Portail HAL Ensta
Not a member yet
11080 research outputs found
Sort by
Des militaires face à l’innovation
No full textInternational audienceFor thirty-five years, the French Army has allowed its soldiers to design equipment that is useful for their work. This "participatory innovation" was carried out quietly. Since 2018 and the creation of the Defense Innovation Agency (AID), it has been the object of more pronounced institutional support, which reconfigures the innovation work undertaken by the military. We propose to analyze the effects of this process of « agencification » of innovation by and for soldiers, and will report on the tensions that such a movement can generate.Desde hace treinta y cinco años, el Ejército francés permite a sus soldados diseñar equipos útiles para el ejercicio de sus ocupaciones. Esta "innovación participativa" se llevaba a cabo discretamente. Desde 2018 y con la creación de l'Agence d'innovation de défense (AID), se da lugar a un apoyo institucional más notorio, el cual reconfigura el trabajo de innovación emprendido por los militares. Aquí proponemos analizar los efectos de este proceso de agencificación de la innovación por y para los soldados y daremos cuenta de las tensiones que tal movimiento puede generar.Depuis trente-cinq ans, l'Armée française permet à ses soldats de concevoir des matériels utiles à l'exercice de leurs métiers. Cette « innovation participative » se réalisait à bas bruit. Depuis 2018 et la création de l'Agence d'innovation de défense (AID), elle fait l'objet d'un soutien institutionnel plus marqué, lequel reconfigure le travail d'innovation entrepris par les militaires. Nous proposons ici d'analyser les effets de ce processus « d'agencification » de l'innovation par et pour les opérationnels et rendrons compte des tensions qu'un tel mouvement peut générer
Detection of OFDM modulations based on the characterization in the phase diagram domain
No full textInternational audienceSignal modulation identification is of high interest for applications in military communications, but is not limited only to this specific field. Some possible applications are related to spectrum surveillance, electronic warfare, quality services, and cognitive radio. Distinguishing between multi-carrier signals, such as orthogonal frequency division multiplexing (OFDM) signals, and single-carrier signals is very important in several applications. Conventional methods face a stalemate in which the classification accuracy process is limited, and, therefore, new descriptors are needed to complement the existing methods. Another drawback is that some features cannot be extracted using conventional feature extraction techniques in practical OFDM systems. This paper introduces a new signal detection algorithm based on the phase diagram characterization. First, the proposed algorithm is described and implemented for simulated signals in MATLAB. Second, the algorithm performance is verified in an experimental scenario by using long-term evolution OFDM signals over a software-defined radio (SDR) frequency testbed. Our findings suggest that the algorithm provides good detection performance in realistic noisy environments
A Smart Home Digital Twin to Support the Recognition of Activities of Daily Living
No full textInternational audienceOne of the challenges in the field of human activity recognition in smart homes based on IoT sensors is the variability in the recorded data. This variability arises from differences in home configurations, sensor network setups, and the number and habits of inhabitants, resulting in a lack of data that accurately represent the application environment. Although simulators have been proposed in the literature to generate data, they fail to bridge the gap between training and field data or produce diverse datasets. In this article, we propose a solution to address this issue by leveraging the concept of digital twins to reduce the disparity between training and real-world data and generate more varied datasets. We introduce the Virtual Smart Home, a simulator specifically designed for modeling daily life activities in smart homes, which is adapted from the Virtual Home simulator. To assess its realism, we compare a set of activity data recorded in a real-life smart apartment with its replication in the VirtualSmartHome simulator. Additionally, we demonstrate that an activity recognition algorithm trained on the data generated by the VirtualSmartHome simulator can be successfully validated using real-life field data
The Robust Semantic Segmentation UNCV2023 Challenge Results
No full textInternational audienceThis paper outlines the winning solutions employed in addressing the MUAD uncertainty quantification challenge held at ICCV 2023. The challenge was centered around semantic segmentation in urban environments, with a particular focus on natural adversarial scenarios. The report presents the results of 19 submitted entries, with numerous techniques drawing inspiration from cutting-edge uncertainty quantification methodologies presented at prominent conferences in the fields of computer vision and machine learning and journals over the past few years. Within this document, the challenge is introduced, shedding light on its purpose and objectives, which primarily revolved around enhancing the robustness of semantic segmentation in urban scenes under varying natural adversarial conditions. The report then delves into the top-performing solutions. Moreover, the document aims to provide a comprehensive overview of the diverse solutions deployed by all participants. By doing so, it seeks to offer readers a deeper insight into the array of strategies that can be leveraged to effectively handle the inherent uncertainties associated with autonomous driving and semantic segmentation, especailly within urban environments
Agents autotéliques linguistiques ouverts avec apprentissage par renforcement profond et modèles de langage
No full textAI has made immense progress in the past 10 years, brought about by the increasing availability of computation, data, and by the invention of flexible algorithmic paradigms to leverage both: machine learning, neural networks, deep reinforcement learning, and large-scale self-supervised learning. However, current AI systems are still missing one of the fundamental drives of human beings: the drive to invent one’s own new problems and learn from striving to achieve them. This drive for creative curiosity is at the center of childrens’ play, scientists’ inventions, artists’ exploration of new forms, and underlies a large part of humanity’s cultural progress. In this thesis we use the framework of autotelic agents, agents that build a repertoire of skills by setting their own goals and learning to achieve them, to take first steps towards truly open-ended AI systems. We argue for language as a support for creative imagination of goals, and as a way to easily access open-ended behavior, as well as a domain where language goals are easy to ground and where we can stand on the shoulders of pretrained models to build capable agents without starting from scratch. We present several empirical contributions to this effect, among others: studying linguistic autotelic agents in the complex ScienceWorld text-based environment showing the impact of sampling goals of intermediate difficulty and appropriate social feedback; demonstrating how to collect open-ended goal repertoires with language models with LMA3, and in the realm of program synthesis, how to generate an open-ended diversity of programming puzzles with ACES, and finally how to frame autotelic learning as a 2-player game with Codeplay. We end the manuscript with a discussion of our results leading to a roadmap for future advances in linguistic autotelic agents and of how the framework could be pushed to build truly creative, open-ended AI systems.L'IA a fait d'immenses progrès au cours des dix dernières années, grâce à la disponibilité croissante des calculs et des données, et à l'invention de paradigmes algorithmiques flexibles permettant d’en tirer parti : apprentissage automatique, réseaux neuronaux, apprentissage par renforcement profond et apprentissage auto-supervisé à grande échelle. Cependant, les systèmes d'IA actuels ne disposent toujours pas de l'une des motivations fondamentales de l'être humain : la motivation d'inventer ses propres problèmes et d'apprendre en tentant de les résoudre. Cet élan de curiosité créative est au centre des jeux des enfants, des inventions des scientifiques, de l'exploration de formes nouvelles par les artistes, et sous-tend une grande partie du progrès culturel de l'humanité. Dans cette thèse, nous utilisons le cadre des agents autotéliques, des agents qui construisent un répertoire de compétences en fixant leurs propres objectifs et en apprenant à les atteindre, pour faire de premiers pas vers des systèmes d'IA véritablement ouverts. Nous défendons le langage comme support pour l'imagination créative des buts, et comme moyen d'accéder facilement à un comportement ouvert, ainsi qu'un domaine où les objectifs de langage sont faciles à ancrer et où nous pouvons nous appuyer sur des modèles pré-entraînés pour construire des agents capables sans partir de zéro. Nous présentons plusieurs contributions empiriques à cet effet, entre autres : l'étude d'agents linguistiques autotéliques dans l'environnement textuel complexe de ScienceWorld montrant l'impact de l'échantillonnage d'objectifs de difficulté intermédiaire et d'un retour social approprié ; une démonstration de comment utiliser de grands modèles de langage pour l'invention créative de répertoire de buts ouverts avec LMA3; la recherche d’une diversité de puzzles de programmation avec ACES, et une perspective avec l'implémentation du cadre autotélique comme un jeu multi=agent avec Codeplay. Nous terminons le manuscrit par une discussion sur nos résultats qui débouchent sur un plan d’attaque pour de futures avancées dans le domaine des agents linguistiques autotéliques par une discussion sur la façon dont ce cadre pourrait être poussé à construire des systèmes d'IA véritablement créatifs et ouverts
Hydrogen Production by Photocatalysis under Solar Light: Development of Nanocomposites based on Metal Organic Frameworks (MOFs)
No full textInternational audienc
Modeling a Novel Laser-Driven Electron Acceleration Scheme: Particle-In-Cell Simulations at the Exascale
No full textInternational audienc
Modeling a novel laser-driven electron acceleration scheme with Particle-In-Cell simulations on exascale-class supercomputers.
No full textInternational audienc
Étude des mécanismes de vieillissement hydrique à l’échelle locale et de leur influence sur le comportement mécanique d’assemblages collés
No full textStructural bonding is an increasingly used industrial assembly technique. In the naval, or more generally maritime, field, in the presence of seawater, the prediction of the effect of humidity on the mechanical behavior of bonded assemblies is of paramount importance. The aim of this thesis is to study water diffusion in a two-component epoxy adhesive material, and to propose a simplified modeling strategy for the mechanical behavior of such an adhesive in a single-lap joint bonded assembly, considering the effects of water aging. First, experimental analysis of water absorption in the adhesive is performed using gravimetry and X-ray tomography. The presence of pores in the polymer joint allows us to understand the local evolution of diffusion kinetics, leading to the final choice and validation of a diffusion model. Then, based on experimental tests, the long-term behavior of the investigated adhesive is characterized for different amounts of absorbed water. An enriched 1D finite element model is then developed, enabling the multi-physics modeling of a single lap joint robustly and efficiently. Finally, 2D finite element simulations and tests are carried out for validation purposes.Le collage structural est une technique d’assemblage de plus en plus utilisée dans le milieu industriel. Dans le domaine naval, ou plus généralement maritime, en présence d’eau de mer, la prédiction de l’effet de l’humidité sur le comportement mécanique des assemblages collés est d’une importance capitale. L’objectif de cette thèse est d’étudier la diffusion de l’eau dans un matériau adhésif époxy bicomposant et de proposer une stratégie de modélisation simplifiée du comportement mécanique d’un tel adhésif au sein d’un assemblage collé de type joint à simple recouvrement, en prenant en compte les effets du vieillissement hydrique. Tout d’abord, une analyse expérimentale de l’absorption d’eau dans l’adhésif est réalisée par gravimétrie et tomographie aux rayons X. La présence de pores dans le joint polymère permet une compréhension locale de l’évolution de la cinétique de diffusion, ce qui conduit au choix final et à la validation d'un modèle de diffusion. Ensuite, sur la base d’essais expérimentaux, le comportement à long terme de l’adhésif étudié est caractérisé pour différents niveaux d’absorption d’eau. Un modèle élément fini 1D enrichi est ensuite développé, permettant la modélisation multi-physique d’un joint à simple recouvrement, de manière robuste et efficace. Des simulations par éléments finis 2D et des essais sur assemblages sont enfin réalisés à des fins de validation
H-matrix accelerated FEM-BEM coupling for dynamic analysis of naval structures in pulsating potential fluids
No full textThis article addresses one of the components of our ongoing work towards an efficient computational modeling methodology for evaluating all effects on a submerged structure of a remoteunderwater explosion. Following up on a previous study devoted to computing the transient acoustic fields induced by the shock wave initially sent by the blast on a rigid submarine, we focus here on the second stage of the underwater event, namely solving the transient fluid-structure interaction (FSI) between the structure and the incompressible potential flow induced by the delayed, and slower, oscillations of the gas bubble created by the remote blast.The boundary element method (BEM) is the best-suited approach for handling potential flow problems in large fluid domains (idealized as unbounded), whereas the finite element method (FEM) naturally applies to the transient structure analyses. To perform the FEM-BEM coupling we use a sub-cycling approach that alternates fluid and solid analyses with Neumann boundary conditions. The transient nature of the coupled analysis and the recourse to sub-cycling together make the overall procedure rely on a large number of BEM potential flow solutions, while the complexities of the wet surface and of the solid transient response imply a need for large BE models for the flow potential. This combination of reasons mandates accelerating the BE component.Accordingly, our main contribution is to study the feasibility and effectiveness of coupling the Hierarchical-matrix accelerated BEM (H-BEM) and the FEM for the FSI problems of interest. Inparticular, we show that the same integral operators can be used at all time instants in spite of the expected global motion of the submerged structure, a feature that the H-BEM can exploit to full advantage. The proposed original treatment is validated against analytical solutions for the case of a motionless or mobile rigid spherical immersed object, and then tested on a complex configurationrepresentative of target applications