Archive ouverte HAL-LAAS
Not a member yet
12189 research outputs found
Sort by
ETSI TC SmartM2M; Smart Machine-to-Machine Communications; Demonstration of Performance Evaluation and Analysis for oneM2M Planning and Deployment
No full textThe full document is available onhttps://www.etsi.org/deliver/etsi_tr/103800_103899/103842/01.01.01_60/tr_103842v010101p.pdfThe present document describes and present different demonstrators build on top of the simulation tool and the profiler tool developed in ETSI TR 103 841. More precisely, the simulation tool is a OMNeT++ library implementing either the deployment model and the case studies. The profiler is a standalone software to be run together with a real open source oneM2M implementation to measure and record the impact on computer resources like memory or processor. Those real measures are used as input for the simulator. The present document provides the configuration elements of three demonstrators, highlights classical behaviors in IoT system deployment and illustrates features of the simulation tools in terms of KPI analysis
Reducing Stochastic Games to Semidefinite Programming
No full textInternational audienceWe present a polynomial-time reduction from max-average constraints to the feasibility problem for semidefinite programs. This shows that Condon’s simple stochastic games, stochastic mean payoff games, and in particular mean payoff games and parity games can all be reduced to semidefinite programming
Light-based 3D printing and post-treatments of moulds for PDMS soft lithography
No full textInternational audiencePolydimethylsiloxane (PDMS) chips are still the workhorses of academic microfluidics. Their production requires the fabrication of moulds, commonly produced using clean-room technologies. Light-based 3D printing and in particular, vat photopolymerization, material jetting and two-photon polymerization are rising techniques for the fabrication of moulds for PDMS replication, thanks to their accessibility, fast prototyping time, and improving resolution. Here, we are first reviewing the possibility opened by 3D printing for soft lithography, with a focus on mould designs. Then, inhibition of PDMS curing by photosensitive resins will be discussed as the main technical hurdle of 3D printed moulds. Fortunately, mould post-treatments are efficient solutions to eliminate this curing inhibition, which we gathered in a large database of post-treatment protocols from the literature
Balancing Accuracy and Energy: An Empirical Study of Optimal Subset Size Selection
No full textThis is the version i submitted to the conference.International audienceThis paper investigates the trade-o between model performance and energy eciency in supervised learning by varying the size of the training data, motivated by the increasing energy consumption of machine learning models. We establish empirical energy scaling laws for neural network training and develop optimization frameworks for balancing accuracy and energy consumption through stratied random sampling. We demonstrate that energy consumption per training epoch scales linearly with dataset size for CNNs, with architecture-dependent coecients ranging from 2.143×10 -5 kJ/sample (MNIST) to 1.356×10 -4 kJ/sample (CIFAR-10). Critically, we show that model complexity, not dataset characteristics, determines energy scaling patterns. Our comprehensive analysis across MNIST, Fashion-MNIST, and CIFAR-10 reveals optimal subset sizes for energy eciency: 5% for MNIST and Fashion-MNIST and 24% for CIFAR-10 when maximizing accuracy per unit energy consumed. We extend these ndings to federated learning with 20 clients, validating the framework's generalizability across distributed training scenarios. For practitioners targeting specic accuracy thresholds, we provide energy budgeting strategiesfor example, achieving 80% accuracy on MNIST requires only 7.5% of the data while saving 82% energy. The derived energy models enable solving three classes of optimization problems: minimizing energy subject to accuracy constraints, maximizing accuracy within energy budgets, and optimizing eciency metrics. This work provides actionable guidance for sustainable AI development and establishes a methodology for deriving energy scaling laws for new datasets and architectures.</div
Micro separator of aerosol particles based on the thermophoretic effect
No full textInternational audienceThe micro-separator of aerosol particles based on thermophoretic effect is proposed. The first step in designing such a micro-device is the development of the mathematical model. The three-dimensional model is proposed and several key parameters are tested such as the length and width of the microchannel, the intensity of the temperature gradient and the velocity of the carrier gas. The next steps will be the fabrication of a prototype and then testing its efficiency
Informatique Fondamentale et ses Mathématiques : Une photographie en 2025
No full textInternational audienceL'informatique Fondamentale et ses Mathématiques (IFM) regroupe des thématiques relevant a priori de tous les domaines de l’informatique, pourvu que la démarche de recherche ou le point de vue adoptés soient « mathématisables ». Cet ouvrage réunit les cours proposés à l’édition 2025 de l’École des Jeunes Chercheuses et Chercheurs en Informatique Fondamentale et ses Mathématiques (EJCIFM), organisée à Caen du 16 au 20 juin 2025 dans le cadre du Groupe de recherche en Informatique Fondamentale et ses Mathématiques (GdR IFM) du CNRS. Présentant de manière équilibrée différentes thématiques du domaine IFM, cet ouvrage en propose une photographie.Cinq thèmes sont présentés par sept spécialistes du domaine : l’informatique quantique, la sémantique, la combinatoire énumérative, l’évaluationsymbolique numérique et la compilation de connaissances. Ces chapitres ont été écrits par Miriam Backens, Kostia Chardonnet, Pierre Clairambault,Julien Courtiel, Denis Arzelier, Mioara Joldes et Bruno Zanuttini
Étude de cavité laser hybride émettant autour de 965nm pour la génération de peigne de fréquences
No full textInternational audienc
Méthodes d'apprentissage de raisonnement géometrique pour la planification de tâches et de mouvements multi-robots.
No full textNational audienceThis thesis, titled "Learning Geometric Reasoning for Multi-Robot Task and Motion Planning", focuses on improving the efficiency of Task and Motion Planning (TAMP) in robotic manipulation by integrating learned geometric reasoning. Task and Motion Planning is a fundamental challenge in robotics, particularly in scenarios involving object manipulation within complex and cluttered environments. The difficulty of TAMP stems from the need to simultaneously reason about high-level task sequences and low-level motion feasibility. Traditional approaches rely heavily on computationally expensive geometric planners, making them inefficient in real-time applications. The primary objective of this research is to address these limitations by developing learning-based models capable of predicting action feasibility, thereby reducing the need for costly geometric verification.A key contribution of this thesis is the development of the Action and Grasp Feasibility Prediction Network (AGFPNet), a convolutional neural network designed to assess the feasibility of robot actions and grasps in cluttered 3D environments. AGFPNet leverages visual scene representations to predict whether a given action is feasible, significantly reducing the number of geometric planner calls required during task planning. This predictive capability enhances the efficiency of TAMP by prioritizing actions that are more likely to succeed, ultimately leading to improved success rates and reduced planning times. This thesis also proposes extensions of action and grasp feasibility predictions to multi-robot settings and mesh-shaped objects.Building on AGFPNet, the thesis introduces Geometric Reasoning Networks (GRN), a more advanced model that incorporates graph-based representations of 3D environments. Unlike image-based approaches which suffer from occlusions as the number of objects increases, GRN represents the environment using graphs where nodes correspond to objects and edge represent spatial relationships between them. Moreover, in addition to predicting action and grasp feasibility, GRN is also capable of predicting the reasons of infeasibility, namely inverse kinematics feasibility and grasp obstructions by other objects. This offers richer geometric feedback to the planner that could be used to rectify infeasibility, enabling more informed decision-making. In addition to its increased accuracy, this model stands out for its interpretability, scalability, ability to generalize to more complex environments, and adaptability to real-world applications on robots such as the Franka Emika Panda and PR2.Additionally, the thesis presents a novel feasibility-informed task and motion planning algorithm that integrates learned geometric reasoning into the planning process. By using feasibility predictions as heuristics, the planner can efficiently explore the search space, pruning infeasible branches and prioritizing promising actions. This approach is further enhanced by an informed backtracking mechanism, which records reasons for infeasibility as planning constraints, ensuring the avoidance of repeated geometric planning failures as well as a better estimation of the cost-to-goal. This method significantly outperforms previous traditional and learning-based TAMP approaches, achieving a 100% success rate in multiple benchmark scenarios while drastically reducing computation time.In conclusion, this research provides an innovative and scalable approach to robot task and motion planning by leveraging deep learning for geometric reasoning. The proposed models and algorithms demonstrate substantial improvements in efficiency, interpretability, and generalization, making them suitable for real-world applications ranging from industrial automation to domestic robotics. By reducing the reliance on computationally expensive geometric planning and introducing learning-based heuristics, this thesis paves the way for more adaptive and capable autonomous robotic systems.Cette thèse, intitulée "Learning Geometric Reasoning for Multi-Robot Task and Motion Planning", se concentre sur l'amélioration de la planification des tâches et des mouvements (Task and Motion Planning - TAMP) en robotique en utilisant l'apprentissage automatique pour le raisonnement géométrique. La planification des tâches et des mouvements est une problématique clé en robotique, particulièrement pour la manipulation d'objets dans des environnements complexes et encombrés. Cette tâche est difficile en raison de la complexité combinatoire des problèmes de planification et du coût computationnel élevé des planificateurs géométriques. L’objectif principal de cette thèse est de réduire cette complexité et d'améliorer l'efficacité des algorithmes de planification en intégrant un raisonnement géométrique basé sur l’apprentissage automatique.L’une des principales contributions de cette recherche est le développement d'AGFPNet, un réseau de neurones convolutif permettant de prédire la faisabilité d'actions et de prises dans des environnements 3D encombrés, à partir de représentations visuelles de ces derniers. Grâce à cette prédiction, il est possible de réduire le nombre d’appels aux planificateurs géométriques coûteux, en priorisant les actions les plus susceptibles d’être réalisables. Ce modèle améliore ainsi considérablement les performances en terme de taux de succès et d’efficacité computationnelle. Afin d’élargir l’applicabilité de cette approche, la thèse explore également son extension aux problèmes multi-robots ainsi qu’à la manipulation d’objets en maillage 3D.En parallèle, la thèse introduit les Geometric Reasoning Networks (GRN), un modèle avancé basé sur des graphes modélisant les relations géométriques entre les objets dans un environnement. Ce modèle surmonte les limitations des approches basées sur l’image en résolvant les problèmes d’occlusion. GRN offre également des informations plus riches à travers la prédictions des raisons d'infaisabilité d'une action, notamment la prédiction de l'existence de solution de cinématique inverse et l'estimation des obstructions de prises par d'autres objects. Ces informations peuvent ainsi être utilisées afin de rectifier l'infaisabilité et mieux guider le planificateur vers une solution géométriquement faisable. En plus de sa précision accrue, ce modèle se distingue par son interprétabilité, scalabilité, sa capacité de généralisation à des environnements plus complexes et par son adaptabilité aux applications réelles sur des robots tels que Franka Emika Panda et PR2.En outre, cette thèse propose également une méthode novatrice pour intégrer ces prédictions de faisabilité et de raisons d'infaisabilité dans un algorithme de planification des tâches et des mouvements via un backtracking informé. Cette technique améliore considérablement la recherche de solutions en grâce à un pruning plus efficace de l’arbre de recherche et une estimation plus précise du coût pour atteindre l’objectif. Les expérimentations démontrent que cette approche permet de résoudre efficacement des problèmes de planification complexes, atteignant un taux de succès de 100 % sur plusieurs benchmarks. De plus, cette méthode surpasse considérablement plusieurs approaches TAMP précédentes.En résumé, cette thèse propose une approche innovante et performante pour la planification des tâches et des mouvements en robotique. En exploitant l’apprentissage profond pour prédire la faisabilité des actions et optimiser la recherche de solutions, elle ouvre de nouvelles perspectives pour la robotique autonome et collaborative, avec des applications potentielles dans des secteurs variés allant de l’industrie manufacturière à la robotique domestique
Dimensionnement et Commande optimisées d'une motorisation tout électrique pour fonction non-propulsive critique à objectifs de performances disparates sous fortes contraintes
No full textInternational audienceL'article présente une motorisation électrique d'ouverture de porte d'aéronef. Les nombreuses contraintes dimensionnelles, fonctionnelles et sécuritaires ont conduit à une approche globale de dimensionnement de la chaîne cinématique, du moteur et de son architecture de commande. De cette étude, deux prototypes de moteurs synchrones à aimants permanents ont été réalisés. Une commande en cascade de ces moteurs a été mise en oeuvre pour assurer le suivi de trajectoires préétablies par commande optimale. Ces trajectoires prennent en compte les profils de missions exigeants par optimisation sous fortes contraintes. Des premiers résultats expérimentaux en quasi-statique ont permis de valider le dimensionnement des moteurs et sa commande
Apprentissage par renforcement pour le robot H1-2
No full textHumanoid robots are continually advancing in their capabilities over time. These advancements are driven by the development of high-performance hardware, like GPUs (graphics processing units), and the progress in machine learning methods such as Reinforcement Learning (RL). This latter allows robots to learn tasks such as walking, balancing, and interaction with environment by trial and error, but is not enough due to the fact that once the model of the robot is trained, it is needed to transfer it on real robots (sim-to-real transfer), which remains a significant challenge. This is difficult because of model differences, sensor noise, and environmental factors that can lead to failure when these algorithms are deployed.This thesis explores various approaches to addressing these challenges by comparing a traditional reinforcement learning algorithm such as Proximal Policy Optimization (PPO), implemented using the Robotics System Lab RL (RSL RL) library, with a novel constrained reinforcement learning method, Constraints as Terminations (CaT), implemented via the CleanRL library. The primary focus is to adapt the existing PPO implementation, originally developed for the H1-1 robot, to the latest version of the robot, H1-2.The goal of this thesis is to compare different approaches to bridge the gap between simulation and real-world deployment, offering insights into how a constrained RL technique can be adapted to different robotic systems, particularly on humanoid robots