Completed projects

Master thesis

Visual odometry with deep learning

Candidate:  Sara Lucia Contreras Ojeda
Supervisor: Prof. Marcello Chiaberge
Date:  Settembre 2022

Visual Odometry (VO) is a technique that allows knowing accurately the position of a robot over time, useful, for instance, for motion tracking, obstacle detection, avoidance, and autonomous navigation. To do these tasks requires the use of images captured by a monocular or stereo camera on a robot. From these images, it is needed to extract features to figure out how the camera is moving. This can be done in three different ways: feature matching, feature tracking, and calculating the Optical Flow. Once the key feature points are found is possible to do a 3D to 3D, 3D to 2D, or 2D to 2D motion estimation. Over the years many implementations of visual odometry have been done, a common denominator is that they need to be specifically fine-tuned to work in different environments and there is needed for prior knowledge of the space to recover all the trajectory done by the camera. To create a more generalized implementation, able to adapt to distinct environments, and improve the accuracy of the pose estimation, deep learning techniques have recently been implemented to overcome the limitations previously mentioned. Convolutional Neural Networks (CNNs) have proven to give good results for artificial vision tasks; however, VO is not a task that has been solved with this technique. On the other hand, CNNs have been able to solve with good results tasks such as feature detection and Optical Flow, these are included in some approaches to VO estimation, obtaining an improvement in the results. 

Considering this, for the purpose of this work CNNs were used for the estimation of the Optical Flow. This work presents an approach to solving the Visual Odometry problem using Deep-Learning in one of the stages as a tool to calculate the trajectory of a stereo camera in an indoor environment. To achieve this goal there were implemented Convolutional Neural Networks such as RAFT and The Flownet to calculate the optical flow from two consecutive frames, also was calculated the depth map from the right and left camera images of each frame using an OAK-D camera. The aim of this procedure was to extract key feature points from the images over time. The key points of the left image in the first frame were found with a key point feature extractor that in this case was the Fast Algorithm for Corner Detection. Once gotten, the optical flow was used to find the same feature points of the previous left image in the left image of the consecutive frame. 

Then, from the depth map was obtained the disparity and with this value were located the same key feature points in the right images of the two frames. The key feature points were used to do triangulation and find the 3D points, with them was possible to obtain the transformation matrix that has the information on the pose of the camera along the period of the measure. The proposed method has been implemented with a prototype robot that is in development at the Service Robotic Center of the Politecnico di Torino (PIC4SER), which will have the task of measuring the levels of CO2 in indoor environments with the aim to create an autonomous system capable of purifying the air of these spaces when is needed. The camera was put on the robot and with this system indoors courses were done. The movement of the robot was controlled by a person with a joystick and the odometry was captured using ROS2. The success of the Visual Odometry estimated from the proposed methodology in this work was compared with the odometry of the robot, obtained with ROS2, and plotted using MATLAB.

Path planning algorithm for an autonomous air sanitizing mobile robot in indoor scenarios

Candidate:  Carlo Barbara
Supervisor: Prof. Marcello Chiaberge
Date:  Settembre 2022

Robotics is now a well-established but constantly evolving sector which has reached multiple field ranging from large manipulators for industry applications to smarter mobile robots for housework that requires greater human interaction capabilities. An interesting branch in the robotics world is the Service Robotics, whose main purpose is to help people and improve their quality of life. Mobile robots are widely used in this area, they have to carry out tasks that require them to navigate in known and unknown spaces, thus are of paramount importance the abilities to measure the space around them with different kind of sensors (Lidar, camera, TOF, . . . ) and to try to have a perception of their position in the 3D-space (Odometry). In the past two years, one of the biggest problems that has involved our society is Covid-19, a dangerous virus that makes the air its main vector of contagion. 

The objective of this thesis work, which is part of a collaboration between the PIC4SeR (Politecnico di Torino Interdepartmental Center for Service Robotics) and the Innovation Center of Intesa Sanpaolo, is to develop an autonomous mobile platform, equipped with CO2 sensor and an air sanitizer, capable of performing the sanitization of the air in a working environment (offices, meeting rooms, etc.) to limit the risks of Covid-19 contagious in indoor scenarios. In details, a simulated environment with different rooms and corridors was created. Then, after creating a detailed map of the working environment, a path planning algorithm has been developed using ROS2 and NAV2. This custom algorithm optimizes the path that the robot has to follow when it moves from one room to another, but the algorithm selects also the shortest path to perform the sanitization process of a room. In addition, in the path planning algorithm, the state of charge of the robot’s battery has been taken into account as well as the possible presence of closed doors between different rooms. The final goal is to perform a shorter and better optimized sanitization.

Person-aware autonomous navigation for an indoor sanitizing robot in ROS2

Candidate:  Vittorio Mayellaro
Supervisor: Prof. Marcello Chiaberge
Date:  Settembre 2022

In recent decades, intelligent systems have increasingly become part of our everyday lives to the point that robots able to perceive their surrounding environment and interact with it are not a dream anymore. Nowadays, robots are employed extensively in a variety of industries, including manufacturing, packing, transportation, search and rescue, healthcare and surgery. The usage of robots in social contexts, on the other hand, is still in an earlier stage. The areas of localization, mapping, and exploration for autonomous mobile robots have been the subject of substantial research, mainly focused on unknown environments where the robot is able to build 2D map based on its sensor’s output. In particular, in the last few years there has been an increase in research on autonomous navigation in social environments, which is noteworthy in terms of how much attention this subject is receiving recently. Person-aware indoor navigation focuses on the ability of the robot to automatically detect a person, its position and velocity in real time. Indeed, this skill is crucial for people-aware indoor mapping, obstacle avoidance and path planning. This thesis project aims to improve the navigation strategy of a sanitizing robot by introducing a person-aware module. 

Through the usage of ROS2 the local and global costmap of Nav2 are modified to address the presence of people and maintain an acceptable social distance. A research about the autonomous navigation problem in social contexts has been carried out and the solution we propose uses computer vision to detect people and distinguish them from static obstacles. The developed algorithms are subsequently used to determine the optimal path through the environment by combining multiple probabilities of success based on each sensor’s output. This project is linked to a collaboration between the Interdepartmental Centre for Service Robotics of Politecnico di Torino (PIC4SeR) and the Innovation Centre of Intesa San Paolo. The proposed solution intends to build the groundwork for more extensive and focused approach to tackle the issue of autonomous navigation in social contexts.

Design of a behavior-based navigation algorithm for autonomous tunnel inspection

Candidate:  Riccardo Tassi
Supervisor: Prof. Marcello Chiaberge
Date:  Settembre 2022

The modern world is characterized by the presence of a significant amount of subterranean infrastructures, including networks of tunnels, caverns, and the urban underground. Each of these environments offers intricate settings that could create difficulties for exploration, inspection, and several other activities. Conditions might deteriorate and vary over time, and there may be various risks. Most of the time, this confluence of difficulties and dangers creates circumstances that are too dangerous for workers. Situations like gas leaks, explosions, rock falls, confinement, and prolonged exposure to dust are all potentially lethal. Robotic solutions are thus required to operate when and where human risk is too high. Autonomous robots can lower risks by taking over potentially hazardous jobs for workers. For instance, an autonomous robot can do tasks like assessing the air quality or checking the conditions in hazardous mines.

To achieve this goal, a robot platform must be developed with a series of sensors and algorithms that allow it to navigate autonomously inside the tunnel, collecting the necessary data without any human intervention. This thesis project intends to contribute to this field by designing a robotic platform with the least amount of sensors and algorithms to autonomously accomplish the tunnel inspection task. Specific requirements and environmental difficulties guided the design phase: the rover must be able to cover the entire unknown tunnel plan to perform a good inspection, in the presence of challenging terrains, low-light visibility, and GPS-denied environments. The selected robotic platform consists of a Clearpath Husky rover. A LiDAR sensor is employed to perceive walls and obstacles, while localization is achieved by fusing IMU and encoder data. The main effort was focused on developing the navigation algorithm. It consists of a behaviour-based navigation algorithm that does not require a global map of the environment to work. A state machine interprets LIDAR data and processes specific instantaneous paths tracked by the robot through a Pure Pursuit controller. The paths generation is designed to let the robot always keep the left wall, in such a way, the exploration of the entire tunnel plan is ensured. The navigation algorithm has been tested in simulation with three tunnel models with different sizes and characteristics. In all models tested the robot successfully covered the entire tunnel plans and returned to the starting point. Moreover, the entire system was also successfully deployed and tested on the robotic platform in a real environment.

Prototyping of energetically-autonomous UWB anchors for rover localization in lunar environment

Candidate:  Michelangelo Levati
Supervisor: Prof. Marcello Chiaberge
Date:  Settembre 2022

The space race began in the 1950s, the world’s great powers competed to reach farther and farther frontiers, mainly for political reasons related to the Cold War. From launching satellites into orbit to putting space stations into orbit, from the first man on the moon to conquer the Solar System. On July 20, 1969, Neil Armstrong and Buzz Aldrin left the first footprint on the Lunar surface, while Eugene Cernan, Ron Evans, and Harrison Schmitt were the last humans to leave Earth orbit in December 1972. To this day, the desire to return to the Moon continues, and more and more missions are being planned and completed for the grand return. This thesis has its roots in an expedition to the Lunar surface during which a mobile rover must be able to pinpoint its location relative to a Lunar base. An Ultra Wide-Band Anchor network is chosen for localization. This technology uses wide-band radio waves to transmit information, which can be exploited for high-precision real-time localization by triangulating at least three signals. 

The greater the number of Anchors, the greater the accuracy as noise effects are reduced. The network consists of several Anchors, each equipped with a UWB module, to be placed at strategic locations with the rover itself so that it can communicate at any time with at least three Anchors. Thus, the goal is to prototype a series of Anchors, devices capable of working autonomously after placement on the Lunar ground. The work was divided into four macro areas, including the identification of the best power supply system for the devices, the mechanical design of the Anchors, the electrical design of the Motherboard to manage the systems, and the distribution of the Anchors on the lunar ground. The four design phases were carried out simultaneously to ensure that they matched. For the power supply, a system was designed consisting of five solar panels and five Li-Po batteries, which ensure standby operation of the device even during the period of no light; the design was completed to minimize the size during transportation; the Motherboard is responsible for managing the charging system and powering up the UWB module under specific conditions; and for the method of distributing the Anchors, on the other hand, a code was developed in Python based on QR code recognition to identify the device and pick and place through a six-joint mechanical arm. In addition, the device was realized with the ability to close up to be moved, so that multiple missions of the same type could be carried out in different locations with the same Anchors.

Robust autonomous landing on a high-speed platform

Candidate:  Leoluca Rigogliuso
Supervisor: Prof. Marcello Chiaberge
Date:  Settembre 2022

Research on autonomous landing methods of UAVs on a moving platform has experienced rapid growth in recent years and found applications in civil and military sectors. The extreme precision of drone landing is required to overcome problems related to the low battery autonomy of drones, through landing in mobile charging stations, but at the same time also finding applications in the most varied sectors ranging from parcel delivery to rescue operations. This work implements an autonomous algorithm that allows for the landing onto a vehicle that is moving at high speed, through the use of different sensors, chosen depending on the relative drone-rover position. The result is a robust three state machine that makes use of GPS measurements when the drone-rover distance is large, UWB when the rover is nearby and the fusion of information from the camera and the UWB when the drone is landing. The relative position, computed from the UWB sensors with a Least Square algorithm, must be rotated from the rover’s mobile system to the NED reference frame. 

Therefore, a correct estimate of the orientation of the rover and a consistency between the UAV and UGV compasses is of vital importance for an autonomous landing at high speed. This limit is overcome by mounting a camera on the drone that computes the orientation of the apriltag with extreme precision and this information replaces the noisy one of the rover compass. Kalman filter manages the information coming from the various sensors and generates an estimate of the relative position and relative speed. These are then passed to a PID speed controller that allows accurate and fast tracking and landing on the moving target. Through a purely proportional control over long distances of the rover and a proportional-integrative-derivative control when UAV and UGV are close together, the drone speed value is computed and this is passed to the autopilot which in turn generates the correct thrust of the motors corresponding to that speed. Since the rover landing occurs with a vertical descent after the engines are turned off, a predictive control must be implemented so that the drone predicts the progression of the rover in the next timesteps. The adoption of a predictive control system, the introduction of new sensors and the correction of the misalignment between the compasses of the drone and rover made it possible to reach landing speeds above 30 km/h.

A Deep Learning approach to Instance Segmentation of indoor environment

Candidate:  Riccardo Tesse
Supervisor: Prof. Marcello Chiaberge
Date:  Settembre 2022

Nowadays, mobile robots are frequently used in both indoor and outdoor situations, including agriculture, transportation in industries, surveillance, and cleaning buildings. These are being developed for several applications where long-term capabilities would be advantageous. The primary goal of mobile robotics is to build fully autonomous machines, meaning that they must be able to carry out their jobs without assistance from humans. Their industrial and technical use is continuously becoming more significant, especially when reliability (the uninterrupted and dependable completion of tasks like surveillance), accessibility (the inspection of locations that are inaccessible to humans, such as confined spaces, hazardous environments, or remote sites), or cost are considered. Computer vision is playing a vital part in making these projects more efficient due to the enormous strides that Machine Learning and Deep Learning have achieved in the sector. These innovations significantly altered how tracking and detecting issues are tackled, making real-world applications considerably more practical and successful. The goal of this thesis is to investigate a system that can segment floor plans into individual rooms. 

Several robotics activities depend on this, including topological mapping, semantic mapping, place categorization, human-robot interaction, and automated commercial cleaning. Different map partitioning strategies can be used to complete this task. The Mask R-CNN model has been used to fulfill this target successfully. This network, an extension of Faster R-CNN, enables the prediction of an object mask in conjunction with the branch already in place for bounding box recognition. Since there is not a reasonably large, publicly accessible dataset of floor plans, this thesis’s research involved creating one with the corresponding annotations. The entire dataset containing 4224 images is then used to train the Mask R-CNN model, allowing us to obtain a neural network capable of performing an instance segmentation task on them. Once the model has been trained and validated, a new floor plan map is produced using measurements from a LIDAR sensor. Then, the map is processed using computer vision software to create a crisper and cleaner map of the surrounding area and to prepare it for the segmentation method. This work can be used as a foundation for creating more sophisticated systems capable of automatically classifying rooms (for instance, by including various room typologies in the dataset), or by integrating the algorithm onto a mobile robot to perform segmentation after mapping an entire area.

Hand Gesture Recognition for Home Robotics

Candidate:  Davide Guarneri
Supervisor: Prof. Marcello Chiaberge
Date:  Settembre 2022

Robotics is a sector in deep ferment and constant change. The great interest that this area attracts is due to the ability of robots to carry out demanding and repetitive tasks with higher speed and precision than a human operator, a reason which has led to the strong growth in the development and adoption of large machinery belonging to the important sub-category of industrial robots. However, world society has also undergone stark changes thanks to rapid technological development in all areas, with the result that, although new lifestyles, new needs, and new problems have arisen, novel solutions that can make the life of people easier have also come to light. From this point of view, a particularly active and lively segment called Service Robotics is coming to the fore, ready to bring clear improvements mainly in contexts such as medicine, precision agriculture, logistics, security, the office, the home, and smart cities, settling down as one of the most promising emerging technological trends. The fascinating side of the development of this sector is the incessant propensity to bring robots closer to humans, making them increasingly collaborative and demonstrating over time that they can perform tasks better and better. 

The actual bridge between these two worlds can deservedly be represented by Artificial Intelligence, another technology that is becoming increasingly popular nowadays and allows smartly solving intricate conceptual problems characterized by complex mathematical and computer algorithms behind them. The project presented in this thesis work is an example of the union of these two cutting-edge disciplines and consists of the development of a deep learning model capable of classifying some types of dynamic hand gestures; the interpretation of the performed gesture provided as output by this model will then be used to make a wheeled robot, designed for a domestic environment, perform some specific maneuvering procedures. To achieve this result, recognizing and classifying a frame-by-frame sequence of hand landmarks coordinates, a 2D Convolutional LSTM Deep Neural Network architecture has been chosen, using a softmax layer as the output layer. The advantages offered by this solution mainly reside in the absence of communication interfaces, such as touch screens and joysticks, for controlling the robot and in the reduced amount of data to be processed by an algorithm that is also relatively light in terms of size and required computational capacity; these features allow to obtain a remarkable rapidity in classifying hand gestures and executing actions, that makes this solution combinable with other models for better usability and scalable for different contexts in which gesture recognition can be functional.

Learning Odometric Error in Mobile Robots with Neural Networks

Candidate:  Alessandro Navone
Supervisor: Prof. Marcello Chiaberge
Date:  Settembre 2022

An accurate indoor localization for mobile robotic platforms is fundamental to accomplish autonomous navigation tasks. The most common source of odometric signal for wheeled Unmanned Ground Vehicles (UGV) is provided by the wheels’ encoder sensors and an inertial measurement unit (IMU) if present. However, it is strongly sensible to slip conditions, accumulating error in the robot’s positioning and becoming inaccurate after a short time since it is done integrating over the encoders’ data. Visual odometry techniques can represent a precise alternative for an indoor environment with various features and suitable lighting conditions. However, when the environment offers few or repetitive visual patterns, it is necessary to consider an alternative for reliable UGV localization. Nonetheless, Ultra-WideBand (UWB) anchors recently emerged as a promising solution for indoor robot localization. UWB signals present high precision in Line of Sight (LOS) conditions. On the other hand, they are heavily subjected to a positive bias error in Non-Line of Sight conditions (NLOS) due to obstacles obstruction.

This work aims to study a Machine Learning-based solution to correct and improve robot localization techniques, relying only on raw sensor data. In particular, the study is focused on the correction of two cost-effective sensor signals: wheel odometry and Ultra-WideBand (UWB) anchors ranges. First, a dataset is collected running a Jackal UGV at the PIC4SeR (PoliTO Interdepartmental Center for Service Robotics) laboratory and recording the ROS topics published with the data measured by wheel encoders, IMU, UWB ranges of four antennas, and an Intel t265 visual odometry camera, used as ground truth and validated with the use of a Leica Absolute Tracker AT930. Different neural networks (NN) architectures are considered to correct the received odometry signals. A classic feed-forward dense neural network (NN) for the UWB localization and a Long Short-Term Memory network (LSTM) for the encoder odometry. In both cases, the temporal sequence of previous N sensor data has been adopted to feed the NN models, demonstrating better performances than a single data sample. A grid search for hyper-parameters tuning has been carried out to select the best models on a validation data set. Finally, for both the wheel odometry and UWB, an experimental test set is used to compare the selected best models with an Extended Kalman Filter (EKF) and an analogous NN trained with a single input instead of a temporal sequence of sensor data.

Academic Year 2020/2021

VIO indoor navigation

Candidate:  Chiara Bonanno
Supervisor: Prof. Marcello Chiaberge
Date:  December 2021

Master thesis project, entitled “Visual Odometry Technique for challenging environment with focus on low-texture”, has been carried out at PIC4SeR (Politecnico di Torino Interdepartmental Centre for Service Robotics). Service robotics is based on creating autonomously or semi-autonomously systems useful for the human wellbeing. It can have several applications ranging from precision agriculture to space and to services for people with disabilities. The latter is the field of this research project: visual sensor is mounted on a wheelchair that is moved in indoor environment. Different cameras and Visual Odometry algorithms are evaluated in order to obtain a precise real-time position variation, focusing on low-texture challenge. Generally, indoor environments, such as hospital, do not have many features, so placing a front-facing camera may not solve low-texture problems. The idea of placing the visual sensor facing downwards as well as frontally is proposed, in order to focus exclusively on the floor texture and obtain the desired Odometry with higher accuracy. The purpose of this research is to localize a robot within limited indoor environment through Visual Odometry. 

The algorithm discussed in this Master thesis project consists of two parts. In the first one, consecutive images captured by the camera are considered. Thus, landmarks in the surrounding environment are examined and feature detection and matching algorithms are applied, evaluating the change in position of that feature between one frame and the following. In the second part, the Essential Matrix and the Fundamental Matrix are calculated, obtaining an estimate of the robot’s position point by point. At the end of the algorithm, the trajectory of the wheelchair is obtained. After having created a simulation environment on Gazebo and once transcribed the above-mentioned algorithms in Python language, ROS 2 Foxy is used to simulate the behaviour of the system. The correctness of the robot’s trajectory is verified by comparing the Odometry obtained from Gazebo and the Odometry obtained through the algorithm, by means of a Jackal 3D model. Afterwards, once evaluated the efficiency of the algorithms, system is simulated in real world using a Turtlebot3 Burger robot. Through this experimental approach, it can be demonstrated that a low-texture environment compromises the accuracy of the trajectory; however, with the camera facing downwards, features are greater and the odometry is more accurate.

Face recognition system for service robotics applications with deep learning at the edge

Candidate:  Musaab Awouda
Supervisor: Prof. Marcello Chiaberge
Date:  December 2021

In this thesis, we introduce the problem of facial recognition, which is an active research field in computer vision; many real-life applications like identification, access control, and human-machine interactions require an accurate, fast, and stable face recognition algorithm. With the rise of Deep convolutional neural networks (CNNs), and the massive advancement in relevant technologies (e.g.,smartphones, digital cameras, GPU,…), this task has gained a significant performance improvement. The first stage in the face recognition pipeline is to identify a model capable of detecting and locating a face, if present, in an image or video frame, and then locate the face landmarks (left eye, right eye, nose, upper lip, and lower lip). One of the best tools for face detection is Dlib, an open-source library that provides the best environment for developing software based on machine learning in C++. After face detection, we crop the face, normalize the image, and then feed it to the second stage. In the second stage, we extract features from the detected face in the form of a vector called the embedded vector, which contains enough information to distinguish between different identities. After studying the state-of-the-art models and algorithms, we decided to adopt ArcFace (Additive Angular Margin Loss for Deep Face Recognition). 

The key feature in this model is the design of appropriate loss functions that enhance discriminative power for the extracted features. This model is trained using CASIA, VGGFace2, and MS1MV2 datasets and achieved a verification performance of 99.83% when tested on the LFW dataset. In the last stage, we use the embedded vectors extracted from all the images in our local dataset to carry out the classification and identification. When a new image is to be identified, we extract its embedded vector and calculate its distance from the other vectors in the dataset. Based on the distance value, the system can identify it. Moreover, to distinguish between a real face and an image of a face in front of the camera, we utilize the depth map information extracted from the Intel RealSense depth camera. Finally, a user-friendly GUI (Graphical User Interface) is developed to allow users to add and remove IDs from the Dataset, and to launch the application to do real-time face recognition. This thesis is developed in collaboration with PoliTO Interdepartmental Centre for Service Robotics (PIC4SeR). It is a lightweight application that can run locally and without an internet connection. The application is developed and deployed on the Nvidia Jetson AGX Xavier developer kit, and can be beneficial for many service robotics applications.

Seamless Indoor-Outdoor Autonomous Navigation for Unmanned Ground Vehicles

Candidate:  Andrea Ostuni
Supervisor: Prof. Marcello Chiaberge
Date:  December 2021

Precision Agriculture (PA) concept has been gaining popularity in the last few years, and it is seen as a possible solution to meet the needs of the growing world population. New techniques and approaches are being developed; these innovations are meant to increase productivity and maximize profit while being a foundation of sustainable agriculture. Autonomous robots have a crucial role in this application. In particular, this work is intended to investigate the problem of Autonomous Navigation to provide a solution of seamless Indoor and Outdoor Navigation. An example of employment should be the forage distribution in large farm stables, where the recharge point is usually outside the building. The Navigation problem and solution for the two environments differ. For the Outdoor navigation, a combination of Global Navigation Satellite System (GNSS) and Inertial Measurement Unit supplies enough data for estimating the position and orientation. In Indoor scenarios, the vehicle needs a static map and a range sensor (LiDaR) data to localize itself. Robot Operating System’s (ROS) tools, as services and messages, have been used as a framework to develop and run the application due to its modularity and simple interfaces. In order to obtain good performances, the sensor data are processed by two different filters: a Particle Filter and an Extended Kalman Filter. The first is the foundation of the Adaptive Monte Carlo Localization (AMCL) utilized in indoor environments, while the second is for outdoor application. Then, to navigate seamlessly, a switching algorithm selects which pose estimate to use from the two filters. The feasibility and accuracy of this approach have been tested through several simulations and then deployed on a real rover. The results of these experiments are illustrated in the last chapters of this work.

Design of a system for the detection and monitoring of falling waste

Candidate:  Gabriele Rodovero
Supervisor: Prof. Marcello Chiaberge
Date:  December 2021

Nowadays, waste management is inefficient because of the inadequacy of the waste differentiation and recycling process. This issue causes an increase in costs in economic and especially in environmental terms. Indeed, since it is more economically convenient to produce new objects by using raw materials instead of recycled ones, the waste continues to increase, and so do the polluting emissions of carbon dioxide. This thesis work has been carried out in collaboration with ReLearn, an innovative Italian startup whose mission is to optimize the waste treatment process and consider waste no longer a problem but a resource. To simplify recycling and improve waste management, they have developed a smart bin called Nando that is able to automatically differentiate all the waste inserted inside it. By using robotics and artificial intelligence, Nando recognizes the material of which the waste is composed and then sorts it into the correct bin. 

This thesis project aims to design a system to estimate the volume of objects placed inside Nando. Volume measurements can be useful to monitor the bin fill level and detect possible objects that get stuck falling into the appropriate container. In order to avoid adding additional hardware that would entail additional costs, the measurement has been performed using only the camera already present inside the bin that was used to recognize the waste material. Taking advantage of the recent progress in the field of depth estimation achieved by deep learning methods, it has been possible, starting from a single RGB image of the object captured inside the bin, to predict the corresponding depth image. The Deep Convolutional Neural Network (DCNN) used to estimate depth has been trained on a dataset specifically built for the purpose of the thesis. Subsequently, from the depth image obtained in the previous step, the volume of the waste has been computed with an estimation algorithm specifically developed. Finally, the outcome is a system able to estimate the volume of an object starting exclusively from an RGB image that portrays it. The results of the performed simulations show good scalability of the algorithm and reliable estimation results despite using only a simple camera.

Estimating Depth Images from Monocular Camera with Deep Learning for Service Robotics Applications

Candidate:  Luisa Sangregorio
Supervisor: Prof. Marcello Chiaberge
Date:  December 2021

Estimating depth information from images is a fundamental and critical job in computer vision, as it may be utilized in a large range of applications such as simultaneous localization and mapping, navigation, object identification, and semantic segmentation. Depth extraction can be faced with different techniques: geometry-based (stereo-matching, structure from motion), sensor-based (LiDAR, structured-light, TOF), and deep learning-based. In particular, monocular depth estimation is the challenge of predicting a depth map using just a single RGB image as input. 

This significantly reduces the cost and the power consumption for robotics and embedded devices. However, it is frequently described as an ill-posed problem, since an infinite number of 3D scenes might actually correspond to a single 2D view of a scene. Recently, thanks to the fast development of deep neural networks, monocular depth estimation via Deep Learning (DL), using Convolutional Neural Networks (CNN), has garnered considerable attention, and demonstrated promising and accurate results. With the aim of enabling depth-based real-time applications, this work focuses on lightweight networks, comparing two different CNN models. At the same time, a great effort has been also devoted to the choice of models able to reach state-of-the-art performance on widely used datasets, such as NYU depth and KITTI. Since the behaviour of a neural network highly relies on the training data and the intended context for this network is indoors, two unique datasets have been gathered capturing photos from a mobile robot (TurtleBot3) to better fit the target environment. 

The first one, made with RealSense D435i is composed of pairs of aligned RGB and depth images. The second dataset, made with ZED 2, is composed of left and right stereo pairs which come along depth and disparity ground truth. Fast Depth, is devoted to obtaining a dense depth map in a supervised manner while keeping the complexity and computational effort low. The supervised approach has some weaknesses since requires quality depth data in a range of environments. Monodepth faces depth extraction as an image reconstruction problem, it learns how to predict a disparity map, starting from a single image to reduce the photometric error. In this case, the learning process is fully unsupervised, and it needs only a stereo pair as input. Moreover, to improve the predicted depth map for collision-avoidance tasks, the loss function has been regulated by a weighting map that considers the information of the nearest obstacles. To enhance the effectiveness of these frameworks in my specific case study, the networks have been trained with the collected custom datasets, both from scratch and fine-tuning the pre-trained weights. Finally, the results have been analysed qualitatively and quantitatively, using the main evaluation metrics such as RMSE, RMSE log, Abs Rel, Sq Rel, and Accuracy between the estimated depth and the ground truth. The effectiveness of the predicted depth map has been tested by developing a simple navigation algorithm for obstacle avoidance with a TurtleBot3. The natural progression of this work may be the integration of the depth estimator with advanced autonomous navigation systems in support of indoor service robotics tasks.

Indoor navigation with vocal assistant: Alexa VS Low-power vocal assistant at the edge

Candidate:  Giulia Bertea
Supervisor: Prof. Marcello Chiaberge
Date:  December 2021

Among the numerous human-machine interaction methods, vocal communication has become very popular in the latest years. When initially brought onto the market, vocal assistants were strictly integrated on portable devices; nevertheless, nowadays it is becoming clear that they can be a useful feature for service robotics. In particular, driving a robot vocally constitutes a more inclusive mean of communication, which guarantees a faster and more straightforward way of asserting a command. Indeed, this technology is beneficial since it allows to tackle the needs of some social groups such as the elderly, visually-impaired or physically-limited people. The main purpose of this thesis work is to analyze and compare two different approaches to vocal navigation, while developing and deploying both on a robotic platform for domestic environments. The first approach exploits Amazon Alexa and the AWS cloud system, to which it needs to connect. This aspect represents the greatest drawback of this approach, since it brings out many issues related to privacy and security; moreover, it requires constant internet service availability. A valuable alternative can be a low-power vocal assistant at the edge, which is therefore locally integrated on the robotic platform, that has been implemented by a team of researchers at PIC4SeR (PoliTo Interdepartmental Centre for Service Robotics). 

This vocal assistant is a compound of different machine learning models for speech recognition and processing. An algorithm for the navigation of the robotic platform is developed and integrated with both vocal assistants. The main functions implemented allow the robot to follow basic navigation instructions and steer towards predefined sets of coordinates, which identify rooms and goals in a hypothetical map. Furthermore, an analysis of the meaning extraction methods exploited by both approaches is presented. Regarding the low-power vocal assistant at the edge, a more powerful and precise module for the action classification, based on natural language processing algorithms, is proposed and integrated into the application. The described module exploits advanced machine learning techniques, such as transformers and Deep Attention Neural Networks, for encoding and classifying sentences into predefined categories of instructions. Finally, after extensively simulating in a virtual environment a service robot guided by the two vocal assistants, some real-world tests are run with the goal of highlighting the limitations and advantages of both approaches. The results obtained open up to various future implementations and show how service robotics can highly benefit from vocal assistants at the edge, especially in indoor environments for assisting elderly, visually-impaired or physically-limited people.

Motion Control Architecture for Service Robotics

Candidate:  Nunzio Villa
Supervisor: Prof. Marcello Chiaberge
Date:  December 2021

In recent years, there has been an increase in the use of robots in everyday human activities, both domestic and professional, which go beyond the industrial sector. The IFR (International Federation of Robotics) defines service robotics as “a robot that operates autonomously or semi-autonomously to perform services useful for the well-being of human beings, excluding the manufacturing sector”. The main applications in the home are the cleaning of domestic environments, care of the elderly, and entertainment of children. In the professional field, on the other hand, the applications are innumerable both in natural environments (land, sea, and space) and in artificial environments (offices, hospitals, and urban infrastructures). Examples of these applications are rescuing in hostile areas, monitoring and data collection actions, inspection, and maintenance, in the logistics field for storage and movement of material, and the medical one for surgery and rehabilitation. The use of robots in these areas intensified in 2020 during the COVID-19 pandemic and is expected to further increase in the coming years as evidenced by the 2020 IFR report [1]. This thesis aims to develop both a software and hardware architecture that can be used in most of the aforementioned cases. Therefore, the common needs of these robots will be analyzed first and then they will be implemented in the final design. The final architecture resulting from this analysis will have to respond to a very important requirement such as the versatility that will be the cornerstone of the project.

Remote Sensing-based vineyard image segmentation with deep computer vision for precision agriculture

Candidate:  Cesar Andres Sierra Pardo
Supervisor: Prof. Marcello Chiaberge
Date:  December 2021

The incorporation of remote sensing in precision agriculture technologies has greatly helped to optimize the productivity and efficiency of both farming and agricultural production processes, especially in applications where the monitoring of an extensive area of land is required and the use of automation and robotics can help reduce the time and costs of it. The inspection of vineyard fields is one of the precision agriculture applications where remote sensing and deep learning techniques are combined into systems that can, autonomously, assess the state of the crop. Recent research incorporates satellite imagery as well as drone-captured images as inputs for Convolutional Neural Networks (CNN) capable of quickly processing them to output the information relevant to the specific application, ranging from the growth state of the plantation to a list of commands that an unmanned ground vehicle can use to move through the field. For both cases, a clear understanding of where the vineyard rows are located is crucial, so a system in charge of the vine identification is needed. This thesis aims to present a deep learning approach to solve the problem of segmenting vineyard rows in remotely sensed RGB images. The proposed goal has been achieved by implementing two different CNN that allow comparing traditional and cutting-edge architecture performances. Since, contrary to most studies in the field, that focus on particular cases, generality is a desired characteristic, a dataset consisting of aerial vineyard images of different grape varieties from several wine-growing regions was gathered to account for variable factors such as the illumination condition, the resolution of the images or the growth stage of the crop. Finally, the proposed solutions have been tested with images describing different scenarios with good results, for which a qualitative and a quantitative comparison is done. However, several issues can be further addressed to increase the model efficiency and performance, making this topic interesting for future work development.

Performance evaluation of operational space control and visual servoing for complex 3D robotic arm applications

Candidate:  Luca Marchionna
Supervisor: Prof. Marcello Chiaberge
Date:  December 2021

Nowadays a lot of effort is put into trying to reproduce human behavior. Thus, neuroscience know-how can be leveraged to create a variety of robotic applications in which neural substrates of motor skill learning are transferred to artificial machines in terms of perception and dexterity. An important benchmark for testing analogies and differences in learning techniques can be presented through a game. In this project, control systems are investigated to allow a robotic arm to play Jenga. In particular, this master’s thesis aims at comparing motion control in Operational Space and Visual Servoing control techniques. Therefore, keywords are path planning, force interaction, mechanical design and sensor fusion. The control systems development is performed through proprioceptive and exteroceptive sensors. Such information allows building the overall strategy for the game, characterized by an analytical footprint combined with empirical considerations. Forces acting on a generic block are studied from a theoretical viewpoint that includes geometrical dimensions, material properties and physical constraints. Afterward, the results of such analysis are used to provide a tactic for playing Jenga. To this end, a force sensor is used with 3-D printed support, directly mounted on the manipulator to provide real-time measurements. In addition, a RealSense camera is attached to the robot’s end-effector in a well-known configuration, also called eye-in-hand. It allows the construction of control systems based on visual information.

 Also in this case, the task involves the design of the camera support equipment. In particular, two control techniques are tested: operational space control and visual servoing. The first scheme consists of a priori planned trajectory in Cartesian space which, receiving a 3-D point inside the robot workspace, generates the waypoints to be followed in order to achieve such pose. This control method guarantees the convergence to the desired pose through a PID controller that ensures a small tracking error. For doing this, the functionalities of MoveIt, a planning framework in Robotic Operating System (ROS), are leveraged through a customized inverse kinematic solver and planning adapter. The second control method is a real-time feedback control law, designed to improve task accuracy. Indeed, it has been designed to respond quickly to world noise, lack of measurements and kinematic tolerances. The eye-in-hand configuration provides the feedback information for the position-based visual servoing, a control scheme for actuating the manipulator according to the object pose. In this case, the controller has to be designed to accept the velocity data generated from the visual control loop. The general approach for system validation involves unit testing on individual components. Then, the different modules are combined and tested together. In conclusion, the experimental results are reported in the last chapter, highlighting analogies and discrepancies with respect to similar works. The main differences arise in the type of manipulator employed, Jenga’s tactic and block extraction.

Person following and visual relocalization for indoor service robotics

Candidate:  Paolo Ruggeri
Supervisor: Prof. Marcello Chiaberge
Date:  December 2021

In the last decades the average quality of life improved all over the globe, leading to an increase of elderly people, that often need assistance in their everyday life. Contextually the field of service robotics has been trying to fulfill this social need, providing solutions which can safely coexist in the house. Inside this scenario, PIC4SeR (PoliTO Interdepartmental Centre for Service Robotics) has been putting efforts in a wide project, with the aim of assisting elder people in their daily routines. In particular, the goal of this thesis was to develop a series of software applications to be run on an Unmanned Grounded Vehicle (UGV), designed to roam safely in a domestic environment. Person following is the task of navigating while maintaining the user framed as he walks. To do so, the robot detects the presence of a person in the camera stream and recognizes the main parts of the human body, exploiting an already trained, light and low resource demanding neural network: PoseNet. Adopting SORT (Simple Online and Real time Tracking), the presence of the same person in multiple successive frames is tracked. Finally, the relative distance between the robot and the person is computed. 

The other main subject of this work was relocalization, correcting the odometry error accumulated while roaming. At first it has been performed via markers. Adopting the April tag system, knowing a priori position and orientation of the marker, it is possible to derive the robot global position and orientation very precisely. Another approach to relocalization was then tried in the last part. Acquiring a database of images when the robot turns on, it can later be compared with a query frame. Finding features matching in two frames, an estimation of the relative pose difference between them can be estimated. This framework, although less precise, has the advantage of working in an unstructured environment, unlike the previous one. It also leaves space for improvements in the way matching couples are selected.

Semantic Scene Segmentation for Indoor Robot Navigation

Candidate:  Daniele Cotrufo
Supervisor: Prof. Marcello Chiaberge
Date:  December 2021

Scene Segmentation is an important component for robots which are required to navigate in an indoor environment. Obstacle avoidance is the task of detecting and avoiding obstacles and represents a hot topic for autonomous robots. To obtain a collision free motion, a robust module for obstacle detection is needed. The objective of this thesis is to make a robot able to navigate autonomously, relying only on visual perception, performing real-time segmentation of the indoor scene. In accordance with the state of the art, the proposed method is based on a Deep Learning model for Semantic Scene Segmentation. A Pyramid Scene Parsing (PSP) Net with a ResNet-34 as a backbone is chosen as a model to train. At first, the backbone has been pre-trained on ImageNet dataset, then, maintaining these weights fixed, the PSP Net is trained on the labeled dataset for semantic segmentation. In order to detect the viable part of the scene with high robustness, binary segmentation is chosen, so pixels al labeled as floor (1) or not floor (0). A 91% Intersection over Unit (IoU) score is achieved on the test set with this approach. Scene Segmentation is an important component for robots which are required to navigate in an indoor environment. Obstacle avoidance is the task of detecting and avoiding obstacles and represents a hot topic for autonomous robots. To obtain a collision free motion, a robust module for obstacle detection is needed. The objective of this thesis is to make a robot able to navigate autonomously, relying only on visual perception, performing real-time segmentation of the indoor scene. In accordance with the state of the art, the proposed method is based on a Deep Learning model for Semantic Scene Segmentation. 

A Pyramid Scene Parsing (PSP) Net with a ResNet-34 as a backbone is chosen as a model to train. At first, the backbone has been pre-trained on ImageNet dataset, then, maintaining these weights fixed, the PSP Net is trained on the labeled dataset for semantic segmentation. In order to detect the viable part of the scene with high robustness, binary segmentation is chosen, so pixels al labeled as floor (1) or not floor (0). A 91% Intersection over Unit (IoU) score is achieved on the test set with this approach. Once the image is correctly segmented, a post-processing is applied, in order to obtain a “pixel-goal” for navigation purposes. Navigation is performed through a proportional controller which links the steering angle with the coordinates of the pixel-goal. The linear velocity is handled by the navigation algorithm too. A ROS2 net with a segmentation and a navigation node is built. The model and the ROS2 net are then deployed on a Jetson AGX Xavier platform and the pipeline is tested on a Turtlebot3 robot. The coefficient of the proportional control is tuned directly with real world tries and multiple tests are performed to analyze the performance, mainly from a qualitative point of view. The best results are achieved in corridor scenarios, with the robot able to avoid obstacles along its path, while stays far enough from the walls. In situation with multiple objects with more complex shapes, such as people and chairs in an office, performances are worse, but the robot still often exploit obstacle avoidance correctly.

UAV Precise ATOL Techniques using UWB technology

Candidate:  Gennaro Scarati
Supervisor: Prof. Marcello Chiaberge
Date:  December 2021

Autonomous landing of an UAV on a mobile platform is currently one of the most explored research areas. It emerges as a powerful solution in many sophisticated civil and military applications where human intervention is not always available or sufficiently responsive. Examples of this are continuous flight tasks or long distances to be covered, where mobile charging stations are needed. It is therefore clear that in all these operations the position accuracy of both UAV and mobile platform is of vital importance. This thesis examines the development of an autonomous landing system based on ultrawide-band ranging sensors. A pose estimation filter and a robust control algorithm are proposed, enabling precise tracking and autonomous landing both on the stationary or moving platform. Firstly, they are tested in a large number of simulated scenarios, where it is possible to model both UAV and UGV dynamics and all sensors with their realistic noise. Finally, the algorithms are implemented on the real embedded system, allowing a landing accuracy from 5 to 10 cm. Ultrawide-band is a suitable technology for service robotics because of its high ranging precision, obstacle penetration capabilities and robustness against interference. 

It is in fact possible to accurately compute the relative position of both UAV and UGV by installing UWB ranging sensors on the two systems. Positioning is achieved through a least squares multilateration algorithm, which takes as input the distances given by the UWB devices and returns the relative position of the two systems in the rover reference frame. This information is very noisy and needs to be rotated in the drone reference frame for control purposes. Therefore, it is then fused with UAV sensors and UGV compass data in a loosely coupled Kalman filter, allowing up to 5cm accuracy when the drone is within 1m from the rover. The filtered relative position estimate is then passed to a gain scheduling PID speed controller, which ensures fast tracking and acceptable overshoot in both chase and landing phases. The first step is handled by a proportional control algorithm, while the second by a proportional-integrative-derivative one. Since the main problem of the proposed landing system is the misalignment between the drone and rover compasses, this control algorithm is designed to be robust with respect to orientation errors, allowing successful landings with errors up to 40 degrees. Finally, a predictive control variant is proposed, in which the indirect UGV speed estimate is used to compute at a 10Hz rate the optimal control setpoint and allow autonomous landings at speeds above 3km/h.

Visual and Tactile Perception to play Jenga with a Robotic Arm

Candidate:  Giulio Pugliese
Supervisor: Prof. Marcello Chiaberge
Date:  December 2021

Visual perception and sensory feedback are key elements in any recent robotic system. Human vision and dexterity have been analyzed and adapted to teach the machines how to perform tasks the way people do it. The scope of this Master’s Degree thesis is the study of computer vision and tactile perception methodologies to realize a multisensory robotic application able to understand its surroundings and to act with precision on the desired target. More in detail, the goal of this project is to enable the robotic arm e.DO, developed by Comau, to play Jenga tower through a control system based on vision and touch. The constraints given by the rules of the game and the accurate physical interaction with objects constitute a challenging framework for this robotic task. The analysis of the task leads to the development of a 3D model of the tower and a dataset of images to train a segmentation neural network. From its detections a visual tracker estimates the pose of the specific objects in the camera. Finally, a force sensor is implemented for feedback. 

The tower is modelled in a synthetic environment with the goal of training an instance segmentation neural network to detect and discriminate each block in the camera image. Using Blender, 2D renderings of the tower’s 3D model constitute an image dataset to train the Yolact neural network and allows it to understand the tower’s configuration from its pictures. Here the game dictates how the tower is built from blocks and how it changes in time by removing them. The employed vision system is the position-based visual servoing, composed by two steps adapted from the library of ViSP, an open source visual servoing platform. The first one consists of a model-based tracker that, given the generic 3D model of an object, finds the real object in the camera frame, learns and detects its keypoints, then follows its movement in subsequent frames; furthermore, the output is the pose (translation and orientation) of the tracked object in each frame. Instead of a single Jenga block, a group of them is tracked for an increased number of visual features, exploiting the tower’s staticity and known geometry. 

The second step, taking that pose as input, is a visual control law computing a twist command (linear and angular velocities) to move the camera and reach the desired pose for the extraction of the tracked object. This is achieved by mounting a RealSense D435 depth camera on the robotic arm in the so-called eye-in-hand configuration to link the movement of camera to the arm. Physical interaction to extract a block by pushing it is studied from the properties of wood material, weights of blocks and their relative positions. After an estimate of friction values a force sensor is mounted on top of the robot’s end effector to take the real measurements. The sensor’s touch feedback allows to abort a potentially disruptive push on the tower’s balance, because the extraction of the real blocks is made harder by their irregularities. These avert from the ideal friction behaviour as some blocks are not free to move but they cannot be detected by vision alone. After developing and testing all software components separately, they are connected through ROS interfaces into a final system which ultimately allows to publish e.DO robot specific commands and actuate its joints. The work presents theory and experiments for each unit and, in conclusion, the obtained results achieved by the complete mechatronic system.

Sensor fusion techniques for service robotic positioning and flight in GNSS denied environments using UWB technology

Candidate:  Cosimo Conte
Supervisor: Prof. Marcello Chiaberge
Date:  December 2021

Drones are usually designed to navigate outdoor spaces, they are often equipped with cameras and other sensors, making them a powerful tool for surveying large areas. In these environments a Global Navigation Satellite System (GNSS) is combined with an Inertial Measurement Unit (IMU) to achieve precise positioning, allowing successful navigation in a 3D open space. During the past years users started to use small size drones in challenging environments, indoor places, inside caves or near bridges, where a GNSS is not always reliable or reachable. Classical positioning techniques are no longer efficient in these cases, so it is necessary to develop new systems to adapt in these situations. Ultra WideBand (UWB) sensors are used to enhance the positioning in closed environments. These sensors allow two tags to exchange signals at high frequency in order to retrieve the distance between each other. It is possible to fuse this information with classical positioning methods to resolve the positioning problem in any scenario. 

The goal of this thesis is to design a system that allows drones to flight in both GNSS enabled and denied environments using UWB tags. This work explores the main positioning techniques used in literature and introduces several sensor fusion algorithms, aiming at comparing them in terms of both accuracy and precision. The final intent is to achieve reliable flight in any situation without disruption of service: referred as seamless flight. 

The capabilities of the proposed method are measured in a simulated environment and then validated on a real quadcopter. In order to easily implement the algorithms, all proposed codes are integrated in the PX4-Autopilot Robotics API. This allows to test the same instances of the system both on the companion computer, a Raspberry Pi 4 on the drone, and on the Gazebo simulation environment on a desktop computer, without changing a single line of code. The results of this work show the outstanding capabilities of Kalman filters to fuse sensor’s information, with a particular focus on their nonlinear variant. Thanks to these methods is possible to obtain a reliable pose estimate using the raw UWB ranging data and augmenting it with predicted velocity estimate, that is vital in achieving stable control.

Deep learning computer vision algorithms for apple localization and tracking-simulation, implementation and validation

Candidate:  Davide Blasutto
Supervisor: Prof. Marcello Chiaberge
Date:  December 2021

The world population is growing faster than ever, with an expectation of 10 billion inhabitants on the planet by 2050. To sustain such rapid growth, the agri-food sector must necessarily improve its production capacity and its efficiency, innovating through technological drivers that aim at optimizing and automating the process, while at the same time embracing approaches that can guarantee the sustainability of the supply chain. In particular, automated and precision agriculture is spreading across the industry, both through fully automated machines and through cooperative robots. Computer vision is the main enabler of this industrial shift, thanks to the enormous improvements the field has experienced through Machine Learning and Deep Learning. These technologies radically changed the way the tracking and detection problems are approached, making real-world applications much more convenient and effective. 

The main industrial applications revolve around localization of crop fruits, assessing of its maturity state and its agricultural needs, building a 3d map of the orchard and being able to navigate autonomously through real-time mapping of the surrounding environment. For every of the listed applications, the foundation of the system is a reliable model that is able to detect and track an object consistently. The best solutions known in the literature to accomplish such a task are Mask-R CNN, SSD and YOLO for the detection part and SORT and DeepSORT for tracking. This thesis aims to study a system capable of detecting and counting the fruits of a crop through an implementation of Y.O.L.O. V4 combined with a SORT tracker. The main goal of the system is to be reliable, fast and flexible. While the system has been tested on the specific case of apple orchards, it is able to perform counts for any type of fruit. Several steps have been taken to develop and validate the system. First, a simulation environment was built to correctly validate the system results in the detection, tracking and counting parts, in order to have a reliable reference. 

This environment was developed with real-world counterparts as a reference, also implementing eight different light conditions representing eight different hours of the day. Second, the performance of the system has been validated in real-world scenarios through videos under different conditions. The last part presents a re-evaluation of the previous work, both virtual and real, by retraining the YOLO model with a dataset provided by PIC4SeR (Politecnico of Turin Interdepartmental Centre for Service Robotics) that features apple orchards located in the countryside near Cuneo. The resulting system shows an average counting error spanning from 7% to 13% both in the simulated and in the real environments, with sensitivity on the measure due to light conditions. After the retraining, the counting error outputs minor improvements in the simulations, while in the real applications it is seeing counting errors spanning from 11% to 6%. This work can be taken as a basis for future developments of more advanced systems, capable of carrying out automatic harvesting, perform crop mapping operations or assess fruit maturity state.

Indoor autonomous navigation for a sanitizing UGV for Intesa Sanpaolo

Candidate:  Zafarion Temurov
Supervisor: Prof. Marcello Chiaberge
Date:  December 2021

Thе COVID-19 pаndemic kеeps sprеading аcross thе wоrld аnd, whilе nаtional gоvernments concentratе on lockdowns аnd rеstrictions to mitigatе thе disastеr, аdvanced tеchnologies cоuld bе еmployed more widely to fight the pandemic. This thesis describes existing robotic solutions that could be employed for pandemic care and presents a systematized description of desired robot properties based on a particular application area and target users. In the first chapter with brief introduction to robots and description of types of sanitizing robots available in the market. 

The choice was made by innovation center of Intesa Sanpaolo bank members after thorough research is Aris K2 robot. In the following chapters this robot was described in detail. The next section describes all possibilities of this robot for further improvements and obstacles to achieve some of them. Finally, discussion ends by describing map segmentation and optimal path planning for disinfection of scanned areas.

Academic Year 2019/2020

Indoor SLAM and Room Recognition with Deep Learning at the edge

Candidate:  Andrea Eirale
Supervisor: Prof. Marcello Chiaberge
Date:  December 2020

In recent years, the development of technology has led to the emergence of increasingly complex and accurate localization and mapping algorithms. In the field of robotics, this has allowed the progressive integration alongside other programs with the most varied functions, from home care to space exploration, designed to provide a service, to support and improve people’s living conditions.
With this goal in mind, the PIC4SeR (PoliTo interdepartmental centre for service robotics) has developed the idea of integrating a SLAM algorithm, for simultaneous self localization and mapping, with a convolutional neural network for room recognition.

The main goal of this thesis project is the development of an algorithm able to lead an unmanned ground vehicle in an unknown, closed domestic environment, mapping it and classifying each room encountered in the process. Low cost sensors and free, open-source software are used to achieve the final result.

For localization and mapping, several techniques are considered, from the classic extended Kalman filter to the more advanced graph-based SLAM. The most adapted ones are further developed, to retrieve a first, raw representation of the environment.\\
The map is then processed with computer vision software in order to obtain a cleaner and clearer plot of the surroundings, and to setting it up for the recognition algorithm.

Finally, a convolutional neural network model is used, alongside to a series of frame images taken by the robot from the environment, to classify each room and provide predictions on the map.

The final algorithm is relatively efficient and lightweight, and opens up to a series of future implementations in the field of service robotics, in domestic environments and in the assistance to elderly and disabled users.

Adaptation of a path planning algorithm for UGV in precision agriculture

Candidate:  Francesco Messina
Supervisor: Prof. Marcello Chiaberge
Date:  December 2020

A great interest in self-driven vehicles has developed in the field of robotics and more generally in the industrial sector in the last few decades, and the number of applications where these vehicles are used is growing strongly. A clear example of this phenomenon are the efforts made by several automotive in terms of research and development, to manufacture increasingly competitive, safe and precise Autopilots. This is also due to cutting-edge on-board computers, the development of highly precise and performing sensors, and the advent of the 5G network, which is also aimed to ensure an unprecedented V2X (Vehicle-to-everything) communication, especially thanks to a new architecture called “network slicing”.
The robotics sector is now focused on providing solutions that may promote and improve the man-machine collaboration in agriculture as well as in other areas closely related to the civil sphere, therefore not only for military uses and space exploration, as it used to be. This process has seen an acceleration this year for the emergence of Covid-19, which has become a global pandemic. This situation has highlighted, now more than ever, the need to rely on autonomous instruments that, for example, transport medical equipment or essential commodities, without running the risk to get in contact with other people or the need for anyone to leave home.
In this context, the aim of this thesis work is the creation of a UGV (Unmanned Ground Vehicle) that can run independently, safely and efficiently through rough and uneven terrain, as for agricultural fields or unpaved roads, by using the RRT* (Rapidly-exploring Random Tree) path planner. The latter manages to accomplish its task also thanks to the collaboration with a UAV (Unmanned Aerial Vehicle)
that, by means of the aerial photogrammetric survey method, can generate highprecision georeferenced orthophotos and DTM (Digital Terrain Model), that are later processed by a GIS (Geographic Information System) software to generate the static binary mask.
In detail, the project is designed to cover different project phases. Starting from the design phase, when the criticalities of the problem are evaluated and possible solutions are identified, to the programming phase, in which a code has been developed that allows the planning of the route adapting to the problems that may arise in this type of terrain, up to the simulation and real test phase in the field.

Visual based local motion planner with Deep Reinforcement Learning

Candidate:  Mauro Martini
Supervisor: Prof. Marcello Chiaberge
Date:  September 2020

This thesis aims to develop an autonomous navigation system for indoor scenarios based on Deep Reinforcement Learning (DRL) technique. Autonomous navigation is a hot challenging task in the research area of robotics and control systems, which has been tackled with numerous contributions and different approaches. Among them, learning methods have been investigated in recent years due to the successful spreading of Deep Learning (DL). In particular, in Reinforcement Learning an agent learns by experience, i.e. through the interaction with the environment where it is placed, avoiding the need of a huge dataset for the training process. Service robotics is the main focus of the research at PIC4SeR (PoliTo Interdepartmental Centre for Service Robotics), where the idea of this thesis project is born as part of a broader project. Under the supervision of Professor M. Chiaberge (PoliTo), member of the group, the thesis embraces the vision of the centre, which is to develop high-tech solutions for peculiar fields such as precision agriculture, surveillance and security, in addition to assist people in their every-day life. As a matter of fact, an autonomous navigation system enables competitive advantages in a wide variety of the applications of interest. Deep Deterministic Policy Gradient (DDPG) is the specific DRL algorithm applied to train an agent in a simulated environment using ROS (Robot Operating System). Training simulations offer different types of scenarios presenting both static and moving obstacles. The main goal of the project is to provide a safe collision-free navigation in an unknown indoor environment. An Artificial Neural Network (ANN) is used to directly select suitable actions for the robot, expressed in terms of linear and angular velocity (ANN output). Input information is composed of robot pose and goal position, in addition to raw images provided by a depth camera. A great focus is also devoted to reduce the computational cost of the model in the training phase, as well as the energy consumption in a potential hardware implementation. For this reason, an efficient architecture of the CNN is studied, paying attention to both desired performances and costs. Firstly, a set of convolutional layers is needed to extract high-level features from depth images. Then fully-connected layers predict the action for the robot. Beside these aspects, also sensor data play a key role in a navigation system. From a research point of view, it is interesting to evaluate the performance of the algorithm when using depth images, compared to other popular implementations based on LiDAR sensor. On the one hand, a camera offers a rich depth information. On the other hand a simple 2D LiDAR is able to cover a wider field of view. The navigation system has been tested in a virtual environment with obstacles. Despite the difficulty of the challenge and the amount of resources required for the development, the system can be considered a good starting point for future works. The implementation of the algorithm on a real robot will be a natural next step for the project.

Deep Learning Methodologies for UWB Ranging Error Compensation

Candidate:  Simone Angarano
Supervisor: Prof. Marcello Chiaberge
Date:  September 2020

Ultra-Wideband (UWB) is being extensively introduced in various kinds of both human and robot positioning systems. From industrial robotic tasks to drones used for search and rescue operations, this high-accuracy technology allows locating a target with an error of just a few centimeters, outperforming other existing low-cost ranging methods like Bluetooth and Wi-Fi. This led Apple to equip the latest iPhone 11 with an UWB module specifically for precise localization applications. Unfortunately, this technology is proved to be very accurate only in Line-Of-Sight (LOS). Indeed, performances degrade significantly in Non-Line-Of-Sight (NLOS) scenarios, where walls, furniture or people obstruct the direct path between the antennas. Moreover, reflections constitute an additional source of error, causing the receiver to detect multiple signals with different delays. The aim of this thesis is to compensate NLOS and multi-path errors and to obtain a precise and reliable positioning system, allowing the development of several service robotics applications that are now limited by unsatisfactory accuracies. Another fundamental goal is to guarantee good scalability of the system to unseen scenarios, where even modern mitigation methods still fail. For this scope, a large dataset is built, taking both LOS and NLOS measurements in different environments and experimenting with different types of obstacles. Then, modern Deep Learning methods are used to design a Convolutional Neural Network that predicts the error of the range estimates directly from raw Channel Impulse Response (CIR) samples. Finally, a positioning test is conducted to verify the effectiveness of the method in a real scenario.

GPS-based autonomous navigation of unmanned ground vehicles in precision agriculture applications

Candidate:  Simone Cerrato
Supervisor: Prof. Marcello Chiaberge
Date:  September 2020

The global population is growing exponentially and the actual agricultural techniques and resources will not be able to feed every person on the Earth in a few years. To account for this serious problem, groups of research are focusing their attention on precision agriculture, because it looks for the improvement of the productivity and efficiency of both agricultural and farming production processes, while reducing the environmental impact, exploiting automation and robotics. The thesis aims to design and develop a solution, based on GPS, for the autonomous navigation problem in precision agriculture, using only few sensors: an Inertial Measurement Unit, a GPS receiver and a depth camera, in order to be cost effective. The proposed goal has been achieved through a system of inter-operating sub-components, that have to share information and collaborate each other in order to provide a complete autonomous navigation. In particular, the main involved entities are: a localization filter, a global and a local path planning algorithms and an obstacle avoidance approach, that have been developed and can cooperate each other by means of the Robot Operating System. Eventually, the proposed solution has been tested in a simulation environment, through different possible scenarios providing good results in each of them. However, it may be considered as a starting point for future improvement in the field of autonomous navigation for precision agriculture.

Academic Year 2018/2019

Deep Reinforcement Learning and Ultra-Wideband for autonomous navigation in service robotic applications

Candidate:  Enrico Sutera
Supervisor: Prof. Marcello Chiaberge
Date: December 2019

Autonomous navigation for service robotics is one the greatest challenges and there’s a huge effort from scientific community. This work is born at PIC4SeR (PoliTo Interdepartmental Centre for Service Robotics) with the idea of facing the aforementioned challenge merging rediscovered and promising technologies and techniques: Deep Reinforcement Learning and Ultra-Wideband technology. Over few past years the world has seen a huge advance in the field of Artificial Intelligence, especially thanks to Machine Learning techniques. The latter include a branch called Deep Reinforcement Learning (DRL) that involves the training of Artificial Neural Network (ANN) from experience, i.e. without the need of huge datasets. Here DRL has been used to train an agent able to perform goal reaching and obstacle avoidance. Ultra-wideband (UWB) is an emerging technology that can be used for short-range data transmission and localization. It can be used in GPS-denied environments, such as indoor ones. In this work UWB has been used for localization purposes. UWB is supposed to be a key technology in future: many giant companies are involved and Apple has already inserted an UWB chip in its latest product. It has been used a differential drive robot as implementation platform. The robot is controlled by an ANN (which has robot pose information, lidar information and goal information as input and linear and angular speeds as outputs) using ROS (Robot Operating System). The ANN is trained using a DRL algorithm called Deep Deterministic Policy Gradient (DDPG) in a simulated environment. The UWB has been used in testing phase only. The overall system has been tested in a real environment and compared with human performances, showing that it is able – in some tasks – to match or even outdo them. There have been satisfying results and it is believed that, although there are strong limitations given by the difficulty of the challenge, the system complies with expectations and constitutes a good baseline for future work.

Optimization of an Ultralight Autonomous Drone for Service Robotics

Candidate:  Simone Silvestro
Supervisor: Prof. Marcello Chiaberge
Date: December 2019

Drone industry is constantly growing and evolving in time. Federal Aviation Administration Aerospace Forecast predicts that drone market volume will be about three times the actual one in 2023. This because UAVs are becoming more and more fundamental in the most various application like agriculture, emergency response, urban planning and maintenance, entertainment, security and so many others. For all these tasks, in the near future a large number of drones will fly over our heads. From here the necessity to build safe, lightweight and autonomous UAVfor the drones colonization of the very lower part of the Earth atmosphere. Furthermore European Aviation Safety Agency fixed to 250 grams the limit of the C0 open UAS, category with quite few restrictions to fly. PIC4SeR (PoliTO Interdepartmental Centre for Service Robotics) wants to keep up with this incoming demand and enter in this market, deciding to invest time and resources on the optimization of an ultralight autonomous drone, with the future intention of full customization with the needs of the end-user for different and specific cases such as service robotics, smart city search and rescue and precision agriculture that are the four fields of action of the interdepartmental centre where this thesis took place. With the base of the first flying prototype and a deep research on the State Of The Art literature on drones, in particular lightweight and autonomous ones, the work done for this thesis was to optimize the drone with special attention to hardware and weight, changing the usual construction method from carbon fiber frame to a newborn design using PCB and 3D printed plastic, and finally to make it intelligent and usable by non-expert users with basic knowledge. The improvements made and overall stability of the new prototype are satisfactory, a solid point from which to develop many ultralight drone based technologies.

Navigation Algorithms for Unmanned Ground Vehicles in Precision Agriculture Applications

Candidate:  Diego Aghi
Supervisor: Prof. Marcello Chiaberge
Date: October 2019

With the rapid growth of the world population over the past years the agriculture industry is asked to respond properly to the exponential augmentation of global demand for food production. To do so, it is easy to infer that the agriculture output must be increased rapidly. However, due to environmental resources limitations, this task is not straightforward, therefore, new techniques aimed to maximize the eciency of every single land in a sustainable way are required. To this purpose, as days pass, more and more researchers are focusing their attention on the development of innovative smart farming and precision agriculture technologies by introducing automation and robotics into the agricultural processes.
This thesis is aimed to provide a low cost solution for an autonomous navigation in an unknown outdoor environment using only a camera sensor. More speci cally, it presents two algorithms able to successfully carry out an autonomous navigation in vineyards adopting machine learning and computer vision techniques. Such algorithms have been developed for applications in vineyards but they can also be employed in orchards and in any other analogous scenarios.
The work proposed in this thesis proved to be very reliable when performing motion planning by elaborating the images acquired by the camera. Nonetheless, it is to be considered as a starting point for further research in this sector.

Person tracking methodologies and algorithms in service robotic applications

Candidate:  Anna Boschi
Supervisor: Prof. Marcello Chiaberge
Date: October 2019

The vital statistics of the last century highlight a sharply increasement of the average life of the world population with a consequent growth of the number of elderly people. This scenario has caused new social needs that the research in the service robotics field is trying to fulfill. Particularly, the idea of this thesis is born at the PIC4SeR (PoliTo interdepartmental centre for service robotics) with the purpose of creating complex service robotics applications to support the autonomous and selfsufficient old people into their house in everyday life, avoiding the task of monitoring them by third parties. This work represents the first steps of a broad project in which many other service tasks will be integrated.
The main argument of this thesis is to develop algorithms and methodologies to detect, track and follow a person in an indoor environment using a small wheeled rover and low cost and available sensors to monitor the target person. Several techniques are explored showing the evolution of these methods along the years: from the classical Machine Learning algorithms to the Deep Neural Network ones. Since the main requirement to be respected is the necessity of real-time results, only few of the analysed algorithms are developed for this project scope and at the end are compared in order to find the best solution with optimal outcomes. The detection and localization are the basis of the person tracking application, done by the robot on which it has been implemented a movement control algorithm and at last it has been introduced an obstacle avoidance algorithm to prevent collisions.

Comparison of Stereo Visual Inertial Odometry Algorithms for Unmanned Ground Vehicles

Candidate:  Roberto Cappellaro
Supervisor: Prof. Marcello Chiaberge
Date: July 2019

PIC4SeR is the Interdepartmental center of PoliTO for service robotics. For its research it emerged the need to investigate indoor localization algorithms, in particular the visual-inertial type. This work aims to study different types of algorithms to assess which one is the best choice for indoor localization with the already available COTS hardware. Although the end application is meant to be UAV, a Jackal UGV is used instead, because it was the vehicle available and it lowered the risks of damages during the testing phase. A MYNTYEYE S stereo camera, with included IMU and IR projector, was available and mounted on the UGV. Three algorithms are considered: a light-weight filter-based VIO framework, ROVIO, and two optimization-based VIO frameworks, VINS-Fusion and OKVIS, that should better accommodate data asynchrony. The algorithms are tested in two different environments making the robot follow two paths multiple times: a linear path in a corridor and a pseudo-rectangular one in a room. The algorithms performance is evaluated by two parameters: the relative error on the total travelled distance and the difference between initial and final position of the UGV. The tests underlined no best algorithm, but a dependence on the environment. As future work, it would be interesting to test a camera using a tight hardware-synchronization with an IMU, since, according to the literature, should be a big source of error.

ROS-Based Data Structure for Service Robotics Applications

Candidate:  Simone Rapisarda
Supervisor: Prof. Marcello Chiaberge
Date: April 2019

The cooperation and communication between different robotic agents is a very powerful tool useful for a lot of implementations, from the mapping of areas (SLAM) to the exchange of data to achieve tasks in a shorter time or with better solutions that can allow to save resources. In particular, the possibility of having an Unmanned Ground Vehicle (UGV) communicating with an Unmanned Air Vehicle (UAV) is something that is in greater demand than ever before. For instance, in agriculture applications this cooperation is quite interesting since the mapping of vineyards can be accomplished by a rover which then shares the data with a drone that can use them to define the parameters needed for its mission planning. The purpose of this thesis is to build a common data structure able to receive information from an agent, regardless the nature of the robot (UGV or UAV), and make these data available for other robots that need to work together to achieve a common task. The first part is an introduction to the ROS environment and explains how to use this tool in order to program the robots and organize the data structure that will be used for the thesis. The second part discusses about UGVs, in particular about the two TurtleBot3 rovers which have been used for this project, the Waffle and the Burger models, and the Jackal rover. In this section it is also explained how these rovers can be controlled using the Pixhawk autopilot, which is usually used with drones and therefore needs to be properly programmed in order to be utilized with UGVs. Finally, the last part focuses on MAVROS and MAVLink protocol and explains how these two tools work and how they can be used as interface to allow the communication between the Pixhawk and the rover.

Academic Year 2017/2018

Remote Monitoring of Robots using Web Based Technologies and ROS Framework

Candidate:  Angelo Tartaglia 
Supervisor: Prof. Marcello Chiaberge
Date: October 2018

World population is growing faster than expected. Every year Earth resources are consumed faster, as proven by the early fall of the Earth Overshoot Day that claims the end of year resources. One of the simplest solutions to this problem would be investing in technologies and innovations towards a smart extraction of what population needs. In the last few years a lot of companies introduced automation and robots to improve production in terms of time and obviously cost. Agricultural world is not distant from this evolution; in fact many of the works attempted are now helped by a set of machines that simplifies human work. Hitherto the application in precision agriculture has been always under the human control; there are a lot of applications that see the participation of a worker and in parallel a machine used for a lot of tasks. In the future may an entire field will be managed by a group of automated robots that work together. Those machines will require a lot of specific features likes autonomous navigation, mapping, visual object recognition and many others. This thesis is part of that project and it regards the detection andclassificationoffruitsforfutureapplicationofauto-harvestingandhealthcontrol. Inamorespecificway apples will be considered in this thesis work for fruit application and methods and algorithms as Y.O.L.O. and Mask R-CNN to do the processing of images will be described.These techniques permits the detection of apples in post processing and in real-time with accuracies that range over 32% to 78%. The final result can be used in future applications for spatial localization of fruits and for the detection of possible diseases. It should be emphasised that, even if the thesis shows the results of the object class apple, the algorithms can be applied in a wide range of objects with the only requirement of a different training images dataset. 

Image processing algorithms for synthetic image resolution improvement

Candidate:  Francesco Salvetti
Supervisor: Prof. Marcello Chiaberge
Date: October 2019

The aim of this thesis is to develop a Machine Learning algorithm for Multi-image Super-resolution (MISR).
Super-resolution is a well known image processing problem, whose aim is to process low resolution (LR) images in order to obtain their high resolution (HR) version. The super-resolution process tries to infer the possible values of
missing pixels in order to generate high frequency information as coherently as possible with the original images.
In general, we can distinguish between two different superresolution approaches: Single-image Super-resolution (SISR) and Multi-image Super
Resolution (MISR). The former tries to build the best LR-HR mapping analysing the features of a single LR image, while the latter takes as input multiple LR images exploiting the information derived from the small differences between the images such as changes in the point of view position and orientation, in the lightening condition and
in the image exposition and contrast.
The thesis had as first objective to take part to the competition called “PROBA-V Super
Resolution”, organized by the Advanced Concept Team of the European Space Agency.
The goal of the challenge was to obtain High-resolution images from low resolution ones
from a dataset of pictures took from the PROBA-V satellite of the ESA.
The work has been developed under the direction of the PIC4SeR (PoliTO Interdepartmental Centre for Service Robotics), which aims to integrate it into its agricultural related projects, for the satellite monitoring of the fields status.
The analysis of the state of the art for what concerns super-resolution reveals that Machine Learning approaches outperform the classical algorithms proposed for this
problem. In particular, Neural Networks have been widely used in literature for Singleimage Super-resolution, while this approach for Multi-image Super-resolution is relatively new. An original model to deal with competition problem has been studied, trained and tested. The obtained results show that the multi-image approach can help in the improvement of existing algorithms for super-resolution. However, several issues can be further addressed to increase the model efficiency and performance, making this particular topic interesting for future work development.

Real-time Monitoring of Robotic Devices by using Smartphones Sensors and the ROS Framework

Candidate:  Davide Brau
Supervisor: Prof. Enrico Masala
Date: Aprile 2019

Nell’ambito della robotica, l’attività di monitoraggio risulta fondamentale per il funzionamento dell’intero sistema, sia esso controllato da un essere umano o da un algoritmo programmato per decidere in maniera autonoma. In questo contesto, con monitoraggio si intende quel processo continuo, o svolto a intervalli regolari, che ha lo scopo di acquisire le informazioni provenienti dalle varie parti del sistema robotico, nello specifico i sensori, utili per poter esercitare un controllo efficace su di esso. Durante questa attività di tesi, l’obbiettivo è stato quello di esplorare le possibilità e le soluzioni adatte a svolgere il monitoraggio, analizzando le funzionalità del frame-work: ROS. Per fare ciò, è stato necessario capire la logica utilizzata all’interno di quest’ultimo nell’interfacciarsi con i vari moduli, ponendo particolare enfasi sull’infrastruttura di comunicazione e sui formati utilizzati per i dati; focalizzandosi in seguito su quelli multimediali, in particolare le immagini. In seguito, si è cercato di mettere in pratica le nozioni acquisite, attraverso lo sviluppo di un’applicazione Android basata su ROS che ha lo scopo di fornire un supporto concreto al monitoraggio remoto. Essa consente di acquisire una grande varietà di informazioni, grazie ai vari sensori presenti sui dispositivi odierni, utili per fornire un supporto aggiuntivo e a basso costo per le operazioni di controllo, nel momento in cui lo smartphone viene messo a bordo di un qualsiasi tipo di drone o rover. Si è quindi cercato di evidenziare le varie criticità dovute sia ai limiti computazionali e di reattività tipici dei dispositivi mobili, sia alle caratteristiche della rete. Questi elementi possono determinare un aumento delle latenze o la perdita dei dati, tollerabili entro un certo limite, nel caso di controllo real-time. Attualmente, l’applicazione consente di acquisire informazioni relative alla cinematica utilizzando: accelerometro, giroscopio, magnetometro e il servizio di localizzazione. Tuttavia, il monitoraggio può anche coinvolgere l’ambiente circostante in cui il robot si muove, per tale motivo vengono acquisiti i dati riguardanti: temperatura, pressione e illuminamento; oltre ad informazioni relative alla rete cellulare, al Wi-Fi e alla batteria. L’applicazione consente anche la visualizzazione da remoto delle immagini acquisite dalle fotocamere utilizzando le codifiche: JPEG e H264. Inoltre, sono stati definiti i vari ROS Service che permettono di regolare i parametri associati ad esse. Durante la fase di sviluppo, sono stati introdotti nuovi tipi di messaggi ROS per il trasporto di informazioni associate alle classi Android, tra cui quelli specifici per il GNSS. È stato possibile sfruttare questi ultimi per la condivisione e l’elaborazione di un insieme di informazioni dettagliate ottenute dai satelliti, note con il nome di “raw measurements” (disponibili negli smartphone più recenti). Questi, se utilizzati in modo appropriato, potrebbero costituire una nuova frontiera nello sviluppo di applicazioni per il posizionamento ad alta precisione, dalle quali molti sistemi robotici potrebbero trarre grande vantaggio. Infine, l’ultima parte ha riguardato le prove con il rover Clearpath Jackal, queste si sono rivelate particolarmente utili per poter verificare in un contesto reale le difficoltà e i limiti che ci ritrova a dover affrontare nello sviluppo di questo tipo di applicazioni.

Machine Learning Algorithms for Service Robotics Applications in Precision Agriculture

Candidate:  Federico Barone 
Supervisor: Prof. Enrico Masala
Date: Aprile 2019

La robotica rappresenta uno dei settori disciplinari maggiormente in crescita nel mondo della ricerca tecnologica. In particolare, la robotica di servizio spicca tra i vari ambiti che studiano tale disciplina. Il presente elaborato si pone l’obiettivo di realizzare un’applicazione web che sia in grado di interagire con un robot da remoto. Questa permette sia di controllare i movimenti effettuati dal robot che le sue potenzialità. Di fatto, il robot presenta alcune caratteristiche dalle quali è possibile ricavare molteplici informazioni, tra cui, principalmente, la sua posizione. Inoltre, mediante l’utilizzo di un dispositivo android, è possibile utilizzare opportuni sensori presenti in esso gestendone i dati multimediali acquisiti. L’interfaccia web, nel suo complesso, ha lo scopo di monitorare da remoto l’ambiente circostante al robot, visualizzando i parametri in opportuni grafici bidimensionali. L’utilizzo del framework ROS (Robot Operating System) è alla base della programmazione robotica di servizio. Con l’emergere dell’Internet of Things (IoT), aumenta l’interesse nel fornire interfacce web che consentano agli utenti di monitorare i loro strumenti da remoto. Pertanto, sfruttando talune tecnologie Web Based è stato possibile realizzare uno strumento, composto dall’ambiente di configurazione ROS e dall’applicativo web, atto al monitoraggio remoto del robot. Usufruendo di uno o più dispositivi android posti sul robot, inoltre, è stato possibile monitorare e analizzare differenti parametri acquisiti dai vari sensori. Effettuando varie misurazioni mediante più dispositivi android, è stato possibile confrontare i dati acquisiti e trarne le dovute considerazioni.

Navigation Algorithms for Unmanned Ground Vehicles in Precision Agriculture Applications​

Candidate:  Jurgen Zoto 
Supervisor: Prof. Marcello Chiaberge
Date: October 2018

Robotics for agriculture can be considered a very recent application of one of the most ancient and important sectors, where the latest and most advanced innovations have been brought. Over the years, thanks to continous improvement in mechanization and automation, crop output has extremely increased, enabling a large growth in population and enhancing the quality of life around the world. Both these factors, as a consequence, are leading to a higher demand for agriculture and forestry output. Precision agriculture defined as the correct management of crops for increasing its productivity and maximizing harvest, is considered the answer to this issue. As a matter of fact, thanks to the development of portable sensors, the availability of satellite images and the use of drones, the collection of data is allowing a vast development in this field. This thesis adresses in general robotics for agriculture in the form of a solution to be applied in order to improve robot mobility, in particular automated path planning in agricultural fields, by proposing a method to classify different parcels of which they are composed and to assign a precise task to the terrestrial unmanned robot.

Obstacle Avoidance Algorithms for Autonomous Navigation System in Unstructured Indoor Areas

Candidate:  Lorenzo Galtarossa 
Supervisor: Prof. Marcello Chiaberge
Date: October 2018

This work aims to implement different autonomous navigation algorithms for Obstacle Avoidance that allow a robot to move and perform in an unknown and unstructured indoor environment.
The first step is the investigation and study of the platform, divided into software and hardware, available at the Mechatronics Laboratory (Laboratorio Interdisciplinare di Meccatronica, LIM) at the Politecnico di Torino, on which it is implemented the navigation algorithm. For what is concerned with the software platform, ROS has been used. The Robot Operating System is an open source framework to manage robots’ operations, tasks, motions. As hardware platform the TurtleBot3 (Waffle and Burger) has been used that is ROS-compatible.
The second step is the inspection of the different algorithms that are suitable and relevant for our purpose, goal and environment. Many techniques could be used to implement the navigation that is generally divided into global motion planning and local motion control. Often autonomous mobile robots work in an environment for which prior maps are incomplete or inaccurate. They need the safe trajectory that avoids the collision.
The algorithms presented in this document are related to the local motion planning; therefore, the robot, using the sensor mounted on it, is capable to avoid the obstacles by moving toward the free area.
Three different algorithms of Obstacle Avoidance are presented in this work, that address a complete autonomous navigation in an unstructured indoor environment. The algorithms grow in complexity taking into consideration the evolution and the possible different situations in which the robot will have to move, and all are tested on the TurtleBot3 robot, where only LiDAR was used as sensor to identify obstacles.
The third algorithm, “Autonomous Navigation”, can be considered the final work, the main advantage is the possibility to perform curved trajectory with an accurate choice of the selected path, combining the angular and the linear velocity (980 different motions), the LiDAR scans 180° in front of the robot to understand the correct direction. The last step is the automatic creation of the map. This map will be analysed and compared with the one defined using the RViz software that is the official software used in ROS environment. The tool is suitable to visualize the state of the robot and the performance of the algorithms, to debug faulty behaviours, and to record sensor data. The improvement of this reactive Obstacle Avoidance method is to successfully drive robots in Indoor troublesome areas. As conclusion we will show experimental results on TurtleBot3 in order to validate this research and provide an argumentation about the advantages and limitations.

Implementation of an Ultralight Autopilot Drone for Service Robotics

Candidate:  Salvatore Romano
Supervisor: Prof. Marcello Chiaberge
Date: December 2018

In this thesis, after a brief introduction on the regulation and classification of UAVs, the main sizing criteria of each component of a multirotor will be shown. Starting from the design constraints and a state of the art of the main flight controllers, the hardware and firmware components chosen for the implementation of an autopilot quadcopter under 250 grams will be described. The sizing of the components will be strongly nfluenced by the weight of each of them and will be flanked by a test on the motor / propeller coupling to evaluate the performance and then to choose the most suitable devices for the purpose. Once the Pixracer
has been identified as the best flight controller for the project, the PX4 firmware and related software for remote control (Mission Planner and QgraounControl) will be described. Finally, after the assembly phase, the evaluation tests of the performance of the aircraft, the problems encountered and the possible solutions
and improvements will be described.