IEEE Robotics & Automation Magazine - September 2017 - 99

scouting, phenotyping, and other tasks. The Hortibot is a slope
mower modified for operation in horticulture. The Armadillo
is a tracked robot with a software architecture based on the
Robot Operating System (ROS). The BoniRob is a highly flexible robot that can change its track width and height. Whereas
these projects focused on the functional design of a robot suitable for a wide variety of tasks, in this article, we focus on the
key software components required for an autonomous agricultural robot and mostly use an existing vehicle platform. Another project describes a robot specifically designed for mechanical
weed control, including a system to guide a robot along crop
rows and a vision system to detect weed plants [1].
In this project, our target environment is large broad-acre
fields. The first robot we developed is a general-purpose platform for performing autonomous robot field experiments
named AgBot. The second robot, named SwarmBot, was custom designed specifically for broad-acre farming environments and notably is lightweight and mechanically simple.
The remainder of this section describes the physical and software components of the AgBot robot platform. These
platforms are depicted in Figure 2.
The software architecture is shown in Figure 3, and components of the robot system are described in more technical
detail in [2]. The AgBot has two i5 personal computers
(PCs); one PC runs the navigation, vehicle controller, and
localization modules, and the other runs the obstacle detection module. All PCs run Ubuntu 12.04 and the ROS Hydro
[24]. The robot's navigation system is based on the ROS
move_base architecture that allows the robot to autonomously plan collision-free paths around obstacles to desired

NVS
GNSS
Strobe
Light
5-m
Boom

goal locations. We have modified this architecture so that the
robot returns quickly to the desired path (crop row) after
avoiding an obstacle. We also added the ability to plan to distant goals outside the planner's map. The global planner
plans long-term kinematically feasible and collision-free
paths through the obstacle detector costmap to a desired
waypoint using the search-based planning library lattice
planner (SBPL) [6]. The local planner uses vector-pursuit to
generate the desired robot velocity commands to follow path
output by the global planner.
The AgBot has a forward-facing stereo camera pair,
encoders on the rear wheel, an inertial sensor, and real-time
kinematic (RTK)-GPS, the details of which are in the caption for Figure 3. Part way through the project, we switched
cameras from Power over Ethernet Gigabit Ethernet Imaging Development Systems GmbH cameras (eV2 EV76C560,
60-dB dynamic range, 12-ke- saturation capacity) to the
newer universal serial bus3 Point Grey (Sony Pregius
IMX174, 72-dB dynamic range, 32-ke- saturation capacity).
This change improved the performance of our algorithms
by reducing the amount of image noise in night time and
low-light conditions, as well as allowing higher contrast
images in daytime bright-light conditions.
Obstacle Detection
Obstacle detection systems have been shown to perform well
in a variety of environments using only structure information
[11], [25]. These methods are popular due to the availability
of laser scanners and are largely independent of illumination.
Structure-based methods are less suitable in environments

Novatel
GNSS
Stereo
Cameras

200-L Tank
and Funnel

Vectornav VT100
LMS-111
Laser Scanner

TOPCON RTK GNSS Antenna
Telemetry
Radio

eBox

10-b
Encoders
TE Gator
Vehicle
(a)

(b)

Figure 2. The two robot platforms that were used in our experiments: (a) The AgBot, an electric John Deere TE gator modified for
autonomous operation using a RoPro Design automation kit. The robot has a 200-L herbicide tank with a level sensor and a 5-m-wide
spray boom. Sensors on the robot include a forward-facing stereo camera pair (Point Grey Grasshopper GS3-U3-23S6-C), encoders
on the rear wheels, an inertial sensor (Vectornav VN-100), and a GPS (NVS08C-CSM Global Navigation Satellite System). The robot
receives RTK corrections over a mobile data network from the SmartNetAus Continually Operating Reference Station network. Ground
truth position is obtained using a dual antenna Novatel FlexPack GPS with tactical grade inertial measurement unit (IMU). (b) The
SwarmBot, a custom robot designed by SwarmFarm Robotics for the farm environment. It has a tricycle wheelbase that is driven and
steered via the rear wheel, with the front two wheels providing differential odometry measurements. The robot uses a forward-facing laser
(SICK LMS111), tilted down at an angle of 11° for obstacle detection. The robot has a GPS (Topcon RTK GNSS receiver) and an inertial
sensor (Vectornav VN-100).

SEPTEMBER 2017

IEEE ROBOTICS & AUTOMATION MAGAZINE

Table of Contents for the Digital Edition of IEEE Robotics & Automation Magazine - September 2017