L'objectif de cette thèse est de définir puis d'étudier les performances de méthodes de guidage coopératif de véhicules aériens autonomes. L'intérêt du guidage coopératif est de confier une mission complexe à une flotte, plutôt qu'à un véhicule unique, afin de distribuer la charge de travail et d'améliorer les performances et la fiabilité. Les lois de guidage étudiées sont distribuées sur l'ensemble des véhicules afin d'une part, de répartir la charge de calcul et d'autre part, d'augmenter la fiabilité en éliminant la possibilité de perte de l'organe central de calcul de la commande.La première partie de la thèse porte sur les possibilités offertes par la règle des plus proches voisins. La loi de guidage développée consiste à ce que la commande de chaque véhicule soit élaborée en combinant les états des véhicules voisins. Afin de transmettre des consignes au groupe de véhicules, des objets dénommés agents virtuels sont introduits. Ceux-ci permettent de représenter des obstacles, d'indiquer une direction ou une cible au groupe de véhicules en utilisant des mécanismes déjà présent dans la loi de guidage.La seconde partie de la thèse porte sur les possibilités offertes par la commande prédictive. Ce type de commande consiste à employer un modèle du comportement du système afin de prédire les effets de la commande, et ainsi de déterminer celle qui minimise un critère de coût en respectant les contraintes du système. La loi de guidage développée emploi un critère de coût tenant compte et arbitrant entre les différents aspects de la mission (sécurité, progression de la mission, modération de la commande), et une procédure de recherche de la commande utilisant jeu prédéfinis de commandes candidates afin d'explorer l'espace de commande de manière efficace. Cette procédure, distincte des algorithmes d'optimisation habituels, génère une charge de calcul faible et constante, ne nécessite pas d'étape d'initialisation et est très peu sensible aux minima locaux. / The thesis objective is to define and study the performances of cooperative guidance methods of autonomous aerial vehicles. The interest of cooperative guidance is to entrust a complex mission to a fleet, instead of an isolated vehicle, to distribute the workload and improve performances and reliability. Studied guidance laws are distributed among all vehicles, on one hand to distribute the computation load, and on the other hand to remove the possibility to lose the centralized organ of command computation.The first part deals with the possibilities offered by the nearest neighbour rule. The developed guidance law consists in elaborating the command of each vehicle by combining the states of neighbour vehicles. To transmit instructions to the fleet of vehicles, objects denominated virtual agents are introduced. These allow figuring obstacles, indicating direction or target using existing mechanisms of the guidance law.The second part deals with the possibilities offered by model predictive control. This type of command consists in employing a behavioural model of the system to predict the control effects, and thus finding the one that minimises a cost criterion while respecting system's constraints. The developed guidance law uses a cost criterion that take into account and arbitrate between the several aspects of the mission (safety, mission evolution, control moderation), and a control search procedure based on a predefined set of candidate controls to explore the control space efficiently. This procedure, different from usual optimisation algorithms, generate a low and constant computation load, needs no initialisation step and is little sensitive to local minima.

The focus of this study was deep learning. A small, autonomous robot car was used for obstacle avoidance experiments. The robot car used a camera for taking images of its surroundings. A convolutional neural network used the images for obstacle detection. The available dataset of 31 022 images was trained with the Xception model. We compared two different implementations for making the robot car avoid obstacles. Mapping image classes to steering commands was used as a reference implementation. The main implementation of this study was to separate obstacle detection and steering logic in different modules. The former reached an obstacle avoidance ratio of 80 %, the latter reached 88 %. Different hyperparameters were looked at during training. We found that frozen layers and number of epochs were important to optimize. Weights were loaded from ImageNet before training. Frozen layers decided how many layers that were trainable after that. Training all layers (no frozen layers) was proven to work best. Number of epochs decided how many epochs a model trained. We found that it was important to train between 10-25 epochs. The best model used no frozen layers and trained for 21 epochs. It reached a test accuracy of 85.2 %.

Ce travail traite de l'évitement d'obstacles pour les robots mobiles à roues. D’abord, deux solutions sont proposées dans le cas d’un seul robot autonome. La première est une amélioration de la technique des champs de potentiel afin de contraster l’apparition de minima locaux. Le résultat se base sur l’application de la définition de l’ «Input-to-State Stability» pour des ensembles décomposables. Chaque fois que le robot mobile approche un minimum local l’introduction d’un contrôle dédié lui permet de l’éviter et de terminer la tâche. La deuxième solution se base sur l’utilisation de la technique du «Supervisory Control» qui permet de diviser la tâche principale en deux sous tâches : un algorithme de supervision gère deux signaux de commande, le premier en charge de faire atteindre la destination, le deuxième d’éviter les obstacles. Les deux signaux de commande permettent de compléter la mission en temps fini en assurant la robustesse par rapport aux perturbations représentant certaines dynamiques négligées. Les deux solutions ont été mises en service sur un robot mobile «Turtlebot 2». Pour contrôler une formation de type leader-follower qui puisse éviter collisions et obstacles, une modification de l’algorithme de supervision précédent a été proposée ; elle divise la tâche principale en trois sous-problèmes gérés par trois lois de commande. Le rôle du leader est adapté pour être la référence du groupe avec un rôle actif : ralentir la formation en cas de manœuvre d'évitement pour certains robots. La méthode proposée permet au groupe de se déplacer et à chaque agent d’éviter les obstacles, ou les collisions, de manière décentralisée / This dissertation work addresses the obstacle avoidance for wheeled mobile robots. The supervisory control framework coupled with the output regulation technique allowed to solve the obstacle avoidance problem and to formally prove the existence of an effective solution: two outputs for two objectives, reaching the goal and avoiding the obstacles. To have fast, reliable and robust results the designed control laws are finite-time, a particular class very appropriate to the purpose. The novelty of the approach lies in the easiness of the geometric approach to avoid the obstacle and on the formal proof provided under some assumptions. The solution have been thus extended to control a leader follower formation which, sustained from the previous result, uses two outputs but three controls to nail the problem. The Leader role is redesigned to be the reference of the group and not just the most advanced agent, moreover it has a active role slowing down the formation in case of collision avoidance manoeuvre for some robots. The proposed method, formally proven, makes the group move together and allow each agent to avoid obstacles or collision in a decentralized way. In addition, a further contribution of this dissertation, it is represented by a modification of the well known potential field method to avoid one of the common drawback of the method: the appearance of local minima. Control theory tools helps again to propose a solution that can be formally proven: the application of the definition of Input-to-State Stability (ISS) for decomposable sets allows to treat separate obstacles adding a perturbation which is able to move the trajectory away from a critic point

In this thesis, control and guidance algorithms for unmanned sea surface vehicles are studied. To design control algorithms of different complexity, first a mathematical model for an unmanned sea surface vehicle is derived. The dynamical and kinematical equations for a sea surface vehicle are obtained, and they are adapted to real life conditions with necessary additions and simplifications. The forces and torques effecting on the vehicle are investigated in detail. Control algorithms for under-actuated six degrees-of-freedom model are designed. PID and LQR controllers are implemented to attain desired surge speed and yaw position. The autopilots are designed and their performances are compared. Based on the autopilots, a guidance algorithm is implemented to achieve desired motions of the vehicle. An obstacle avoidance algorithm is proposed for safe motion among the obstacles. A next-point generation algorithm is designed to direct the vehicle to the most appropriate next way-point if the one ahead is missed. The effects of disturbances on the motion of the vehicle are studied thoroughly on simulation results. PID controller for an unmanned sea surface vehicle is implemented on ArduPilot Mega v1.4 cart controlling a Traxxas Spartan model boat. The performance of the controller is validated. Simulations and experimental results are provided.

The advantages of UAVs in the aviation arena have led to extensive research activities on autonomous technology of UAVs to achieve specific mission objectives. This thesis mainly focuses on the development of a mission-based guidance system. Among various missions expected of UAVs for future needs, autonomous formation flight (AFF) and obstacle avoidance within safe operation limits are investigated. In the design of an adaptive guidance system for AFF, the leader information except position is assumed to be unknown to a follower. Thus, the only measured information related to the leader is the line-of-sight range and angle. Adding an adaptive element with neural networks into the guidance system provides a capability to effectively handle leader's velocity changes. Therefore, this method can be applied to the AFF control systems that use passive sensing methods. The simulation and flight test results clearly show that the adaptive guidance control system is a promising solution for autonomous formation flight of UAVs. The successful flight evaluations using the GTMax rotary wing UAV also demonstrate unique maneuvering aspects associated with rotary wing UAVs in formation flight. In the design of an autonomous obstacle avoidance system, an integrated approach is proposed to resolve the conflict between aggressive maneuvering needed for obstacle avoidance and the constrained maneuvering needed for envelope protection. A time-optimal problem with obstacle and envelope constraints is used for an integrated approach for obstacle avoidance and envelope protection. The Nonlinear trajectory generator (NTG) is used as a real-time optimization solver. The computational complexity arising from the obstacle constraints is reduced by converting the obstacle constraints into a safe waypoint constraint along with an implicit requirement that the horizontal velocity during the avoidance maneuver must be non-negative. The issue of when to initiate a time-optimal avoidance maneuver is addressed by including a requirement that the vehicle must maintain its original flight path to the maximum extent possible. The simulation results using a rotary wing UAV demonstrate the feasibility of the proposed approach for obstacle avoidance with envelope protection.

<p>This thesis consists of four independent papers concerningthe control of mobile robots in the context of obstacleavoidance and formation keeping.</p><p>The first paper describes a new theoreticallyv erifiableapproach to obstacle avoidance. It merges the ideas of twoprevious methods, with complementaryprop erties, byusing acombined control Lyapunov function (CLF) and model predictivecontrol (MPC) framework.</p><p>The second paper investigates the problem of moving a fixedformation of vehicles through a partiallykno wn environmentwith obstacles. Using an input to state (ISS) formulation theconcept of configuration space obstacles is generalized toleader follower formations. This generalization then makes itpossible to convert the problem into a standard single vehicleobstacle avoidance problem, such as the one considered in thefirst paper. The properties of goal convergence and safetyth uscarries over to the formation obstacle avoidance case.</p><p>In the third paper, coordination along trajectories of anonhomogenuos set of vehicles is considered. Byusing a controlLyapunov function approach, properties such as boundedformation error and finite completion time is shown.</p><p>Finally, the fourth paper applies a generalized version ofthe control in the third paper to translate,rotate and expanda formation. It is furthermore shown how a partial decouplingof formation keeping and formation mission can be achieved. Theapproach is then applied to a scenario of underwater vehiclesclimbing gradients in search for specific thermal/biologicalregions of interest. The sensor data fusion problem fordifferent formation configurations is investigated and anoptimal formation geometryis proposed.</p><p><b>Keywords:</b>Mobile Robots, Robot Control, ObstacleAvoidance, Multirobot System, Formation Control, NavigationFunction, Lyapunov Function, Model Predictive Control, RecedingHorizon Control, Gradient Climbing, Gradient Estimation.</p>

This thesis consists of four independent papers concerningthe control of mobile robots in the context of obstacleavoidance and formation keeping. The first paper describes a new theoreticallyv erifiableapproach to obstacle avoidance. It merges the ideas of twoprevious methods, with complementaryprop erties, byusing acombined control Lyapunov function (CLF) and model predictivecontrol (MPC) framework. The second paper investigates the problem of moving a fixedformation of vehicles through a partiallykno wn environmentwith obstacles. Using an input to state (ISS) formulation theconcept of configuration space obstacles is generalized toleader follower formations. This generalization then makes itpossible to convert the problem into a standard single vehicleobstacle avoidance problem, such as the one considered in thefirst paper. The properties of goal convergence and safetyth uscarries over to the formation obstacle avoidance case. In the third paper, coordination along trajectories of anonhomogenuos set of vehicles is considered. Byusing a controlLyapunov function approach, properties such as boundedformation error and finite completion time is shown. Finally, the fourth paper applies a generalized version ofthe control in the third paper to translate,rotate and expanda formation. It is furthermore shown how a partial decouplingof formation keeping and formation mission can be achieved. Theapproach is then applied to a scenario of underwater vehiclesclimbing gradients in search for specific thermal/biologicalregions of interest. The sensor data fusion problem fordifferent formation configurations is investigated and anoptimal formation geometryis proposed. Keywords:Mobile Robots, Robot Control, ObstacleAvoidance, Multirobot System, Formation Control, NavigationFunction, Lyapunov Function, Model Predictive Control, RecedingHorizon Control, Gradient Climbing, Gradient Estimation. / QC 20111121

Η παρούσα διπλωματική ασχολείται με την πλοήγηση κινούμενου ρομπότ. Δεδομένου ενός χώρου με εμπόδια και στόχο, ασχολείται με την δημιουργία ενός αλγορίθμου για την οδήγηση του ρομπότ διαμέσου του χώρου στο στόχο, αποφεύγοντας τα εμπόδια κατά την κίνηση. Επικεντρώνεται σε δίτροχα ρομπότ και αναλύει βήμα βήμα την διαδικασία εύρεση μονοπατιού, δημιουργία τροχιάς και έλεγχο του ρομπότ. / The present thesis deals with the navigation of moving robots. Granted an area with obstacles and target, it deals with the creation of an algorithm for guiding the robot through space at target, avoiding obstacles during movement. It focuses on two-wheeled robots and analyzes step by step the process of finding a path, creating the trajectory and controlling the robot.

Η διατριβή αυτή εξετάζει την χρήση γενετικών αλγορίθμων (ΓΑ) για την επίλυση του προβλήματος του σχεδιασμού κίνησης ρομποτικών συστημάτων τα οποία εκτελούν εργασίες εφοδιαστικής (όπως εργασίες λήψης και μεταφοράς από σημείο σε σημείο, μετακίνησης υλικών επί συνεχούς διαδρομής, κ.α.) στα πλαίσια λειτουργίας τους εντός ενός ευέλικτου συστήματος παραγωγής (ΕΣΠ). Το πρόβλημα του σχεδιασμού κίνησης (ΠΣΚ) είναι ένα υπολογιστικά άλυτο συνδυαστικό πρόβλημα βελτιστοποίησης (έχει αποδειχτεί PSPACE-hard) το οποίο μπορεί να οριστεί ως εξής: «Πως μπορεί ένα ρομπότ να αποφασίσει ποιες κινήσεις πρέπει να αποδώσει προκειμένου να εκτελέσει με επιτυχία επιθυμητές εργασίες στο περιβάλλον εργασίας του;» Προς τον σκοπό αυτό αναπτύχθηκε ένας αριθμός νέων, πρωτότυπων αλγορίθμων εμπνευσμένων από τη Βιολογία των οποίων η απόδοση μετρήθηκε τόσο μέσω πειραμάτων προσομοιωμένων σε υπολογιστή, όσο και σε πραγματικά ρομποτικά περιβάλλοντα στο εργαστήριο του Τμήματος. Συγκρινόμενοι με τις κλασσικές από τη βιβλιογραφία μεθόδους επίλυσης του ΠΣΚ, οι ΓΑ βρέθηκαν ανώτεροι τόσο από πλευράς ποιότητας των λύσεων που παρήγαγαν, όσο και από πλευράς ταχύτητας σύγκλησης (δηλαδή του χρόνου που χρειάστηκαν για τον εντοπισμό αυτών των λύσεων). Επιπρόσθετα, εξετάστηκαν και αντιμετωπίστηκαν με επιτυχία πολύπλοκα προβλήματα κινηματικής που αναφύονται κατά τον σχεδιασμό κίνησης ρομποτικών βραχιόνων σε ένα ΕΣΠ, όπως: Το αντίστροφο κινηματικό πρόβλημα ρομποτικών βραχιόνων με πλεονάζοντες βαθμούς ελευθερίας, η μεγιστοποίηση της επιδεξιότητας του ρομπότ κατά την εκτέλεση των εργασιών του και η παραγωγή με το άκρο εργασίας του ρομπότ ασφαλών και αξιόπιστων τροχιών επί προκαθορισμένων επιθυμητών διαδρομών. Η επίλυση αυτών των προβλημάτων είναι πολύ σημαντική σε πολλές πραγματικές βιομηχανικές εφαρμογές όπως εργασίες συγκόλλησης, βαψίματος ή επάλειψης με ψεκασμό, λείανσης, κ.α. / The use of genetic algorithms (GAs) for the solution of motion planning of robotic systems which perform logistics operations within a flexible manufacturing system (FMS), as well as, logistics tasks in indoors hazardous environments was investigated. Robot motion planning (RMP) is a PSPACE-hard combinatorial problem loosely stated as: How can a robot decide what motions to perform in order to achieve desired tasks in its environment? A number of new biological-inspired approaches were implemented and evaluated on computer simulated environments, as well as, on real industrial environments. In comparison to existing RMP methods, GAs were found superior in terms of both solutions quality and speed of convergence. Furthermore, focusing on RMP of robot manipulators, the proposed approaches tackled with high success difficult kinematics problems such as: the inverse kinematics for robots with redundant degrees of freedom, the maximization of robot’s manipulability, the path following by the robot’s end-effector on demanded trajectories.

Nous observons ces dernières années un besoin grandissant dans l’industrie pour des robots capables d’interagir et de coopérer dans des environnements confinés. Cependant, aujourd’hui encore, la définition de trajectoires sûres pour les robots industriels doit être faite manuellement par l’utilisateur et le logiciel ne dispose que de peu d’autonomie pour réagir aux modifications de l’environnement. Cette thèse vise à produire une structure logicielle innovante pour gérer l’évitement d’obstacles en temps réel pour des robots manipulateurs évoluant dans des environnements dynamiques. Nous avons développé pour cela un algorithme temps réel de génération de trajectoires qui supprime de façon automatique l’étape fastidieuse de définition d’une trajectoire sûre pour le robot.La valeur ajoutée de cette thèse réside dans le fait que nous intégrons le problème de contrôle optimal dans le concept de hiérarchie de tâches pour résoudre un problème d’optimisation non-linéaire efficacement et en temps réel sur un système embarqué aux ressources limitées. Notre approche utilise une commande prédictive (MPC) qui non seulement améliore la réactivité de notre système mais présente aussi l’avantage de pouvoir produire une bonne approximation linéaire des contraintes d’évitement de collision. La stratégie de contrôle présentée dans cette thèse a été validée à l’aide de plusieurs expérimentations en simulations et sur systèmes réels. Les résultats démontrent l’efficacité, la réactivité et la robustesse de cette nouvelle structure de contrôle lorsqu’elle est utilisée dans des environnements dynamiques. / In the field of industrial robots, there is a growing need for having cooperative robots that interact with each other and share work spaces. Currently, industrial robotic systems still rely on hard coded motions with limited ability to react autonomously to dynamic changes in the environment. This thesis focuses on providing a novel framework to deal with real-time collision avoidance for robots performing tasks in a dynamic environment. We develop a reactive trajectory generation algorithm that reacts in real time, removes the fastidious optimization process which is traditionally executed by hand by handling it automatically, and provides a practical way of generating locally time optimal solutions.The novelty in this thesis is in the way we integrate the proposed time optimality problem in a task priority framework to solve a nonlinear optimization problem efficiently in real time using an embedded system with limited resources. Our approach is applied in a Model Predictive Control (MPC) setting, which not only improves reactivity of the system but presents a possibility to obtain accurate local linear approximations of the collision avoidance constraint. The control strategies presented in this thesis have been validated through various simulations and real-world robot experiments. The results demonstrate the effectiveness of the new control structure and its reactivity and robustness when working in dynamic environments.