IEEE Robotics & Automation Magazine - September 2017 - 98

(a)

(b)

Figure 1. The AgBot robot shown in a typical farm environment. The large broad-acre fields have rows of straight and parallel crop
stubble. The robot system underwent significant testing during both (a) day and (b) night.

will become time-consuming to monitor, control, and
resupply. To reduce this burden, we describe a vision-based
docking system for autonomously refilling liquid supplies
and an interface for controlling multiple robots.
Background
To date, increasing productivity has largely been achieved
through mechanization with increasingly large and complex
machinery to optimize the human operator's time and crop
genetics and technological advances such as automated steering with precision global navigation satellite systems (GNSS).
However, these large machines cause significant soil compaction damage and present a single point of failure for timecritical farm operations. Many farming operations are also
very time-consuming and repetitive. Robotics has the potential to help address many of these challenges.
Advances in robot technology have fueled interest in
automating agricultural machines [13], [20], [22], [29].
Researchers have observed that automated vehicles no longer
need to be designed around optimizing the effort of a human
driver [23]. These factors open up the possibility of performing farming operations with many small autonomous
machines rather than a few large human-operated ones.
Smaller autonomous machines cause less soil compaction,
can work 24 h/d, and will be akin to having a large human
workforce able to tend each individual plant.
This article presents our research toward the goal of developing teams of lightweight autonomous agricultural robots
for coverage tasks such as applying herbicide or fertilizer. This
research involved extensive on-farm development and field
testing of our robot platforms during both day and night
(Figure 1) on a working farm near Emerald, Australia.
Replacing large human operated machinery with many
smaller robots introduces several challenges; two of which
we address here. The first challenge is making a large fleet of
autonomous robots economically viable, which requires
using relatively inexpensive sensors for localization and
obstacle detection. With this in mind, we make extensive use
98

IEEE ROBOTICS & AUTOMATION MAGAZINE

SEPTEMBER 2017

of stereo vision since it provides a rich source of information,
including color, texture, and structure information, and is
also relatively cost effective, particularly since we use a single
stereo camera pair as input to the guidance, obstacle detection, and docking systems. We describe a vision-based obstacle detection method that can detect novel obstacles while
continually adapting to environmental and illumination variations. We also present a localization system that combines a
low-cost consumer grade global positioning system (GPS)
receiver with a vision-based crop row tracking method that
functions in fields that vary substantially in appearance.
The second challenge is that a large fleet of lightweight and
small-payload robots will become difficult and timeconsuming for the farmer to monitor, control, and frequently
resupply. To relieve the farmer from frequently replenishing
robots with consumables such as herbicide, fuel, and seed, we
also present an automated docking system. To allow the farmer to effectively monitor and control a large number of robots,
we describe a user interface suitable for a large farm.
We tested components of the system throughout three
months of field testing. In this article, we present results from
a coverage experiment performed over two consecutive days
where two real robots and four simulated robots were
required to simulate applying herbicide to a 59-ha field during the fallow period (with only crop stubble present) while
avoiding obstacles and refilling at a single refill station placed
on the field boundary. This work extends previous work [3]
with the addition of a user interface, improved localization
and obstacle detection, and automated docking for refill, as
well as a second operating robot.
System Design
A large number of agricultural robotic platforms have been
developed for weeding and general-purpose use. The Hortibot
[15], Armadillo [21], and BoniRob [29] projects, and the work
of Bakker et al. [31] focus on the systematic and functional
design of highly flexible and modular robots designed to be
suitable for a wide variety of agricultural tasks such as weeding,

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