Search

International audience; In Task and motion planning (TAMP), symbolicsearch is combined with continuous geometric planning. A taskplanner finds an action sequence while a motion planner checksits feasibility and plans the corresponding sequence of motions.However, due to the high combinatorial complexity of discretesearch, the number of calls to the geometric planner can bevery large. Previous works [1] [2] leverage learning methods toefficiently predict the feasibility of actions, much like humansdo, on tabletop scenarios. This way, the time spent on motionplanning can be greatly reduced. In this work, we generalizethese methods to 3D environments, thus covering the wholeworkspace of the robot. We propose an efficient method for 3Dscene representation, along with a deep neural network capable ofpredicting the probability of feasibility of an action. We developa simple TAMP algorithm that integrates the trained classifier,and demonstrate the performance gain of using our approach onmultiple problem domains. On complex problems, our methodcan reduce the time spent on geometric planning by up to 90%.Index Terms—Task and motion planning, 3D scene represen-tation, Action feasibility prediction, Deep learning

International audience; Reinforcement learning (RL) and trajectory optimization (TO) present strong complementary advantages. On one hand, RL approaches are able to learn global control policies directly from data, but generally require large sample sizes to properly converge towards feasible policies. On the other hand, TO methods are able to exploit gradient-based information extracted from simulators to quickly converge towards a locally optimal control trajectory which is only valid within the vicinity of the solution. Over the past decade, several approaches have aimed to adequately combine the two classes of methods in order to obtain the best of both worlds. Following on from this line of research, we propose several improvements on top of these approaches to learn global control policies quicker, notably by leveraging sensitivity information stemming from TO methods via Sobolev learning, and augmented Lagrangian techniques to enforce the consensus between TO and policy learning. We evaluate the benefits of these improvements on various classical tasks in robotics through comparison with existing approaches in the literature.

This paper proposes a new algorithm called States Path Finder (SPF) to solve the general prehensile manipulation problem for multiple robots and multiple objects. Using a constraint graph to represent the problem, SPF finds a list of transitions from the initial state to the goal state, computes a corresponding set of waypoints and connects the waypoints to form a continuous, collision-free solution path. We benchmark SPF on several problem instances with varying characteristics. The results suggest SPF's generality and also its higher efficiency in comparison to Manipulation-RRT, an existing RRT-like algorithm also based on the constraint graph.

This paper proposes a novel framework for recognizing industrial actions, in the perspective of human-robot collaboration. Given a one second long measure of the human's motion, the framework can determine his/her action. The originality lies in the use of joint angles, instead of Cartesian coordinates. This design choice makes the framework sensor agnostic and invariant to affine transformations and to anthropometric differences. On AnDy dataset, we outperform the state of art classifier. Furthermore, we show that our framework is effective with limited training data, that it is subject independent, and that it is compatible with robotic realtime constraints. In terms of methodology, the framework is an original synergy of two antithetical schools of thought: modelbased and data-based algorithms. Indeed, it is the cascade of an inverse kinematics estimator compliant with the International Society of Biomechanics recommendations, followed by a deep learning architecture based on Bidirectional Long Short Term Memory. We believe our work may pave the way to successful and fast action recognition with standard depth cameras, embedded on moving collaborative robots.

International audience; In this paper we introduce methods and algorithms that will help us solve connectivity queries of parameterized semi-algebraic sets. Answering these connectivity queries is applied in the design of robotic structures having similar kinematic properties (e.g. topology of the kinematic-singularity-free space). From these algorithms one also obtain solutions to connectivity queries of a specific parameter which is in turn related to kinematic-singularity free path-planning of a specific manipulator belonging to the family of robots with these properties; i.e. we obtain paths joining two given singularity free configurations lying in the same connected component of the singularity-free space. We prove in the paper how one reduces the problems related to connectivity queries of parameterized semi-algebraic sets to closed and bounded semi-algebraic sets. We then design an algorithm using computer-algebra methods for "solving" positive dimensional polynomial system depending on parameters. The meaning of solving here means partitioning the parameter space into semi-algebraic components over which the number of connected components of the semi-algebraic set defined by the input system is invariant. The complexity of this algorithm is singly exponential in the dimension of the ambient space. The algorithm scales enough to analyze automatically the family of UR-series robots. Finally we provide manual analysis of the family of UR-series robots, proving that the number of connected components of the complementary of kinematic singularity set of a generic UR-robot is eight.

International audience; Despite great achievements of reinforcement learning based works, those methods are known for their poor sample eciency. This particular drawback usually means training agents in simulated environments is the only viable option, regarding time constraints. Furthermore, reinforcement learning agents have a strong tendency to overt on their environment, observing a drastic loss of performances at test time. As a result, tying the agent logic to its current body may very well make transfer unefcient. To tackle that issue, we propose the Universal Notice Network (UNN) method to enforce separation of the neural network layers holding information to solve the task from those related to robot properties, hence enabling easier transfer of knowledge between entities. We demonstrate the eciency of this method on a broad panel of applications, we consider dierent kinds of robots, with dierent morphological structures performing kinematic, dynamic single and multi-robot tasks. We prove that our method produces zero shot (without additionnal learning) transfers that may produce better performances than state-of-the art approaches and show that a fast tuning enhances those performances.

In the context of mobile robotics, this work reports research addressing scene understanding using embedded multi-modal perception systems. Scene understanding remains a complex problem, essential for high society impact applications such as Intelligent Transportation Systems, involving in particular ADAS and Autonomous Vehicles. Results and outlook based on the conducted research are detailed and structured following two research perspectives., Dans le domaine de la robotique mobile, mes travaux de recherche portent sur l’analyse de scènes dynamiques par des systèmes de perception multimodale embarqués. Ceci demeure un verrou scientifique complexe, mais essentiel pour des applications avec un fort impact sociétal telles que les systèmes de transport intelligents où se situent les systèmes d’assistance à la conduite et les voitures à conduite automatisée. Une synthèse et des perspectives structurées sur deux axes thématiques sont présentés.

Contrary to artificial systems, biological systems are able to perform navigation tasks over large distances and long duration. This ability is partly based on the robustness of the learning mechanisms involved in their cognition, on their ability to exploit couplings between different sensors and to adapt themselves continuously to changes in the environment. The modeling of their cognition could thus be an important vector of innovation in robotics, bringing to existing models new efficient and original mechanisms. In order to propose new robotic architectures, this thesis explores the development of models inspired by mammalian cognition for the autonomous vehicle. This work starts from the "PerAc navigation model", a neurocybernetic model designed to reproduce the observations made in neurobiology in small environments. The model has demonstrated good navigation capabilities on small mobile robots, notably linked to its specific way of representing the environment. However, it is not immediately suitable for large-scale operation because of its high computational and memory cost. In particular, recent work has shown the need to improve the performance of its system of visual localisation, the main bottleneck of the model. In order to overcome this problem, the first part of the thesis was dedicated to the integration of population coding mechanisms in the visual localisation system of the model. Thus, several models of visual information encoding have been proposed, based on a new method of sparse coding inspired by the functioning of the visual cortex. In particular, an effort was made to adopt a conventional method of model evaluation, in order to obtain a strong measure of their performance comparable with the state of the art. Finally, the results obtained demonstrated the efficiency of the proposed methods, with significant gains in computational cost and memory over the original model. The second part of the thesis was devoted to the development of a complete neurorobotic navigation model for autonomous vehicles. The original model has therefore been taken over in its entirety and significantly modified to suit the operation of an autonomous vehicle. This work led to the construction of the MpNav model, a new end-to-end model of visual navigation, ables to learn a trajectory in a few iterations. Thus, its use on an experimental vehicle of VEDECOM has permitted to perform several navigation trajectories in complete autonomy, with a record of 2km on a looping trajectory., Contrairement aux systèmes artificiels, les systèmes biologiques sont capables d’effectuer des tâches de navigation sur de grandes distances et de longues durées. Cette capacité reposerait en partie sur la robustesse des mécanismes d’apprentissage impliqués dans leur cognition, sur leur faculté à exploiter les couplages entre différents capteurs et à s’adapter en permanence aux changements survenant dans l’environnement. La modélisation de leur cognition pourrait ainsi être un vecteur important d’innovation en robotique, apportant aux modèles existants de nouveaux mécanismes efficace et originaux. Dans le but de proposer de nouvelles architectures de robotique, cette thèse explore le développement de modèles inspirés de la cognition des mammifères pour le véhicule autonome. Ces travaux sont basés sur le modèle de navigation "PerAc", un modèle de neurocybernétique conçu pour reproduire les observations faites en neurobiologie chez le rongeur naviguant dans de petits environnements. Le modèle a ainsi démontré de bonnes capacités de navigation sur de petits robots mobiles, liées notamment à sa façon spécifique de représenter l’environnement. Il n’est cependant pas adapté pour fonctionner sur de grandes échelles de fonctionnement à cause de ses coûts computationnel et mémoire trop élevés. Des travaux ont démontré le besoin d’améliorer les performances de son système de localisation visuelle, principal goulot d’étranglement du modèle. En vue de surmonter ce problème, la première partie de la thèse a été consacrée à l’intégration de mécanismes de codage en population dans le système de localisation visuelle du modèle. Plusieurs modèles d’encodage des informations visuelles ont ainsi été proposés, s’appuyant sur une nouvelle méthode de codage parcimonieux inspirée du fonctionnement du cortex visuel. Un effort a notamment été réalisé pour adopter une méthode conventionnelle d’évaluation des modèles, afin d’obtenir une mesure forte de ses performances comparable avec l’état de l’art. Les résultats obtenus ont ainsi démontré l’efficacité des méthodes proposées, avec des gains importants en coûts computationnel et mémoire sur le modèle d’origine. La deuxième partie de la thèse, dans la continuité des travaux entrepris, a été consacrée au développement d’un modèle complet de navigation neurorobotique pour un véhicule autonome. Le modèle d’origine a ainsi été repris dans son intégralité et a été fortement modifié pour convenir aux particularités, en termes de contrôle, d’un véhicule autonome. Ces travaux ont abouti à la construction du modèle MpNav, un nouveau modèle de navigation visuelle, qui, bien que proche du type "end-to-end" n’est pas basé sur un apprentissage profond et permet d’apprendre une trajectoire de navigation en peu d’itérations. Son utilisation sur un véhicule expérimental de VEDECOM a ainsi permis d’effectuer en totale autonomie plusieurs trajectoires de navigation, avec un record de 2km sans intervention de l’opérateur

La paralysie cérébrale (PC) spastique est une pathologie qui touche près d’un enfant sur 500 à la naissance. Cette pathologie crée de nombreux problèmes moteurs notamment lors de la marche comme des mouvements involontaires due à une hypertonie musculaire appelée aussi spasticité. Chaque paralysie cérébrale a sa propre sévérité mesurée par différentes échelles. Actuellement, très peu d’aides motrices sont disponibles pour les enfants, le marché des assistances robotisées se focalisant principalement sur les adultes (anthropométrie constante et pathologie peu évolutive).L’objectif du projet EXOKID, dont cette thèse est issue, est donc de permettre à davantage d’enfants atteints de paralysie cérébrale d’avoir accès à un système robotisé à anthropométrie adaptée. Ce système doit pouvoir évoluer en fonction de leur croissance et de l’évolution de leur pathologie afin d’améliorer leur marche tant dans un contexte de rééducation au sein de structures adaptées que dans leur mobilité quotidienne.L’enfant, atteint de cette paralysie cérébrale, doit être pris en charge le plus tôt possible. Le but est de permettre à ce que l’enfant apprenne à marcher : en effet, la plasticité cérébrale des enfants est plus importante que celle des adultes et permet potentiellement à l’enfant d’apprendre rapidement, malgré d’éventuelles lésions cérébrales.Deux sœurs jumelles, dont l’une est atteinte de PC spastique, ont été étudiées. Cette gémellité nous a permis d’avoir des anthropométries proches et d’utiliser les données de la jumelle saine comme référence pour améliorer la marche spastique de sa sœur.Cette thèse s’articule ainsi autour de deux grands champs : la biomécanique et la robotique. Une étude de l’impact de paralysie cérébrale spastique sur la marche humaine a été réalisée. Cette étude a été suivie par une modélisation de la marche spastique et de l’interaction entre un enfant atteint de PC spastique et un exosquelette dans le but d’améliorer la cinématique de l’enfant.Cette thèse a abouti à la modélisation personnalisée d’une marche spastique ainsi que d’un premier contrôle de cette marche, par le biais d’une interaction exosquelette-humain, afin de la faire converger vers une marche plus saine., Spastic cerebral palsy is a pathology that affects nearly one child in 500 at birth.This pathology creates numerous motor problems, particularly when walking, suchas involuntary movements due to muscular hypertonia, also called spasticity. Eachcerebral palsy has its own severity measured by different scales. Currently, very fewmotor aids are available for children, as the market for robotic aids focuses mainlyon adults (constant anthropometry and pathology that does not evolve much). Theobjective of the EXOKID project, from which this thesis stems, is therefore to allowmore children with cerebral palsy to have access to a robotic system with adaptedanthropometry. This system must be able to evolve according to their growth andthe evolution of their pathology in order to improve their walking both in the contextof rehabilitation within adapted structures and in their daily mobility. The child,suffering from this cerebral palsy, must be taken in charge as soon as possible. Theaim is to enable the child to learn to walk : indeed, the cerebral plasticity of childrenis more important than that of adults and potentially allows the child to learnquickly, despite possible cerebral lesions.Two twin sisters, one of whom has spastic PC, were studied. This twinhood allowedus to have close anthropometries and to use the data of the healthy twin as areference to improve the spastic gait of her sister. This thesis is thus articulatedaround two main fields : biomechanics and robotics. A study of the impact of spasticcerebral palsy on human walking was carried out. This study was followed by amodelling of a spastic walking and the interaction between a child suffering fromspastic cerebral palsy and an exoskeleton with the aim of improving the child’skinematics. This thesis resulted in the personalised modelling of a spastic gait anda first control of this gait, through an exoskeleton-human interaction, in order tomake it converge towards a healthier gait.

The objective of this thesis is to plan the motion of an autonomous vehicle so that it can navigate and interact with other moving obstacles without collision in an urban environment. This involves taking into account constraints that may be specific to the vehicle (dynamic, energetic, linked to perception or forecasting systems), relating to the safety and comfort of users or to the environment close to the vehicle. These last depend on the topology of the road, driving rules and other moving actors in the neigborhood of the ego-vehicle. In the context of this thesis, we focused on: (i) the limited or partial perception of the environment close to the vehicle and in particular the presence of static or mobile obstacles hiding part of its neighborhood ; (ii) uncertainty about the prediction of the motion of other agents; (iii) maneuvers to be carried out in urban scenarios where the lanes can be congested with plenty of other static or moving obstacles; (iv) and the respect of the road signs. All these issues have been addressed through geometric approaches by adapting in different ways the principle of path-velocity decomposition introduced by Kant and Zucker (1986). The CARLA simulator has been adapted to validate the proposed approaches on different scenarios reproducing typical situations of interaction with other vehicles in an urban environment, such as, for example, crossing a crossroads with reduced visibility, overtaking a slow or stopped vehicle in its lane, driving in a section of road regulated by traffic lights, etc., L’objectif de cette thèse est de planifier le mouvement d’un véhicule autonome pour qu’il puisse circuler sans collision dans un environnement urbain en présence d’autres obstacles mobiles. Ceci implique la prise en compte de contraintes propres au véhicule (dynamiques, énergétiques, liées aux systèmes de perception ou de prévision), relatives à la sécurité, au confort des usagers et à l’environnement proche du véhicule. Celles-ci dépendent de la topologie de la route, des règles de conduite et des autres acteurs en mouvement dans le voisinage de l’ego véhicule. Dans le cadre de cette thèse, nous nous sommes intéressés plus particulièrement : (i) à la perception limitée ou partielle de l’environnement proche du véhicule et en particulier à la présence d’obstacles fixes ou mobiles occultant une partie de son voisinage proche ; (ii) à l’incertitude sur la prédiction du mouvement des autres agents ; (iii) aux manœuvres à effectuer dans des scénarios urbains où les voies peuvent être encombrées par de nombreux véhicules arrêtés ou en mouvement ; (iv) et au respect de la signalisation. Toutes ces problématiques ont été traitées par des approches géométriques en adaptant de différentes manières le principe de décomposition chemin-vitesse introduit par Kant et Zucker (1986). Le simulateur CARLA a été adapté pour valider les approches proposées sur différents scénarios reproduisant des situations typiques d’interaction avec d’autres véhicules en milieu urbain, comme, par exemple, le franchissement d’un carrefour avec une visibilité réduite, le dépassement d’un véhicule lent ou arrêté sur sa voie, la conduite sur une portion de route réglementée par des feux de circulation, etc.

International audience; Uncertainty estimation of visual feature is essential for vision-based systems, such as visual navigation. We show that errors inherent to visual tracking, in particular using KLT tracker, can be learned using a probabilistic loss function to estimate the covariance matrix on each tracked feature position. The proposed system is trained and evaluated on synthetic data, as well as on real data, highlighting good results in comparison to the state of the art. The benefits of the tracking uncertainty estimates are illustrated for visual motion estimation.

International audience

Naval mine hunting missions consist in detecting and locating mines. To reduce human risk, this mission can be realized with an autonomous multi-robot system. However, the underwater context imposes extremely strong constraints on communications, which imply major difficulties in establishing cooperation strategies. Moreover, these missions are dynamic by nature because it is necessary to identify the detected suspicious objects. The robots must therefore react to new targets and distribute them within the fleet. Finally, there may be more than one way to achieve an objective, and specific constraints may be imposed by the operator. These elements induce complex dependencies on the tasks to achieve.The distribution of these objectives within the system defines ``Multi-Robot Task Allocation Problems (MRTA)'' to be solved before and during the mission. Given the specificities of the mission, such as communication constraints and complex dependencies that may link tasks, solving an MRTA problem is difficult. The solution of an MRTA problem defines a plan to be executed by the robots. The execution of this plan is, however, not deterministic. Therefore, each robot must supervise the plan execution in order to detect deviations, which must be repaired to not compromise the mission. The implementation of decision, supervision, and repair algorithms that are able to meet these challenges is at the heart of our work. We propose to base these algorithms on methods mixing auction-based allocation and hierarchical planning. Finally, we leverage the rich structures used by the decision algorithm to allow robots to supervise and repair their plans., Une mission de chasse aux mines consiste à détecter et localiser des mines, et peut être réalisée grâce au déploiement d'une flotte de robots sous-marins en coopération. Le contexte sous-marin impose cependant des contraintes sur les communications, impliquant des difficultés à l’établissement des stratégies de coopération. Par ailleurs, cette mission est par nature dynamique puisqu'il faut faire face à des aléas, par exemple la détection d'une nouvelle mine va induire de nouveaux objectifs pour le système multirobots. Enfin, il peut exister plusieurs manières de réaliser un objectif, et des contraintes spécifiques peuvent être imposées par l’opérateur. Ces éléments induisent des dépendances complexes entre les tâches à exécuter.Les répartitions de ces objectifs au sein du système définissent des problèmes d’allocation de tâches multirobots (« Multi-Robot Task Allocation » ou MRTA) devant être résolus avant et pendant la mission. Au regard des spécificités de la mission, telles que les contraintes de communication et les dépendances complexes pouvant lier les tâches, la résolution d’un problème de MRTA est difficile. La solution d’un problème de MRTA définit un plan qui doit être exécuté par les robots. L’exécution de ce plan n’est cependant pas déterministe, chaque robot doit donc superviser l’exécution de son plan afin d’en détecter les divergences. Au même titre que l’intégration de nouveaux objectifs, ces divergences sont des aléas qu’il faut réparer afin de ne pas compromettre la mission. La mise en place d’algorithmes de décision, de supervision, et de réparation, aptes à répondre à ces défis est au coeur de nos travaux. Nous proposons d! e faire reposer ces algorithmes sur des méthodes entremêlant allocation par enchères et planification hiérarchique. Enfin, nous tirons profit des structures riches utilisées par l’algorithme de décision pour permettre aux robots de superviser et réparer leurs plans.

International audience; L’étude et la modélisation des interactions entre piétons et usagers de fauteuil roulant électrique (FRE) est un enjeu majeur pour le développement d’assistance à la conduite de tels fauteuils, notamment pour suivre l’intention de déplacement de l’usager tout en évitant une collision avec un piéton à proximité. Pour mieux comprendre les intentions et les distances de proximité piéton-FRE lors de déplacement en groupe, nous avons demandé à un usager de FRE de rejoindre un groupe de deux piétons, marchant côte à côte. Les marcheurs se déplaçaient d’un point de départ à un point d’arrivée situé en face à 20m. Le participant en FRE arrivait perpendiculairement à la trajectoire des marcheurs. Ils devaient alors tous les trois rejoindre ensemble le point d’arrivée. 9 participants marchant (8 hommes, 1 femme, 27ans) et 3 usagers de FRE (3 hommes, 43 ans) ont participé à cette étude. 2 des usagers de FRE étaient ergothérapeutes, 1 avaient une paralysie cérébrale. Trois sessions incluant 1 usager de FRE et 3 marcheurs ont été réalisées. Chaque paire de marcheurs a réalisé 10 essais, pour un total de 20 essais par marcheur et 30 essais pour l’usager en FRE. L’analyse en cours porte sur les adaptations de mouvement de la personne en FRE et des piétons ainsi que leurs positions relatives. Il s’agit alors d’étudier le comportement d’interaction qui permet aux marcheurs et à l’usager de FRE de former un groupe, les distances de proximité ainsi que de potentielles asymétries marcheur-marcheur et marcheur usager de FRE.

International audience; Contact dynamics relies on the simultaneous satisfaction of constraints at the robot body level and at the contact level. At both levels, various formulations can be chosen that all must lead to the same results, given the same hypothesis, hence the little importance of their details. Yet when using it in an optimal control problem, a particular formulation is often imposed by the task to be performed by the robot. In this paper, we detail the formulation of the contact quantities (force, movement) in an arbitrary frame imposed by the task. In that case, we will show that we are typically not interested in working in the local frame (attached to the robot contact point), nor in the world frame, but in a user-defined frame centered at the contact location with a fixed orientation in the world. The derivations can then be used for 6D, 3D or normal (pure-sliding) contact. We implemented the corresponding derivatives on top of the contact dynamics of Pinocchio in the optimal control solver Crocoddyl. We show that a unique formulation is able to handle several operational orientations by achieving several surfacing tasks in model predictive control with the robot Talos.

In this paper, we present the MESSII Dataset (Manipulator Experimental SyStem Identification and Interaction Dataset). This novel dataset aims at providing numerous sequences of movements of the KUKA iiwa manipulator, in order to evaluate methods for identification and estimation of the robot's dynamic parameters and signals, respectively. Different movements of the manipulator are provided, including trajectories moving one or multiple joints at the same time, trajectories that are specially designed for parameter identification, and trajectories including a payload attached at the end-effector of the manipulator. The information obtained from the propioceptive sensors of position and torque is presented. Furthermore, as physical Human-Robot Interaction (pHRI) is the main application of this robot, sequences with interaction with a person are provided to estimate the forces applied, where a force-torque sensor acts as the ground truth. The dataset can also be used in ROS to evaluate real-time methods as information is presented in rosbags files. Possible applications are given, highlighting the advantage of the dataset to help with state-of-the-art challenges, without the need of the real robot nor doing new and complex experiments. The dataset, as well as more information and tools, are publicly available in https://messii-dataset.enit.fr.

The exploration of unknown environments is an essential application of multi-robot systems, particularly in critical missions, such as hazard detection and search and rescue. These missions share the need to reach full coverage of the explorable space in the shortest time possible. To minimize the completion time, robots in the fleet must be able to reliably exchange information about the environment with one another. One of the main methods to expand coverage is by placing relays. Existing relay-placement algorithms tend to either require prior knowledge of the environment, or they rely on maintaining specific distances between the relays and the rest of the robots. These approaches lack flexibility and adaptability to the environment. This paper introduces the “Connectivity-Aware Relay Algorithm” (CARA), a dynamic context-aware relay-placement algorithm that does not require any prior knowledge of the environment. We compare CARA against a state-of-the-art distance-based relay-placement algorithm. Our results demonstrate that CARA outperformed the state-of-the-art algorithm in terms of the time to completion by a factor of 10 as it placed, on average, half the number of relays.

International audience

International audience; In this paper, we propose a new sampling-based path planning approach, focusing on the challenges linked to autonomous exploration. Our method relies on the definition of a disk graph of free-space bubbles, from which we derive a biased sampling function that expands the graph towards known free space for maximal navigability and frontiers discovery. The proposed method demonstrates an exploratory behavior similar to Rapidly-exploring Random Trees, while retaining the connectivity and flexibility of a graph-based planner. We demonstrate the interest of our method by first comparing its path planning capabilities against state-of-theart approaches, before discussing exploration-specific aspects, namely replanning capabilities and incremental construction of the graph. A simple frontiers-driven exploration controller derived from our planning method is also demonstrated using the Pioneer platform.

To improve human-robot collaborations, more and more researches focus on the study of human behaviour. It is in this context of growing relationships between robotics and biomechanics that theCollaBorative roBot (CoBot) project has emerged. Funded by the French Agence Nationale de la Recherche (ANR), this project (ANR-CoBot) targets a human-humanoid robot collaboration in the context of a load-carriage task. During this collaboration, the robot is aimed to safely and pro-actively assist its human partner to handle a load. The idea of the ANR-CoBot project is, first, to understand the mechanisms at stake during a human-human collaboration and, then, to model and simulate them to finally implement them on a humanoid robot TALOS from PAL Robotics to perform the same task with a human partner.As part of this project, my PhD deals with the transition from the modelling of the human behaviourto the integration of this model in a humanoid robot. The final goal of this thesis is to perform a pro-active human-robot interaction to carry a table. From a biomechanics point of view, my thesis aims to better understand human locomotion in order to enable the robot to anticipate its partner’s behaviour. To achieve this goal, experiments were performed to measure Center of Mass (CoM) human trajectories, first during simple locomotion and then during carriage task. Then, models based on optimal control problems were designed to generate human-like trajectories or even predict them. This prediction may improve the human-robot interactions making the robot acting pro-actively. Then, from a robotics point of view, to achieve a collaborative carriage task, five matters need to be handled:• Localization: the robot needs to locate the table and its human partner in order to interactwith them.• Walk generation: the robot needs to walk toward the table at the beginning of theexperiment and also needs to walk with the table towards an unknown location chosenby its human partner.• Whole-body control: the robot needs to lift the table and handle it while walking.• Balance: the robot should not fall during the experiment.• Safety: the robot should not harm its human partner.Among those matters, my thesis mainly focuses on the walk generation. The other topics are handled by other members of the Gepetto team in LAAS-CNRS. One major contribution of my thesis is to embed the prediction trajectory model into the robot walking pattern generator. This aims to improve the human-robot interaction for two main reasons. First, it may allow the robot to be more reactive by anticipating its partner motions and, then, it may make the interaction less disturbing and more natural for the human as the robot may act in a human-like manner., Ma thèse est financée par l’Agence Nationale de la Recherche (ANR) à travers le projet intitulé ANR-CoBot. Ce projet pluri-disciplinaire a pour but d’étudier les interactions entre 2 individus lors d’un transport de charges collaboratif et de réaliser une collaboration homme-robot pour transporter une charge. Dans le cadre de ce projet, ma thèse s’intéresse spécifiquement à la réalisation d’un transport de table en collaboration entre un humain et un robot de type TALOS produit par l’entreprise PAL Robotics.Lorsqu’on s’intéresse à une telle interaction homme-robot, les deux principaux enjeux à relever sont les suivants : comprendre le comportement de l’humain, d’une part, et contrôler le robot pour qu’il puisse adapter ses actions à celles de son partenaire humain d’autre part. Le premier enjeu est d’ordre biomécanique, tandis que le second est d’ordre robotique. Ma thèse allie donc ces deux disciplines. En biomécanique, mon travail s’est focalisé sur l’étude des trajectoires de Centre de Masse (CdM) d’individus marchant seuls sans contraintes puis marchant en transportant une table en binôme. Le but de cette étude est de modéliser ces trajectoires de marche avec un problème de contrôle optimal afin de pouvoir prédire où va un individu à partir de sa trajectoire passée. Cette information est nécessaire pour réaliser une interaction homme-robot fluide où le robot arrive à anticiper le comportement de son partenaire lors du transport, comme le ferait un être humain. En robotique, cinq aspects sont à maîtriser afin de réaliser un transport de charge :• La localisation: le robot doit pouvoir « voir » la table et son partenaire pour pouvoir interagir avec eux.• La génération de la marche : le robot doit pouvoir marcher jusqu’à l’endroit où l’humain veut amener la table, endroit qui est a priori inconnu au robot.• Le contrôle corps-complet : le robot doit pouvoir soulever la table et marcher en la tenant.• La stabilisation : le robot doit pouvoir réaliser tous ces mouvements sans perdre l’équilibre.• La sécurité : le robot ne doit pas mettre en danger l’intégrité physique de son partenaire humain.Parmi ces différents aspects, ma thèse se focalise sur la génération de la marche tandis que les autres points font l’objet de travail d’autres gens au sein de l’équipe Gepetto du LAAS-CNRS. Lors d’une interaction homme-robot, la génération des trajectoires du CdM du robot et de ses pieds doit être rapide et doit prend en compte les mouvements de l’humain en temps réel. Ainsi, ma principale contribution au générateur de marche du robot, lors de ma thèse, a été de le coupler au modèle de prédiction des trajectoires humaines.

Deep reinforcement learning uses simulators as abstract oracles to interact with the environment. In continuous domains of multi body robotic systems, differentiable simulators have recently been proposed but are yet under utilized, even though we have the knowledge to make them produce richer information. This problem when juxtaposed with the usually high computational cost of exploration-exploitation in high dimensional state space can quickly render reinforcement learning algorithms impractical. In this paper, we propose to combine learning and simulators such that the quality of both increases, while the need to exhaustively search the state space decreases. We propose to learn value function and state, control trajectories through the locally optimal runs of model based trajectory optimizer. The learned value function, along with an estimate of optimal state and control policies, is subsequently used in the trajectory optimizer: the value function estimate serves as a proxy for shortening the preview horizon, while the state and control approximations serve as a guide in policy search for our trajectory optimizer. The proposed approach demonstrates a better symbiotic relation, with super linear convergence, between learning and simulators, that we need for end-to-end learning of complex poly articulated systems.

National audience; Nous proposons un nouveau système de SLAM (Simultaneous Localization And Mapping) qui utilise la segmentation sémantique des images. La segmentation sémantique présente un intérêt car elle contient une information haut niveau pouvant rendre plus précis le SLAM. Notre contribution est double : i) un système de SLAM basé sur ORB-SLAM2 capable de créer une cartographie sémantique. ii) Une modification de l'algorithme d'ajustement de faisceaux pour contraindre la position des points 3D de la cartographie en utilisant des a priori géométriques. Nous évaluons notre approche sur plusieurs jeux de données publics et montrons que notre approche améliore l'estimation de pose de caméra en comparaison de l'état de l'art.

Inverse Optimal Control (IOC) is a problem that exists in many fields. In the context of human motion analysis, many methods for solving this problem have been proposed. This paper presents the Projected Inverse Optimal Control (PIOC) approach which puts forward a simple and complete vision of the problem. PIOC is composed of several independent elements allowing to cover problems of dimensioning, measurement noise and reliability of solution. The first element is the parameterization of the data, allowing to reduce the dimensions of the problem. The second element corresponds to the decoupling of the conditions including the conditions of identifiability of the parameters, singularity of the problem and feasibility of the solution. The last element refers to the choice of solution. We show with this approach that the classical solution methods are in fact projections in the space of trajectories. PIOC also allows us to propose a simple algorithm for the choice of basis, a recurrent problem in IOC

International audience

International audience; L'Étiquetage en Composantes Connexe (ECC) est une opération fondamentale de la vision numérique depuis des décennies. Bien que la majeure partie de la littérature traite des algorithmes en 2D, sur des applications telles que la vidéo-surveillance ou la conduite autonome, il y a de plus en plus besoin d'algorithmes pour traiter des images en 3D, notamment pour des applications médicales. Bien que les algorithmes d'ECC en 2D génèrent un nombre important d'accès mémoire et de comparaisons, le problème empire avec algorithmes en 3D. Il s'agit de la malédiction de la dimension. La mise en place d'un algorithme efficace doit donc résoudre ce problème. Ce papier introduit un algorithme utilisant des segments pour étiqueter des volumes en 3D. Les équivalences entre les étiquettes sont accélérées à l'aide d'une nouvelle stratégie atténuant l'impact de la dimension supérieure. Ce nouvel algorithme surpasse les algorithmes de l'État-de-l'art d'un facteur allant de ×1.5 à ×3.1 sur des images médicales et sur des images aléatoires.

Classical visual simultaneous localization and mapping (SLAM) algorithms usually assume the environment to be rigid. This assumption limits the applicability of those algorithms as they are unable to accurately estimate the camera poses and world structure in real life scenes containing moving objects (e.g. cars, bikes, pedestrians, etc.). To tackle this issue, we propose TwistSLAM: a semantic, dynamic and stereo SLAM system that can track dynamic objects in the environment. Our algorithm creates clusters of points according to their semantic class. Thanks to the definition of inter-cluster constraints modeled by mechanical joints (function of the semantic class), a novel constrained bundle adjustment is then able to jointly estimate both poses and velocities of moving objects along with the classical world structure and camera trajectory. We evaluate our approach on several sequences from the public KITTI dataset and demonstrate quantitatively that it improves camera and object tracking compared to state-of-the-art approaches., Comment: This work has been accepted at IEEE Robotics and Automation Letters

This paper presents an approach to handle a minehunting mission with a decentralized multirobot system. To make efficient decisions, the described framework merges auctions to allocate tasks with hierarchical planning. It is responsive and robust to the occurrence of events during the execution phase, thanks to supervision and plan repair processes., Cet article présente une approche pour accomplir une mission de chasse aux mines avec un système multirobot décentralisé. Afin de prendre efficacement des décisions, l'approche décrite mêle allocation par enchères et planification hiérarchique. Des processus de supervision et de réparation sont également intégrés afin de rendre l'approche réactive et robuste à l'occurrence d'événements lors de l'exécution.

This work deals with the flight of small quadrotors (essentially the Crazyflie, ~10cm, 30g) for indoor environment such as pipes. Three main questions are dealt with in the thesis : the existence and nature of aerodynamic perturbations caused by the interaction of rotors' air flow with the environement, the design of a flight controller fit for environments like pipes and the adaptation of this controller to be embedded onboard a Crazyflie controller. The static componenent of aerodynamic perturbations is measured and we detail a map of these for a given set of locations in the environement. Theses measures are then used to build a model that is able to predict these perturbations at any point of the environement. We propose a MPC flight controller based on the resolution of the optimal control problem, taking the perturbation model into account. This controller is able to plan trajectories considering the perturbations and to reject these perturbations. Solving the optimal control problem being too computationally expensive to be done in real time onboard the Crazyflie, we propose to learn a neural network approximating the MPC flight controller in a supervised way, by imitation. This neural network approximation is able to be executed in real time onboard a Crazyflie., Cette thèse traite du vol de petits quadrotors (en particulier du modèle Crazyflie, ~ 10cm, 30g) dans des environnements confinés comme les tuyaux. Trois problématiques sont principalement abordées dans cette thèse : la présence et la nature de perturbations dues à l'interaction entre les flux d'air déplacés par les rotors et les parois de l'environnement, la conception d'un contrôleur pour le vol dans un tel environnement malgré les perturbations et l'adaptation de ce contrôleur aux contraintes de fonctionnement à bord d'un petit quadrotor comme le Crazyflie. Les perturbations aérodynamiques sont mesurées en régime statique et nous présentons une cartographie de celles-ci en un ensemble de points donné de l'environnement. Ces mesures sont utilisées pour dériver un modèle capable de réaliser des prédictions de ces perturbations en tout point de l'environnement. Le contrôleur proposé pour voler dans un tel environnement est un contrôleur MPC basé sur la résolution de problèmes de contrôle optimal intégrant le modèle des perturbations, permettant ainsi à la fois la planification de la navigation en tenant compte des perturbations à venir et le rejet de celles qui seraient hors modèle. La résolution de ces problèmes de contrôle optimal étant trop coûteuse en calcul pour être réalisée en temps réel à bord d'un petit quadrotor comme le Crazyflie, cette thèse aborde ensuite l'apprentissage par imitation du contrôleur MPC par un réseau de neurones qui permet d'obtenir une approximation de ce contrôleur dont le coût en calcul est compatible avec l'utilisation à bord d'un Crazyflie.

International audience; To control a distributed swarm of UAVs, many solutions aim at reaching global coordination through local interactions at the agent scale. These interactions are based on the use of reactive algorithms that allow a fleet to adapt its collective behavior to specific operating contexts, but not to all. While providing some flexibility, reactive algorithms can lack speed of completion, safety, and robustness in clustered environments. Generally, the best known reactive algorithms are borrowed from behavioral patterns observed in nature, but tend to digress from the adaptation to a rigorous geometrical approach. Our approach aims at implementing a coordination model directly inspired from the interactions and behaviors that a small fish, the rummy-nose tetra (Hemigrammus rhodostomus), adopts to display collective movements. This article focuses on such an implementation in the deployment of a group of UAVs inside a circular arena. After carrying out experiments with a swarm of 5 drones, we develop a simulator to test the implementation in a swarm of drones of an extended version of the fish school model and evaluate its capacity to reproduce the experiments. Finally, after validating the UAV model by comparison with the experiments, the use of such collective motion algorithm in several contexts of operation is discussed.

The work presented in this paper tackles the question of braking efficiently with robots. This is an important feature for collaborative robots. These systems need to ensure safety of their surrounding operators without compromising performance. The proposed approach leverages existing motion planning methods based on trapezoidal acceleration profiles which provide a sub-optimal control policy for braking, assuming minorant constant joint jerk and acceleration capacities. Based on this initial solution to braking, a control policy is learnt by reinforcement using the Proximal Policy Optimization deep learning algorithm. This learnt policy assumes constant torque and torque derivative actuation capabilities and predicts an optimal variation of the sub-optimal control policy which better exploits the actuation capabilities of the robot. The principles of this learning approach are first presented and then evaluated in a simulation with an academic example of a Panda robot restricted to 2 degrees of freedom. The obtained results are promising as they lead to a reduction of the stopping time and energy required to brake.

How do children acquire language through unsupervised or noisy supervision? How do their brain process language? We take this perspective to machine learning and robotics, where part of the problem is understanding how language models can perform grounded language acquisition through noisy supervision and discussing how they can account for brain learning dynamics. Most prior works have tracked the co-occurrence between single words and referents to model how infants learn wordreferent mappings. This paper studies cross-situational learning (CSL) with full sentences: we want to understand brain mechanisms that enable children to learn mappings between words and their meanings from full sentences in early language learning. We investigate the CSL task on a few training examples with two sequence-based models: (i) Echo State Networks (ESN) and (ii) Long-Short Term Memory Networks (LSTM). Most importantly, we explore several word representations including One-Hot, GloVe, pretrained BERT, and fine-tuned BERT representations (last layer token representations) to perform the CSL task. We apply our approach to three diverse datasets (two grounded language datasets and a robotic dataset) and observe that (1) One-Hot, GloVe, and pretrained BERT representations are less efficient when compared to representations obtained from fine-tuned BERT. (2) ESN online with final learning (FL) yields superior performance over ESN online continual learning (CL), offline learning, and LSTMs, indicating the more biological plausibility of ESNs and the cognitive process of sentence reading. (2) LSTM with fewer hidden units showcases higher performance for small datasets, but LSTM with more hidden units is Cross-Situational Learning needed to perform reasonably well on larger corpora. (4) ESNs demonstrate better generalization than LSTM models for increasingly large vocabularies. Overall, these models are able to learn from scratch to link complex relations between words and their corresponding meaning concepts, handling polysemous and synonymous words. Moreover, we argue that such models can extend to help current human-robot interaction studies on language grounding and better understand children's developmental language acquisition. We make the code publicly available * .

Adherence to exercise programs for chronic low back pain (CLBP) is a major issue. The R-COOL feasibility study evaluated humanoid robot supervision of exercise for CLBP. Aims are as follows: (1) compare stretching sessions between the robot and a physiotherapist (control), (2) compare clinical outcomes between groups, and (3) evaluate participant perceptions of usability and satisfaction and therapist acceptability of the robot system. Prospective, randomized, controlled, single-blind, 2-centre study comparing a 3-week (3 hours/day, 5 days/week) physical activity program. Stretching sessions (30 minutes/day) were supervised by a physiotherapist (control) or the robot. Primary outcome: daily physical activity time (adherence). Secondary outcomes: lumbar pain, disability and fear and beliefs, participant perception of usability (system usability scale) and satisfaction, and physiotherapist acceptability (technology acceptance model). Clinical outcomes were compared between groups with a Student t -test and perceptions with a Wilcoxon test. Data from 27 participants were analysed ( n = 15 control and n = 12 robot group). Daily physical activity time did not differ between groups, but adherence declined (number of movements performed with the robot decreased from 82% in the first week to 72% in the second and 47% in the third). None of the clinical outcomes differed between groups. The median system usability scale score was lower in the robot group: 58 (IQR 11.8) points vs. 87 (IQR 9.4) in the control group at 3 weeks ( p < 0.001 ). Median physiotherapist rating of the technology acceptance model was

Combined vision/force arrangements are becoming more and more common in small and medium sized businesses as ''Cobots'' becomes more affordable. Cameras can provide a detailed description of the scene, whereas force sensing can provide local but highly precise information about the contact. However, due to the very different nature of these sensing modalities , obtaining an effective combined use is not straightforward. In this thesis we deal with the coupling of vision and force sensing leveraging their complementarity and allowing a robot manipulator to actively achieve compliant motion along the visual task directions. We have developed advanced control schemes with estimators that cope with the update rate discrepancy between the sensors. furthermore, we contribute to the dynamic model parameter retrieval for manipulators, which are key ingredients both in the proposed controllers, whose take into account the manipulator’s dynamics, as well as for performing more accurate simulations. The reported results, both in the developed simulator and with real experiments show the effectiveness of the proposed techniques., Dispositions combinées vision/force deviennent de plus en plus courants dans les petites et moyennes entreprises à mesure que les ''Cobots'' deviennent plus abordables. Caméras peuvent fournir une description détaillée de la scène, tandis qu’un capteur de force fournis des informations locales mais très précises sur le contact. Cependant, en raison de la nature très différente de ces modalités de détection, il n’est pas simple d’obtenir une utilisation combinée efficace. Dans cette thèse, nous traitons le couplage de la vision et de la détection de la force en tirant parti de leur complémentarité et en permettant à un robot manipulateur de réaliser activement un mouvement conforme aux directions de la tâche visuelle. Nous avons développé des schémas de contrôle avancés avec des estimateurs qui gèrent la différence de taux de mise à jour entre les capteurs. De plus, nous contribuons à la récupération des paramètres du modèle dynamique des manipulateurs, qui sont des ingrédients clés à la fois dans les contrôleurs proposés, qui prennent en compte la dynamique du manipulateur, ainsi que pour effectuer des simulations plus précises. Les résultats rapportés, tant dans le simulateur développé que dans les expériences réelles, montrent l’efficacité des techniques proposées.

his paper focuses on the development and design of a type-2 fuzzy logic controller (T2-FLC) for the control of a variable-speed wind energy conversion system (WECS). In this context, the maximum power point tracking (MPPT) strategy has been used for extracting maximum available power from the wind system under varying wind conditions. Until now, the conventional type-1 fuzzy controller (T1-FLC) has been widely used in several applications, due to its high performance. However, in many modern applications, it is necessary to use controllers that have the ability to cope with wide amounts of uncertainties. Consequently, Type-2 FLCs will have the potential to overcome the limitations of Type-1 FLCs which can improve performance for systems that need handling high levels of uncertainty. To discover the strengths and weaknesses of the proposed controller over to T1-FLC and proportional-integral PI controller, a comparative study was executed. The results obtained confirm the efficiency and high performance obtained by T2-FLC compared to other types of controllers.

Language involves several hierarchical levels of abstraction. Most models focus on a particular level of abstraction making them unable to model bottom-up and top-down processes. Moreover, we do not know how the brain grounds symbols to perceptions and how these symbols emerge throughout development. Experimental evidence suggests that perception and action shape one-another (e.g. motor areas activated during speech perception) but the precise mechanisms involved in this action-perception shaping at various levels of abstraction are still largely unknown. My previous and current work include the modelling of language comprehension, language acquisition with a robotic perspective, sensorimotor models and extended models of Reservoir Computing to model working memory and hierarchical processing. I propose to create a new generation of neural-based computational models of language processing and production; to use biologically plausible learning mechanisms relying on recurrent neural networks; create novel sensorimotor mechanisms to account for action-perception shaping; build hierarchical models from sensorimotor to sentence level; embody such models in robots.

This paper presents the internal architecture of a Modifiable Off-the-Shelf open-source autopilot. We show starting from a set of functional and hardware requirements why most autopilots use as a core thread a main loop acting as a nonpreemptive static scheduler, reacting to external events, some solicited, some unsolicited (but expected). We explain how the type of bus used to communicate with the sensor impacts the nature of the events received from the sensors (solicited or not). We show that depending on the workload that a main loop iteration has to handle, the execution time of an iteration can be larger than the period, creating potential delays in the attitude correction. Finally, we explore the degrees of freedom that can be used to reduce the impact of these overloads by smoothing the periodic workload.

International audience

Human-aware and so-called social navigation abilities of a robot need to be tested and evaluated in real-life experiments with real humans. Such tests are mandatory to validate a mature system. But they are heavy to set up and can become tiresome when debugging the early stages of a system. Simulation could help solve this issue, but the lack of available intelligent human avatars that can exhibit a rational behavior restricts this to simple scenarios. To address this issue, we propose a system providing first an autonomous intelligent human agent specifically designed to act and interact with a robot navigating in a simulated environment. Then, the system provides a GUI to visualize the paths taken by the agents as well as the produced metrics in order to help evaluate the interaction. Moreover, the system also provides a high-level interface to control the agents and run repeatable scenarios. We show through a set of experiments how the proposed system can help answer the lack of intelligent avatars for tuning and debugging social navigation systems before their final evaluation with real humans.

Efficient encoding of visual information is essential to the success of vision-based navigation tasks in large-scale environments. To do so, we propose in this article the Sparse Max-Pi neural network (SMP), a novel compute-efficient model of visual localization based on sparse and topological encoding of visual information. Inspired by the spatial cognition of mammals, the model uses a "topologic sparse dictionary" to efficiently compress the visual information of a landmark, allowing rich visual information to be represented with very small codes. This descriptor, inspired by the neurons in the primary visual cortex (V1), are learned using sparse coding, homeostasis and self-organising map mechanisms. Evaluated in cross-validation on the Oxford-car dataset, our experimental results show that the SMP model is competitive with the state of the art. It thus provides comparable or better performance than CoHog and NetVlad, two state-of-the-art VPR models.

In this paper we present a geometric global descriptor dedicated for place recognition in autonomous vehicles applications that is basedon stereo camera generated point clouds. We first present the approach used to record the 3D structure of the visible space. Then, we proposea parametric optimization to achieve the best performance by coupling the dedicated sensor (Stereo Camera) and the used algorithm. Finally, wepropose a GPU implementation based on CUDA. Compared to a CPU, processing times on GPU are accelerated 7 times on a Jetson AGX Xavierand 30 times using a GeForce RTX 3080., Dans cet article, nous présentons un descripteur global géométrique dédié à la reconnaissance de lieux basé sur des nuages de points issues d'une une caméra stéréo pour des applications de véhicules autonomes. Nous présentons d'abord l'approche utilisée pour enregistrer la structure 3D de l'espace visible. Ensuite, nous proposons une optimisation paramétrique pour obtenir les meilleures performances en couplant les caractérisitiques intrinsèques du capteur utilisé (caméra stéréo) et l'algorithme utilisé. Enfin, nous proposons une implémentation optimisée GPU basée sur CUDA. Par rapport à un CPU, les temps de traitement sur GPU sont accélérés 7 fois sur une Jetson AGX Xavier et 30 fois avec une GeForce RTX 3080.

International audience Quadratic programming (QP) has become a core modelling component in the modern engineering toolkit. This is particularly true for simulation, planning and control in robotics. Yet, modern numerical solvers have not reached the level of efficiency and reliability required in practical applications where speed, robustness, and accuracy are all necessary. In this work, we introduce a few variations of the well-established augmented Lagrangian method, specifically for solving QPs, which include heuristics for improving practical numerical performances. Those variants are embedded within an open-source software which includes an efficient C++ implementation, a modular API, as well as best-performing heuristics for our test-bed. Relying on this framework, we present a benchmark studying the practical performances of modern optimization solvers for convex QPs on generic and complex problems of the literature as well as on common robotic scenarios. This benchmark notably highlights that this approach outperforms modern solvers in terms of efficiency, accuracy and robustness for small to medium-sized problems, while remaining competitive for higher dimensions.

International audience

International audience

Les drones autonomes sont maintenant largement rependus dans les différentes communautés scientifiques et de plus en plus d’applications voies le jour. L’utilisation de ces drones en flotte homogène reste un défi à part entière, et leur utilisation dans des flottes hétérogènes reste encore un domaine à explorer. L’un des grands challenges est également de proposer des drones aux capacités diverses leur permettant d’évoluer dans différents milieux (air, eau, terre). En augmentant ainsi leurs capacités, les drones ne se retrouvent plus contraints à un seul milieu et des nouvelles applications ainsi que des nouvelles stratégies de déploiements peuvent être envisagées.L’article suivant présente la conception d’un drone aérien possédant certains aspects de reconfiguration lui permettant de flotter et de se déplacer aisément sur l’eau. Un modèle de drone pareil pourrait alors trouver son utilité dans une flotte hétérogène en jouant aussi bien le rôle de drone aérien que de relais de communication entre des drones sous-marins et aériens. Les systèmes de communication étant différents, l’interfaçage entre les deux milieux reste encore un milieu de recherche à approfondir et un tel drone pourrait faciliter la mise en place de missions d’explorations et de collectes de données multi-environnement.L’idée est de se baser sur un modèle de drone multi copter à bras pliable trouvable sur le marché et d’intégrer un système mécatronique supplémentaire apportant alors à la structure une fonctionnalité de reconfiguration. Cette dernière consistera à changer l’orientation des moteurs pour générer une poussée horizontale et donc permettre un déplacement en surface de l’eau. La structure résultante pourra donc à la fois fonctionner comme un UAV (Unmanned Aerial Vehicle) et un USV (Unmanned SurfaceVehicle).Les récents travaux sur les VTOLs (Vertical Take Off Landing) prouvent que des drones aériens reconfigurable sont viables pour des missions avec de fortes contraintes. En se servant de ces travaux, l’article présente la structure d’un concept de drone UAV-USV tout en détaillant la partie mécatronique de l’aspect reconfigurable. Plusieurs systèmes sont envisagés, présentés et testés afin de choisir le plus pertinent pour respecter les différentes contraintes mécaniques résultantes d’une telle reconfiguration.La structure finale proposée comprendra donc un système mécatronique permettant la reconfiguration tout en portant une attention particulière sur le dimensionnement du drone qui respectera un cahier des charges précis permettant une mission envisageable dans le monde réel.