and inter-field path planning for agricultural trans-
VI. C ONCLUSIONS AND F UTURE W ORK port units,” Computers & Industrial Engineering, vol. 63, no. 4, pp.
In this paper, we proposed a concept of Agri-Cost-Maps, a 1054–1061, 2012.
[13] W. Winterhalter, F. V. Fleckenstein, C. Dornhege, and W. Burgard, “Crop
general framework for incorporating application-specific envi- row detection on tiny plants with the pattern hough transform,” IEEE
ronmental constraints into the standard ROS-based navigation Robotics and Automation Letters, vol. 3, no. 4, pp. 3394–3401, 2018.
systems for agricultural robots. The benefits of the proposed [14] G. Avina-Cervantes, M. Devy, and A. Marin-Hernandez, “Lane extrac-
tion and tracking for robot navigation in agricultural applications,” in
framework was demonstrated in simulated experiments in Proceedings of ICAR, 2003.
three different scenarios and validated in two real agricultural [15] D. Ball, B. Upcroft, G. Wyeth, P. Corke, A. English, P. Ross, T. Patten,
settings. The additional constraints enable safer and more re- R. Fitch, S. Sukkarieh, and A. Bate, “Vision-based obstacle detection and
navigation for an agricultural robot,” Journal of field robotics, vol. 33,
liable navigation for agricultural robots. This work introduces no. 8, pp. 1107–1130, 2016.
a basic concept of handling environmental constraints but also [16] D. V. Lu, D. Hershberger, and W. D. Smart, “Layered costmaps for
opens many interesting directions for future work. In our work, context-sensitive navigation,” in 2014 IEEE/RSJ International Confer-
ence on Intelligent Robots and Systems. IEEE, 2014, pp. 709–715.
we have only used a single type of constraints in the form of [17] D. Ferguson and M. Likhachev, “Efficiently using cost maps for planning
straight lanes and further studies will look into other types complex maneuvers,” Lab Papers (GRASP), p. 20, 2008.
of such information. For the wheel tracks involving curved [18] B. P. Gerkey and M. Agrawal, “Break on through: Tunnel-based ex-
ploration to learn about outdoor terrain,” in ICRA Workshop on Path
lanes, the performance can be improved by introducing a Planning on Costmaps, 2008.
curvature model using polynomial equations. The proposed [19] P. Regier, S. Oßwald, P. Karkowski, and M. Bennewitz, “Foresighted
feature can be extended to other crop fields although the ability navigation through cluttered environments,” in Intelligent Robots and
Systems (IROS), 2016 IEEE/RSJ International Conference on. IEEE,
of traditional lane detectors handling the variance efficiently 2016, pp. 1437–1442.
throughout the cropping season are limited. We also used [20] T. D. Le, V. R. Ponnambalam, J. G. Gjevestad, and P. J. From, “A low-
very crude feature detectors for generating the cost-maps in cost and efficient autonomous row-following robot for food production
in polytunnels,” Journal of Field Robotics, 2019.
the presented work. We are currently investigating how the [21] J. P. Fentanes, B. Lacerda, T. Krajnı́k, N. Hawes, and M. Hanheide,
feature detector’s performance affects the overall benefits of “Now or later? predicting and maximising success of navigation actions
the proposed feature. Other directions might include extending from long-term experience,” in 2015 IEEE international conference on
robotics and automation (ICRA). IEEE, 2015, pp. 1112–1117.
this concept beyond the agricultural settings.

220

Authorized licensed use limited to: ULAKBIM UASL - Sakarya Universitesi. Downloaded on February 02,2021 at 14:50:32 UTC from IEEE Xplore. Restrictions apply.

You might also like