PERCEPTION OF WHEELED MOBILE ROBOTS IN AGRICULTURE: PHENOTYPING AND MOBILITY ASSESSMENT
Abstract
Agricultural production must double by 2050 in order to meet the expected food demand
due to population growth. Within the technologies employed and developed to
achieve this goal, research on development of intelligent automated or semi-automated
mobile robots capable to navigate in agricultural scenarios acquiring physiological
data about the plants has proven to produce successful results in terms of efficiency
and productivity. In these vehicles, environmental perception is a keypoint to obtain
information not only about the crops, but also about their surroundings and the mobility
status of the robot itself. Within this sensing problem, two perspectives are
identified: agricultural and robotics. The first considers the robot sensors as means to
measure or estimate diverse parameters of the plants, in a phenotyping scheme. The
robotics perspective, on the other hand, aims to use the acquired information for the
robot navigation. This Thesis provides a comprehensive study and real applications
of both perspectives.
Sensors that can be mounted on a robot and used for crop phenotyping are first
reviewed and two specific tests cases are presented. Both provide novel applications
for structural and physiological assessment of crops. The first studies the effects of
using incomplete data acquired from a 2D laser range finder to estimate the treetop
volume of fruit trees. The other application case presents the development and
validation of a sensor fusion methodology to get 3D and thermal representations of
trees. The final result is a point cloud where each point has a temperature value
associated, providing a tool to jointly assess structural and physiological parameters
of the tree.
The robotics perspective focuses on the characterizing the terrain and its effects
on the mobility of the vehicle. As agricultural environments are in general off-road,
traversability of the robot can easily become tough and dangerous. Terrain perception
is then studied using descriptive and dynamic approaches. It is proposed a terrain
classification system to first descriptively characterize the terrain in front of the robot.
As the total cost of the solution is an important matter for commercial adoption, a
low cost sensor was employed. The proposed implementation showed to be robust in
field testing with changing illumination conditions, yielding high accuracy rates.
The dynamic terrain characterization is addressed by off-line identifying a kinematic
model that accounts for non-zero slippage. The parameters of this model are
considered as random variables whose posterior distributions are approximated using
a Particle Markov Chain Monte Carlo method. Contrary to traditional approaches
where punctual estimations are obtained, this point of view can enable a probabilistic
motion assessment with uncertainty propagation not only to the robot positioning but
also to other variables (e.g., wheel slip velocities). Extensive simulation and experimental
tests were used to validate and to compare the proposed methodology with the
Integrated Perturbation Error Dynamics approach. Results showed that the proposed
methodology provides specially satisfactory results when driving an earthmoving
machine through changing terrains.
Collections
- Arq_paso [212]