Titolo della tesi: Aerial Swarm Robotics for Precision Agriculture
Aerial swarm robotics represents a highly scalable solution to autonomous systems with collaborative
Unmanned Aerial Vehicles (UAVs) by means of decentralized decision making. The deployment of
such system leads to numerous challenges and most of the developed research is limited to validation
within simplified simulation environments. This produces solutions that cannot be readily applied
to real world applications despite the promising features and demonstrated advantages of swarm
robotics as shown in laboratory settings. In this thesis, our interest lies on the application of aerial
swarm robotics in precision agriculture, with the purpose of using collaborative UAVs to optimize
data gathering and resources utilisation in farms. Moreover, we focus on the active monitoring and
mapping of crop fields implementing on board data analysis decision making. The overall goal is to
deploy swarms of UAVs and achieve an overall error reduction of weed detection over time that is
faster than baselines approaches based on pre-determined mission plans, where UAVs are exploited
as passive sensors.
In this thesis, we seek to bring a significant step into the transition from simulation to real
world deployment. To this end, we develop simulation tools that provide photo-realistic rendering
and high fidelity UAV dynamics. An experimental setup is presented for validation of all simulation
components which are eventually combined into a comprehensive simulation environment.
Deployment of real world hardware is considered to replicate realistic parameters to the developed
simulators.
We first focus on validation of a photo-realistic simulation based on the Unity video game
engine to ensure adequate UAV perception of crop fields. The results are validated by comparing
classification performance of Convolutional Neural Networks (CNNs) on real data after being trained
separately on synthetic and real images. We considered the simulation of sugar beets, sunflowers
and potatoes crops and simulation of Near Infrared Sensor (NIR) data for the sugar beets and
sunflowers crops. The implemented CNNs trained with synthetic data achieve an Intersection Over
Union (IoU) classification performance similar to the one obtained with real data.
Secondly, we develop a simplified a simulation environment to validate a novel swarm strategy for
monitoring and mapping of crop fields. This strategy is based on exploiting the available knowledge
about the field to measure the Information Gain (IG) from visiting new areas, that is, the expected
reduction of uncertainty from further observation of a given area. This results is a parameter-free
control system achieving a significant improvement over the state of the art Reinforced Random
Walk (RRW) approach.
Finally, we implement an information-based monitoring and mapping strategy into a combined
simulation with highly-realistic UAV perception and motion dynamics. The final simulation environment
allowed us to simulate a swarm of UAVs within a field while doing online image classification
and collaborative decision making. The previously developed strategy is adapted to fit into the new
realistic conditions of the simulator and experimental setup. Furthermore, we seek to exploit the
information available by observing areas of interest from multiple Point of view (POV). Thus, the
UAVs prioritize positions where residual uncertainty of information of observation is maximal. This
results in a better performance than a simplified random walk and the pre-planned optimal trajectory.
Preliminary testing with real world UAVs is made to provide validation of the parameters
established on the final simulation environment.