HAL Portal UTC Université de Technologie de Compiègne
Not a member yet
11652 research outputs found
Sort by
Fault tolerant decentralized collaboration for simultaneous localization and prior map update with stable 2D features
No full textInternational audienceAccurate and reliable localization is crucial for safe navigation of autonomous robots. In environments where GNSS performs poorly, robots can rely on stable 2D features stored on prior vector maps, such as OpenStreetMap, which are detected by on-board sensors. This paper proposes a fault tolerant decentralized collaborative method for online localization and map update. We focus on the case of indirect collaboration, where robots only collaborate through the observation of common landmarks. The method relies on the Schmidt–Kalman filter, which handles correlations between robots while minimizing the number of communications. A Kullback–Leibler Average data fusion is also employed. To ensure the integrity of the state estimation, a fault detection, isolation, and recovery approach is proposed, enabling the detection of sensor faults and the correction of map faults through collaboration. The method is evaluated using both simulation and experimental data. Results show that collaboration improves accuracy and consistency of localization and mapping, and improves the ability to detect faults
S-TCO: A Sustainable Teacher Context Ontology for Educational Support Systems
No full textInternational audienceThe management of teacher support systems faces various challenges as higher education institutions undergo rapid digital transformation. We can summarize these challenges to improve recommendation accuracy, protect user privacy, and reduce environmental impact. Traditional recommender systems (RS) necessitate gathering private information into centralized servers, which increases the risk of data breaches and consumes a lot of energy because of the high data transmission. In order to explore sustainable governance and social responsibility, this paper introduces a semantic sustainable approach, S-TCO (Sustainable Teacher Context Ontology). This approach offers a new theoretical methodology that combines semantic representation and federated learning. This approach transfers the model to data instead of transferring data to the model. We propose a semantic similarity coalition algorithm based on the Teacher Context Ontology (TCO) to address convergence instability and guarantee model aggregation among pedagogically similar teachers. Theoretical simulations demonstrate that S-TCO has the potential to reduce network carbon footprint while respecting privacy restrictions, thus would empower teachers with sustainable governance over their data.</div
Functional random field data analysis using nonparametric L 1 -modal regression
No full textInternational audienc
Acceptabilité social du toucher : un agent virtuel peut-il toucher un humain ?
No full textInternational audienceSocial extended reality (XR) systems increasingly integrate affective and interpersonal cues to support social interaction in immersive environments. Among these cues, social touch is both meaningful and sensitive, raising questions about when and by whom touch is acceptable in virtual reality (VR). This work investigates the acceptability of empathic forearm touch as a function of touch execution and perceived agency of the interaction partner (human vs. autonomous agent). We present a VR interaction framework based on a narrative-driven storytelling task, in which empathic responses and, in some conditions, synchronized vibrotactile feedback, are delivered during predefined emotional moments. To validate the paradigm, we first conducted a real-world pilot study with a trained actor, identifying a narrative scenario that reliably elicited emotional engagement and supportive interpretations of touch. We then outline an ongoing 2×2 VR experiment using a Wizard-of-Oz (WoZ) approach to control partner behavior while manipulating perceived agency and touch presence. This work contributes an evaluation protocol and early insights for designing and assessing socially augmented XR systems involving affective touch.Les systèmes de réalité étendue (XR) intègrent de plus en plus de retours affectifs et interpersonnels pour favoriser l'interaction sociale dans les environnements immersifs. Parmi ces retours, le toucher social est à la fois significatif et sensible, soulevant des questions quant à son acceptabilité en réalité virtuelle (RV) : quand et par qui ? Cette étude porte sur l'acceptabilité du toucher empathique au niveau de l'avant-bras en fonction de son exécution et de la perception de l'agentivité du partenaire d'interaction (soit un humain ou agent autonome). Nous présentons un cadre d'interaction en RV basé sur une tâche narrative, dans lequel des réponses empathiques et, dans certaines conditions, un retour vibrotactile synchronisé, sont délivrés lors de moments émotionnels prédéfinis. Pour valider ce paradigme, nous avons d'abord mené une étude pilote en situation réelle avec un acteur entraîné, identifiant un scénario narratif qui suscitait de manière fiable un engagement émotionnel et des interprétations positives du toucher. Nous décrivons ensuite une expérience en RV 2×2 en cours, utilisant une approche de type « Magicien d'Oz » (WoZ) pour contrôler le comportement du partenaire tout en manipulant la perception de son agentivité et la présence tactile. Ce travail propose un protocole d'évaluation et des premières pistes de réflexion pour la conception et l'évaluation de systèmes XR à réalité augmentée sociale impliquant le toucher affectif
Motion as a language: transformer-based classification of antimicrobial peptide conformational dynamics
No full textInternational audienceAntimicrobial peptides (AMPs) represent a promising alternative to traditional antibiotics against which many bacteria are rapidly gaining resistance. Today, databases containing tens of thousands of AMPs, along with their properties and biological activities, can be screened to select lead candidates for a given application. The conformational plasticity of AMPs has been proven crucial for the recognition of their targets. However, the volume, complexity and recalcitrance to classification of conformational data, obtained from e.g. molecular dynamics (MD) simulations, prevents it from being included in databases, let alone used as a criterion for the screening of AMPs. This work applies the transformer neural network architecture (which powers large language models such as ChatGPT) to the detection of temporal and spatial context in time series of AMP conformations from MD simulations. It shows how the representation of AMP conformational space learnt by the network can be leveraged for the unsupervised classification of AMP plasticity, which can subsequently be used alongside conventional properties for the screening of databases. Thus, it reveals how deep learning can pave the way toward restoring conformational dynamics to its legitimate importance within drug design pipelines
Adaptive Federated Reinforcement Learning for Network Management in Complex Communication Networks
No full textInternational audienceModern consumer communication networks exhibit dynamic conditions, with variable link quality, fluctuating resource demand from video streaming, AR/VR, and IoT devices, and evolving topologies in home and edge networks. In this paper, we propose a federated multi-agent reinforcement learning (F-MARL) framework that leverages adaptive multi-path routing to proactively mitigate congestion and enhance user-perceived quality of service. We model the network as a decentralized, partially observable system, where selected edge or home routers operate as RL agents to optimize traffic-splitting ratios and route selection, while preserving privacy through federated coordination.Our event-driven orchestration triggers federated learning rounds only under congestion, with energy-and utility-aware client selection to ensure lightweight participation on consumergrade devices. Implemented in OMNeT++ with realistic traffic from consumer applications, the framework reduces congestion, improves streaming throughput, and lowers computation overhead compared to OSPF and non-federated MARL baselines. These results highlight the potential of privacy-preserving, adaptive intelligence for next-generation consumer networking.</div
Mechanical dewatering of wet compacts containing binary system of microporous particles
No full textInternational audienceThis study presents the development of a physical model for the mechanical dewatering of wet compacts composed of a binary mixture of soft microporous particles of two different sizes. The layer of microporous particles is conceptualized as a double-porosity system, comprising extraparticle and two distinct intraparticle spaces, each characterized by unique permeability and compressibility properties. Analytical solutions for pressure distributions and consolidation ratios within each space were derived using Heaviside's operational method with Laplace and Fourier integral transformations. The results highlight the impact of particle size and structure on dewatering efficiency and elucidate the role of internal liquid flow within the compressible wet compact. The model also accounts for the consolidation behaviour of both intra- and extraparticle spaces, enabling more accurate predictions of the mechanical dewatering process
AI-enhanced RUL prediction of PEMFCs under dynamic operating conditions using XGBoost-based HI extraction and hybrid transformer-GRU model
No full textInternational audienceProton Exchange Membrane Fuel Cells (PEMFCs) are critical for zero-emission energy systems, particularly in electro-hydrogen generators (GEH2). Accurate Remaining Useful Life (RUL) prediction is crucial for ensuring operational reliability and enabling predictive maintenance. However, dynamic operating conditions present a significant challenge for existing prognostic approaches, particularly in extracting robust Health Indicators (HIs). Conventional HIs, often based on voltage or power, are highly sensitive to mission profiles and fail to generalize in real-world conditions. To address this limitation, we propose a novel data-driven approach based on XGBoost regression to extract a degradation-specific HI directly from raw voltage measurements. This method effectively filters out transient fluctuations caused by varying power demands, isolating the true degradation trend without requiring complex preprocessing or domain expertise. Leveraging the extracted HI, we introduce a hybrid deep learning model that combines Transformer networks and Gated Recurrent Units (GRUs) to capture temporal dependencies and provide accurate RUL predictions under dynamic conditions. Explainable AI techniques are integrated to interpret the model’s predictions and analyze the influence of operational variables on fuel cell degradation. The proposed framework is validated on a real-world industrial dataset from four PEMFC stacks operating in GEH2 systems. Experimental results demonstrate superior accuracy, robustness, and generalizability compared to state-of-the-art methods, highlighting the potential of this scalable and interpretable approach for predictive maintenance in complex industrial environments
Amidinium salt bridges for real-world applications of molecularly imprinted polymers
No full textInternational audienceGünter Wulff and Klaus Mosbach pioneered the molecular imprinting technique in the 70s and 80s, respectively, using complementary approaches. An important development in noncovalent molecular imprinting was the use by Wulff of strong salt bridge interactions displayed by amidinium-phosphate and amidinium-phosphonate complexes in catalyticallyactive enzyme-mimicking molecularly imprinted polymers (MIPs). Inspired by their work, others used amidine monomers and their strong pairing with carboxylate groups in the aim of generating molecular imprints in polar organic solvents and in aqueous solutions. The strong directional salt bridge association between amidine and carboxylate template yields MIPs with high affinity and specificity for their target, allowing applications even in challenging complex environments such as human sweat, urine, and cellular and tissue extracts. Herein, we review the use of amidine monomers as well as related guanidine monomers in molecular imprinting, from the time of Wulff to nowadays, highlighting recent advances and applications from the literature, and discuss future directions in this area
Vers une meilleure évaluation de l'aptitude à l’utilisation des logiciels dispositifs médicaux intégrant l'intelligence artificielle
No full textMedical device software solutions integrating artificial intelligence (AI) are intended for healthcare professionals and patients equally. The diversity of users and use applications makes it essential to assess the usability of these devices during their development. However, the regulatory and normative framework remains unclear, particularly for small organisations. This thesis proposes to analyse existing standards and investigate practices in the field in order to develop a methodological tool to guide start-ups in assessing the suitability for use of software incorporating AI.Les solutions de logiciels dispositifs médicaux intégrant l'intelligence artificielle (IA) s'adressent aussi bien aux professionnels de santé et qu'aux patients. La diversité des utilisateurs et des cas d'usage rend essentielle l'évaluation de l'aptitude à l'utilisation lors du développement de ces dispositifs. Or, le cadre réglementaire et normatif demeure encore peu clair, en particulier pour les petites structures. Ce mémoire propose d'analyser les référentiels existants et d'enquêter sur les pratiques de terrain, afin d'élaborer un outil méthodologique permettant de guider les start-up dans l'évaluation de l'aptitude à l'utilisation des logiciels intégrant l'IA