APRÈN. Portal de la producció docent del professorat de la UPC

Automatic Control problems can be extremely challenging for any non-adapting algorithm to solve. Control involves the real-time evaluation of system responses under uncertain circumstances, and can be encountered in different areas of engineering, ranging from simple valve controls to State-of-the-Art autonomous driving. While classical control algorithms have been proven to work, in a trend of creating more complex and larger systems these previous methods cannot be always applied. This is usually due to incomplete data to design the mathematical model, or having multiple controllable elements in a system. Seeing how these types of problem can be modeled with Game Theory, we propose approaches coming from Reinforcement Learning to solve these sorts of situations in a model-free data-driven methodology. Concretely, we present a study on various Multi-Agent Reinforcement Learning algorithms and apply them on Urban Drainage Systems to test their efficacy. Urban Drainage Systems (UDS) are networks of interconnecting pipes and storage facilities that transport wastewater to Waste Water Treatment Plants, before releasing the water into nearby water bodies, removing them of dangerous chemicals.

The monitoring of biomechanical data during the crutch gait is essential when assessing the progression of the patient during rehabilitation. However, the complexity of this data may hidden embedded patterns on the user gait that could provide additional information about their health condition, and may affect the evolution of the patient. During the last years, research focusing on the implementation of ML techniques for biomechanical analysis have been gaining relevance, due to their capacity for identifying gait disorders, detecting specific gait events or classifying user activity. The main objective of this project, therefore, is to classify 4 different gait patterns through the extraction of relevant spatio-temporal parameters applying ML techniques. To achieve this objective, gait data were collected from a set of instrumented crutches, which allowed to record the trials of 27 healthy participants, along a 10 meter corridor. The collected gait signals were processed and segmented in cycles, by the identification of swing and stance phases using the pitch angle signal of the crutch. The final dataset obtained after this segmentation was composed by a total of 74 extracted features and 6014 samples. The acquisition of this dataset allowed to build a ML classifier to accurately identify the different gait patterns. In order to reduce the data dimensionality, a feature selection process is performed, comparing the most relevant methods applied in literature, and demonstrating the importance of forward acceleration, pitch and force signals for crutch gait identification. The reduced dataset was evaluated against different classifiers through a hyperparameter optimization process, resulting in a SVM classifier able to identify the crutch patterns with a 93.85% of accuracy. Additionally, a solution to reject samples not belonging to one of the expected gait patterns is presented

This project has been developed under the B Modality defined in ETSEIB, as a Graduate project developed while doing an Internship. This thesis is divided in two parts, the theoretical research and the practical implementation. The theoretical research section aims to comprehend and analyze the state-of-the-art regarding the use of RGB and Depth images in instance segmentation, in order to find the models with the highest accuracy. The practical implementation will take place once a candidate algorithm is chosen. The selected algorithm will be tested and further analyzed, and if needed, it will be rewritten to maximize efficiency and to respect the company's code structure. Finally, the best architecture found during the broad research will be tested on real images and compared with the current software solution used by the company, to see whether there is a considerable improvement or not. More information about this project can be found on the preface 1 and an introduction to the company and the main objectives can be found on 2.

This project is focused on improving a Multi Camera People Tracking System that responds to the need to track doctors in a operation room to use their coordinates with a radiation model and compute the uptake they are receiving. Some surgical operations require the use of radiation to be performed such as X-Ray. The radia- tion exposure is defined and controlled by the machine specifications for the patients. However, the clinical staff, for reflection, receives radiation and performs everyday operations so that there is a need to know how much exposed they are. The project PODIUM (Personal Online DosImetry Using computational Methods) was created to open a new way to control the radiation without the use of dosimeters and the Multi Camera People Tracking System is the part responsible of capturing the position of the people in the room. This thesis aims to improve the Multi Camera system by increasing the recording frame ratio, add more flexibility and robustness in the calibration process between cameras, introduce a new pattern format for the world calibration and present a new graphical option for plotting the bodies. The first objective was accomplished by reducing the time delay for the communication client- server. Some problems were encountered due to the printing strings and processing data while working in a higher frequencies so some were deactivated or reduce and the code cleaned and better enclosed. Furthermore, the study showed that the computer specifications were not de- terminant but the number of Kinect cameras was. The second objective consisted in changing the method of calibration from capturing five frames to using a video. The challenges faced were related to filtering and proving that the resulting calibration was correct. To do so, some filters were added and a new test procedure to ensure the calibration quality was created. The validation of the new method has been studied by taking the error mean of the test for different number of frames used and also comparing the transformation matrices calculated with another using a different method. The third objective of adding a new pattern was successfully accomplished for ArUco. This new method for world calibration allowed to work with longer distances because the relation size of 3:1 using QR or ArUco for the same distance was proven. The fourth and last objective was introducing a new online display for the fused information coming from the Kinects. However, the new graphical interface was used and worked, it could not be implemented online. The main difficulty was the fact that the programming languages used were different, from C# to Python, then a required new socket was needed and it was out of the scope of the project. This study added more utility and reliability to the original project and opened a new window to improvements specially related to the graphical interface but also for other works that might use multi-Kinect systems and 3-D positioning.

This project is focused on improving a Multi Camera People Tracking System that responds to the need to track doctors in a operation room to use their coordinates with a radiation model and compute the uptake they are receiving. Some surgical operations require the use of radiation to be performed such as X-Ray. The radiation exposure is defined and controlled by the machine specifications for the patients. However, the clinical staff, for reflection, receives radiation and performs everyday operations so that there is a need to know how much exposed they are. The project PODIUM (Personal Online DosImetry Using computational Methods) was created to open a new way to control the radiation without the use of dosimeters and the Multi Camera People Tracking System is the part responsible of capturing the position of the people in the room. This thesis aims to improve the Multi Camera system by increasing the recording frame ratio, add more flexibility and robustness in the calibration process between cameras, introduce a new pattern format for the world calibration and present a new graphical option for plotting the bodies. The first objective was accomplished by reducing the time delay for the communication client-server. Some problems were encountered due to the printing strings and processing data while working in a higher frequencies so some were deactivated or reduce and the code cleaned and better enclosed. Furthermore, the study showed that the computer specifications were not determinant but the number of Kinect cameras was. The second objective consisted in changing the method of calibration from capturing five frames to using a video. The challenges faced were related to filtering and proving that the resulting calibration was correct. To do so, some filters were added and a new test procedure to ensure the calibration quality was created. The validation of the new method has been studied by taking the error mean of the test for different number of frames used and also comparing the transformation matrices calculated with another using a different method. The third objective of adding a new pattern was successfully accomplished for ArUco. This new method for world calibration allowed to work with longer distances because the relation size of 3:1 using QR or ArUco for the same distance was proven. The fourth and last objective was introducing a new online display for the fused information coming from the Kinects. However, the new graphical interface was used and worked, it could not be implemented online. The main difficulty was the fact that the programming languages used were different, from C# to Python, then a required new socket was needed and it was out of the scope of the project. This study added more utility and reliability to the original project and opened a new window to improvements specially related to the graphical interface but also for other works that might use multi-Kinect systems and 3-D positioning

A diari les persones invidents es mouen pels carrers de Barcelona, molts d'aquests en condicions perilloses per qualsevol persona, ja sigui nen o adult. Quan es parla de la seguretat dels ciutadans, és de màxima prioritat oferir un sistema fiable, capaç de garantir tranquil·litat a tothom, sobretot si es tracta de les persones amb discapacitat visual. Per aquest motiu, l'objectiu d'aquest projecte és molt clar: dissenyar un sistema, incorporat o no a algun accessori, capaç de connectar-se a una càmera que capti les imatges dels semàfors de Barcelona. La opció d'estar contínuament connectat a la càmera brinda la possibilitat d'enviar missatges d'avís, acústics, de forma pràcticament instantània, tot incloent la informació de l'estat del semàfor, ja sigui que està verd, vermell, o bé, no n'hi ha cap de semàfor davant. Partint d'una sèrie de variables que donin dimensió a característiques de disseny de l'artefacte en qüestió, i de la definició de les necessitats dels possibles usuaris s'estableixen les funcions a incorporar a aquest dispositiu. D'aquesta forma, sorgeixen tres solucions amb característiques diferents però amb la mateixa finalitat. L'usuari haurà de poder fer-ne ús a través de botons que permetin engegar l'aparell i pujar o baixar el volum dels missatges. Amb tot això, el dispositiu comptarà amb un baix cost econòmic i energètic, fet indispensable per resultar accessible a tots els potencials clients. Després de comparar analíticament les diferents alternatives, es proposa com a solució definitiva un dispositiu dividit en dues peces, d'una banda la càmera que es pugui acoblar a unes ulleres de sol, i de l'altra el microordinador que processa les dades juntament amb la resta de components. Aquesta solució és molt completa ja que compleix amb els criteris bàsics definits: baixa dificultat de muntatge, ergonomia, flexibilitat, fiabilitat, autonomia i baix cost econòmic. A l'últim capítol, es desenvolupa aquesta idea de manera precisa i detallada per tal de poder fer funcionar un prototip provisional i realitzar una demostració de la funcionalitat de l'aparell en un entorn proper al real.

Robotics is changing the world in a fast manner, simplifying complex and tedious tasks not only in the industrial processes and military fields but also in a social scope and in other areas as, for instance, the farming field that, until now, have been using rudimentary technologies and techniques for conducting its purposes. Earth Rover Europe S.L. is a company within the technological agriculture sector which is developing novel solutions with the aim of helping to feed a growing global population by improving yields, reducing on farm crop waste and at the same time reducing CO2 emissions. Among these contributions is the development of unmanned ground vehicles which will work autonomously inside a farm, carrying out a set of activities previously performed by humans in a less efficient and ecological manner. At this point is where the project presented in the current document takes place. This thesis, which has been conducted inside the previously mentioned company, will present the fundamentals of the development of the central system that will be in charge of the supervision of the entire farm and its elements as well as of the orchestration of the fleet of robots that will be working in it. In order to achieve this, a meticulous description of the entire system, its components and the relations between them must be done before getting on with its implementation. This central system includes a global planning algorithm which optimizes the routes taken by the different robots within the farm to accomplish with the assigned missions. Furthermore, this system must be integrated within an operational environment, therefore, this document will also present the developed simulation for testing purposes and also the real implementation of one rover for doing some real experiments.

During the development of the project, new extensions of the BRIEF descriptor are defined based in the use of color information in different color spaces. They are evaluated against each other and the original BRIEF descriptor (that uses intensity pixel values in grayscale images) by means of a matching test of different points of two different images of the same scene. The selected extension is used to recognize objects in a real time robotic application by means of a classification method. Therefore, the target is to improve the BRIEF descriptor to be applicable in color images by defining the different color extensions and making a comparative evaluation to ensure selecting the best one that most improve the basic BRIEF descriptor with the use of color information. The speed and the recognition rate are computed for every extension and the basic BRIEF to compare the performance of each descriptor. Ones an extension is selected, it is used the Bag Of Features model to create an algorithm able to recognize objects. The BOF model detects the keypoints with the FAST algorithm, describes them by means of the selected extension of the BRIEF descriptor and classifies them using a linear Support Vector Machine. Finally, the algorithm is tested in different situations and conditions, varying illumination, a little bit rotation and background, and checking performance with occluded objects

Learning by demonstration is a machine learning technique which, in the robotics field, can be used for teaching robots new tasks without the need to manually program them. These kind of methods require a model from which the robot acquires the new knowledge. In the present paper, learning by demonstration is applied in a Baxter cobot using human movement as model. To achieve this, a Kinect V2 is used to record the cartesian trajectory of a human hand moving. The framework developed captures the relative coordinates of the right hand with respect to the subject's torso, avoiding the need of a precise camera setup for each recording. After choosing and end-effector orientation, the cartesian data is then processed into a valid 7 DOF trajectory for the Baxter arm, which is used as model trajectory. The model trajectory is fed into a DMP (Dynamic Movement Primitives) algorithm with new starting and final trajectory positions as parameters. Once the DMP trajectories are generated, they are tested both in simulation and in the real Baxter robot. For this, a set of experiments have been designed to prove that the generated DMP trajectories can be used as approaching moves for different pickup operations. The setup used has been a kitchen environment with common objects placed in tables and shelves. The results show that the framework developed is able to create a valid model trajectory for the Baxter robot from human data. The DMP trajectories generated comply with the velocity and arm configuration limits of the robot, are smooth and continuous and fulfill the new requested start and goal positions for movements that do not require major orientation changes.

The current Master's Thesis in Automatics, Control and Robotics covers the development and implementation of an Action Intention Recognition algorithm for proactive human assistance in domestic environments. The proposed solution is based on the use of data provided by a real time RGBD Object Recognition process which captures object state changes inside a defined region of interest of the domestic environment setup. A background analysis is performed to analyze state of the art approaches to both real time RGBD object recognition and action intention recognition methods. The preliminary analysis serves as the base for the proposal of a new volume descriptor for object categorization and an improved formalism for Activation Spreading Networks in the context of action intention recognition. Several tests are performed to study the performance of the proposed solution and its results are analyzed to define the conclusions of the project and propose future work. Finally, the project budget and environmental impact as well as the project schedule are presented and briefly discussed

The AURORA project aims at developing new strategies to take assistive robotics a step further. In order to provide extended services to potential users, engaging a team of simpler robots is often preferable to using a unique, super-capable robot. In such multi-robot systems, the coordination for task execution is one of the major challenges to overcome. The present thesis proposes a solution to the specific issue of multi-robot task allocation within an heterogeneous team of robots, with additional inter-task precedence constraints. The main elements of the state of the art that support this project are reported, including useful taxonomies and existing methods. From the analyzed solutions, the one that better fits with the constraints of the project is an iterated auction-based algorithm able to manage precedence constraints, which has been modified to handle heterogeneity and partial scheduling. The selected solution has been designed and implemented with the particular purpose of being applied to the robotized kitchen setup of the AURORA project; it is however flexible and scalable and can therefore be applied to other use-cases, for instance vehicle-routing problems. Several evaluation scenarios have been tested, that demonstrate the good functioning and characteristics of the system, as well as the possibility to integrate humans into the task assignation process.

This thesis presents how to register point clouds from RGB-D sensors to perform real-time 3D reconstruction of an indoor scene with the aim to build a representation of the environment useful for robot planning. Several common techniques have been explored and compared to determine benefits and drawbacks of each method for our setup. A more specific method was then developed based on this knowledge and has been compared with other methods in terms of rotation and translation errors. Finally these results are used to evaluate the method accuracy and determine in which conditions these results can be replicated

It proposes the creation of an Android application that allows the creation of sociograms from a process involving a group of students. This solution will keep the data of minors securely and keep a record of all the users and processes carried out.

The huge amount of people who suffer from some form of disability implies that more research has to be done in order to develop efficient robotic technologies for them. Yet, research costs can not be neglected. In this matter, software simulators play a significant role as an alternative approach to minimize such expenses. This project aims to make use of OpenSim API and Simbody API, two C++ libraries in order to prove the feasibility of coupling and simulating an exoskeleton model with a musculoskeletal model. Due to the few information related about how to get it, this thesis embraces many others sub-goals strictly related to the achievement of the primary purpose. The first sub-objective has been to learn OpenSim and Simbody API. As solution proposed in this project has been to create a double pendulum model in OpenSim. The second one has been design a model of an exoskeleton. As a result, it has been implemented in OpenSim, and a controller has been designed to test it. The successful achievement of the previous sub-objectives have allowed to create the simulation of the exoskeleton coupled to the musculoskeletal model and simulate it in OpenSim. Last but not least, the documentation throughout the project in conjunction with the successful achievement of the proposed goals has made possible the creation of this thesis which indeed aims to keep working with the exoskeleton coupled to musculoskeletal model or even become a reference in future robotic orthosis projects build in OpenSim and Simbody APIs.

Aquest projecte es basava inicialment en introduir una millora de precisió o velocitat a una impressora 3D de codi obert, durant la fase de desenvolupament tècnic del projecte, es va arribar a la conclusió de la necessitat de refer l'objectiu degut a la complexitat del firmware i les limitacions de càlcul del microcontrolador incorporat a les impressores 3D de baix cost. El nou objectiu consisteix a obtenir una nova plataforma de control, que ofereixi simplicitat en el firmware i augmentar la capacitat de càlcul del processador, mantenint o incrementant en poc el preu del nou sistema. S'estudiaran un seguit d'opcions, però l'escollida consistirà a mantenir la placa Ramps 1.4 (conté l'electrònica de potència i l'adaptació de senyals de les lectures analògiques), substituir la placa de control i fabricar una placa per adaptar els senyals al nou hardware. Amb aquest sistema s'aconsegueix incrementar molt poc el preu de la plataforma, però es millora substancialment la potència de càlcul, el nou sistema introdueix un sistema operatiu (Linux), que possibilita dividir el firmware en processos i així facilita la comprensió i la modularitat del programari. El punt negatiu de la nova plataforma, és que no ofereix totes les funcionalitats que tenia implementades el sistema anterior, i només han estat implementades les funcionalitats essencials per dur a terme una impressió en 3D.

Project oriented to the creation of an object recongition system for a home kitchen using a Kinect 2. The objects are in format of kitchen containers. The system includes the stages of segmentation, feature descriptors and machine learning methods selection, and finally the recognition of the object

The problem of natural objects recognition with different shapes and colors is still an open one, objects can appear in different viewpoints, lighting, rotated, cluttered, occluded,... making their recognition a difficult task. This work is part of an ongoing project to automate a robotic kitchen

In this final degree project, a study is carried out regarding the design, implementation and validation of a real-time system that allows the detection and recognition of traffic signs by means of computer vision algorithms.