Abstract
To tackle wildfires and improve forest biomass management, cost effective and reliable mowing and pruning robots are required. However, the development of visual perception systems for forestry robotics needs to be researched and explored to achieve safe solutions. This paper presents two main contributions: an annotated dataset and a benchmark between edge-computing hardware and deep learning models. The dataset is composed by nearly 5,400 annotated images. This dataset enabled to train nine object detectors: four SSD MobileNets, one EfficientDet, three YOLO-based detectors and YOLOR. These detectors were deployed and tested on three edge-computing hardware (TPU, CPU and GPU), and evaluated in terms of detection precision and inference time. The results showed that YOLOR was the best trunk detector achieving nearly 90% F1 score and an inference average time of 13.7ms on GPU. This work will favour the development of advanced vision perception systems for robotics in forestry operations.

Abstract
Bin picking is a challenging problem that involves using a robotic manipulator to remove, one-by-one, a set of objects randomly stacked in a container. When the objects are prone to entanglement, having an estimation of their pose and shape is highly valuable for more reliable grasp and motion planning. This paper focuses on modeling entangled tubes with varying degrees of curvature. An unconventional machine learning technique, Inductive Logic Programming (ILP), is used to construct sets of rules (theories) capable of modeling multiple tubes when given the cylinders that constitute them. Datasets of entangled tubes are created via simulation in Gazebo. Experiments using Aleph and SWI-Prolog illustrate how ILP can build explainable theories with a high performance, using a relatively small dataset and low amount of time for training. Therefore, this work serves as a proof-of-concept that ILP is a valuable method to acquire knowledge and validate heuristics for pose and shape estimation in complex bin picking scenarios.

Abstract

Abstract
The estimation of a 3D sensor constellation for maximizing the observable surface area percentage of a given set of target objects is a challenging and combinatorial explosive problem that has a wide range of applications for perception tasks that may require gathering sensor information from multiple views due to environment occlusions. To tackle this problem, the Gazebo simulator was configured for accurately modeling 8 types of depth cameras with different hardware characteristics, such as image resolution, field of view, range of measurements and acquisition rate. Later on, several populations of depth sensors were deployed within 4 different testing environments targeting object recognition and bin picking applications with increasing level of occlusions and geometry complexity. The sensor populations were either uniformly or randomly inserted on a set of regions of interest in which useful sensor data could be retrieved and in which the real sensors could be installed or moved by a robotic arm. The proposed approach of using fusion of 3D point clouds from multiple sensors using color segmentation and voxel grid merging for fast surface area coverage computation, coupled with a random sample consensus algorithm for best views estimation, managed to quickly estimate useful sensor constellations for maximizing the observable surface area of a set of target objects, making it suitable to be used for deciding the type and spatial disposition of sensors and also guide movable 3D cameras for avoiding environment occlusions.

Abstract
Robotics in agriculture faces several challenges, such as the unstructured characteristics of the environments, variability of luminosity conditions for perception systems, and vast field extensions. To implement autonomous navigation systems in these conditions, robots should be able to operate during large periods and travel long trajectories. For this reason, it is essential that simultaneous localization and mapping algorithms can perform in large-scale and long-term operating conditions. One of the main challenges for these methods is maintaining low memory resources while mapping extensive environments. This work tackles this issue, proposing a localization and mapping approach called VineSLAM that uses a topological mapping architecture to manage the memory resources required by the algorithm. This topological map is a graph-based structure where each node is agnostic to the type of data stored, enabling the creation of a multilayer mapping procedure. Also, a localization algorithm is implemented, which interacts with the topological map to perform access and search operations. Results show that our approach is aligned with the state-of-the-art regarding localization precision, being able to compute the robot pose in long and challenging trajectories in agriculture. In addition, we prove that the topological approach innovates the state-of-the-art memory management. The proposed algorithm requires less memory than the other benchmarked algorithms, and can maintain a constant memory allocation during the entire operation. This consists of a significant innovation, since our approach opens the possibility for the deployment of complex 3D SLAM algorithms in real-world applications without scale restrictions.

Abstract
Reinforcement Learning (RL) is a well-suited paradigm to train robots since it does not require any previous information or database to train an agent. This paper explores using Deep Reinforcement Learning (DRL) to train a robot to navigate in maps containing different sorts of obstacles and which emulate hallways. Training and testing were performed using the Flatland 2D simulator and a Deep Q-Network (DQN) provided by OpenAI gym. Different sets of maps were used for training and testing. The experiments illustrate how well the robot is able to navigate in maps distinct from the ones used for training by learning new behaviours (namely following walls) and highlight the key challenges when solving this task using DRL, including the appropriate definition of the state space and reward function, as well as of the stopping criteria during training.