Archive ouverte HAL-LAAS
Not a member yet
12189 research outputs found
Sort by
Ordonnancement en ligne robuste de coflows à partir de prédictions
No full textIn this paper, we introduce an online scheduling framework for minimizing the total weighted average completion time of coflows, in scenarios where both flow sizes and release times are unknown. Instead of relying on exact flow sizes, the framework leverages potentially inaccurate predictions. At each decision point, it selects a subset of coflows by solving a Minimum Unscheduled Weight Problem (MUWP), and then applies a variant of the Sincronia algorithm-adapted to operate directly on the predicted sizes. We prove that the proposed approach achieves provably strong competitive guarantees, and extensive simulations demonstrate that it outperforms baseline methods both on average and in worst-case instances
Réseaux de neurones graphiques explicables pour la détection et le diagnostic d'anomalies dans les séries temporelles multivariées
No full textInternational audienceIndustrial systems consist of multiple interconnected sensors generating multivariate time series (MTS), where anomaly detection is critical for ensuring reliability and safety. While recent Graph Neural Network (GNN)–based approaches effectively detect when anomalies occur, they often fail to explain which sensor measurements are truly symptomatic and why anomalies emerge.We propose an explainability framework that combines attention-based GNN anomaly detection with consistency-based model-based diagnosis (MBD). Under the hypothesis that anomalies arise not from faulty sensors themselves but from altered inter-sensor relationships, we analyze attention changes between a normal model and a fault-adapted model to identify influential relational components responsible for system failures. By integrating attention variation analysis with conflict-set generation and minimal hitting set computation, our method produces ranked candidate diagnoses corresponding to faulty system influences
Addressing Long-Horizon Failure in Language-Grounded Robotics via Structured Interaction
No full textDeploying autonomous robots in human-centric environments requires consistent reasoning, memory, and failure recovery across many sequential actions. Although vision-language action systems have shown impressive low-level generalization, their performance degrades rapidly as action sequences grow longer and constraints evolve. Our insight is that these failures arise from a structural mismatch between training supervision and the states that agents encounter during execution, leading to compounding errors and inconsistent behavior over time. To address this, we introduce MAGMA-GEN, a structured interaction-based data generation framework, where a language model-based agent generates supervision signals directly from its own long-horizon executions. MAGMA-GEN provides structured preference and scoring signals for partial completion, failure, and recovery, enabling continued training without expert demonstrations or trajectory-level human labeling once task components are defined. We further analyze a dual-agent instantiation that separates high-level tool selection from persistent task and safety memories, improving stability over extended action sequences. We evaluate our approach on long-horizon manipulation tasks with evolving constraints, designed to expose long-horizon failure modes. Under a fixed training budget, models trained with our method outperform human-labeled and synthetic-augmentation baselines, while continuing to improve with additional self-generated experience, demonstrating more stable scaling behavior
Postdoc report (PO2.1): Détection dans l'Internet des Objets
No full textLe postdoc 2.1 du PEPR Superviz vise à travailler sur la détection d’intrusion dans le cadre de réseaux composés d’objets connectés. Ce travail comporte deux volets principaux, correspondant à deux tâches T1 et T2.— Tâche T1 : compléter les travaux d’une précédente thèse réalisée au LAAS-CNRS sur la détection d’intrusion embarquée dans des contrôleurs Bluetooth Low Energy (BLE).— Tâche T2 : proposer une approche alternative aux travaux précédents en considérant une implémentation des mécanismes de détection d’intrusion dans une pile BLE en source libre
Fifty years of research on resource-constrained project scheduling explored from different perspectives
No full textInternational audienceThe resource-constrained project scheduling problem is one of the most investigated problems in the project scheduling literature, and has a rich history. This article provides a perspective on this challenging scheduling problem, without having the ambition to provide a complete overview. Instead, the article does aim to summarize a number of reasons why this problem has been so intensely investigated from different perspectives.It will be shown that this scheduling problem has many faces, and therefore deserves a lot of research time from a computational and theoretical point of view as well as from a practical point of view. An overview of possible extensions to other problems and a detailed overview of the used (both heuristic and exact) solution methods will be given. In addition, the data used will be discussed and interesting avenues for further research will be mentioned throughout the different sections.</div
La PLL-Emmental pour l'apprentissage neuro-symbolique efficace de contraintes & fonction objectif
No full textInternational audienceLa PLL-Emmental pour l'apprentissage neuro-symbolique efficace de contraintes & fonction objecti
Détection d'anomalies dans la télémesure satellite basée sur la fonction de Christoffel
No full textInternational audienc
Coupled Local and Global World Models for Efficient First Order RL
No full textRL has demonstrated strong performance in locomotion through robust sim-toreal transfer from parallelized simulators to hardware. In contrast, extending this simulator-centric approach to real-world robotics manipulation is hindered by high data demands, sim-to-real discrepancies in intricate interactions, and the difficulty of engineering rewards or accurate physics for tasks involving deformable or irregular objects. We introduce a framework that replaces simulators with a diffusion-based world model trained on real robot image data, which captures complex dynamics from data without rule-based modeling, thus addressing the limitations that prevent simulator-style RL from succeeding in manipulation. Our method enables feasible policy training despite the high cost of trajectory unrolls in large-scale image models, through a novel decoupled first-order gradient (FoG) approach: the full world model generates accurate forward trajectories, while a lightweight latent-space surrogate model is learned to approximate local dynamics around the policy's trajectories, providing efficient, low-variance gradients via backpropagation. This surrogate operates on compressed representations, avoiding direct differentiation through pixel-level diffusion processes. Unlike previous model-based RL approaches, this decoupling ensures high-fidelity forward unrolling alongside computationally tractable backward differentiation. Evaluated on a real robotic arm, our method achieves high success rate on the Push-T task, strongly outperforming PPO in sample-and time-efficiency while offering a scalable alternative to simulator-dependent RL for real-world manipulation
A Gaussian surrogate of partially observed stochastic processes using Wasserstein metric
No full textInternational audienceApproximating the evolution of probability measures for nonlinear stochastic differential equations (SDEs) and the associated nonlinear filtering problems is a challenging problem, as it involves solving high-dimensional differential equations. In contrast to classical variational inference methods, which address this challenge by minimizing the Kullback-Leibler (KL) divergence between the true and approximate distributions, we propose a Wasserstein-based variational framework for approximating the laws of stochastic systems. In particular, instead of minimizing the KL divergence, our approach minimizes the Wasserstein-2 () distance between the joint probability distributions of the state and observation processes. This formulation respects the underlying transport geometry and results in evolution equations for Gaussian parameters that provide an approximation of the dynamics of the true measure. An illustration is provided for some of our results with the help of an academic example
