CEA LIST

The rapid growth of consumer applications such as video streaming, cloud gaming, Augmented Reality/Virtual Reality (AR/VR), and Internet of Things (IoT) deployments has intensified the demand for intelligent and adaptive congestion control in communication networks. Existing centralized congestion control solutions often struggle with scalability and responsiveness under highly dynamic traffic conditions. In this paper, we present a decentralized traffic engineering framework based on Multi-Agent Reinforcement Learning (MARL), tailored to optimize routing and traffic balancing in real time. The framework introduces two types of agents: (i) Routing Agents, which dynamically adjust link weights to steer flows across alternative paths, and (ii) Balancing Agents, which regulate traffic splitting ratios over precomputed multi-path routes. Beyond agent design, we address a key yet underexplored challenge: agent placement strategy. We compare random deployment versus traffic-aware deployment (based on node degree) and evaluate their effectiveness in mitigating congestion. Using an OMNeT++ simulation environment with realistic consumer traffic patterns on the Abilene topology (a well-known Internet2 research network), we test the framework across diverse congestion scenarios. Results show that traffic-aware placement consistently improves throughput, reduces latency and packet loss compared to both random placement and traditional Equal-Cost Multi-Path (ECMP) routing baselines. These findings highlight the potential of MARL-driven, placement-aware agents to support future consumer network- ing services requiring low-latency, high-reliability, and scalable resource management.

A review of advances in modeling for concrete ultrasonic inspection is proposed. Various well-known (Non-Destructive Evaluation) NDE methods based on different principles are usually used to assess the integrity of concrete structures. Classical ultrasonic techniques usually exploit the coherent transmitted wave propagating through the concrete and can be classified as methods using linear acoustics. The multiple scattering approaches (both analytical and numerical) for modeling this coherent wave have been reviewed, as well as their existing validations. Such validations in two dimensions highlight the need to consider in modeling with a finer attention the correlation between scatterers and the actual dimensionality of the problem. The state of the art related to the influence in modeling of the Interfacial Transition Zone (ITZ) - the thin porous layer surrounding each aggregate - and of the porosity is also reviewed for the case of sound (undamaged) concrete. Finally, different perspectives of modeling strategies are discussed.

This paper presents an efficient approach for determining the joint configuration of a multifingered gripper solely from the point cloud data of its poly-articulated chain, as generated by visual sensors, simulations or even generative neural networks. Well-known inverse kinematics (IK) techniques can provide mathematically exact solutions (when they exist) for joint configuration determination based solely on the fingertip pose, but often require post-hoc decision-making by considering the positions of all intermediate phalanges in the gripper’s fingers, or rely on algorithms to numerically approximate solutions for more complex kinematics. In contrast, our method leverages machine learning to implicitly overcome these challenges. This is achieved through a Conditional Variational Auto-Encoder (CVAE), which takes point cloud data of key structural elements as input and reconstructs the corresponding joint configurations. We validate our approach on the MultiDex grasping dataset using the Allegro Hand, operating within 0.05 milliseconds and achieving accuracy comparable to state-of-the-art methods. This highlights the effectiveness of our pipeline for joint configuration estimation within the broader context of AI-driven techniques for grasp planning.

Refreshable braille displays give visually impaired people access to text and graphics. Existing technologies typically use an actuator for each braille dot, making full-page displays complex, cumbersome and costly, and ultimately limiting their use. This paper presents a novel methodology for actuating a large number of braille pins with a small number of actuators. Each braille pin comprises a bistable mechanism that is actuated through time reversal focusing of elastic waves. The design of such a system is studied in detail, and a mock-up device has been constructed in order to implement the method. The system is able to operate several braille pins one at a time from remote actuators. Although the repeatability of the process still requires improvements, the measurements show a good agreement with theory and allow us to envisage a full-page display based on this method.

2148

Inspired by statistical physics, Ising machines promise to efficiently address complex optimization problems by leveraging topologies that enable local information processing. At the algorithmic level, they emulate Boltzmann probability laws to explore the state space and tend toward optimal configurations. Guided by thermodynamic principles, they minimize collective energy while preserving the entropy-driven ability to move to neighboring states at a given temperature. Among various approaches to implementing Ising machines, analog accelerators built from asynchronous networks of nanodevices or small circuits most closely resemble the microscopic model of directly interacting spins, operating without explicit update instructions or temporary storage of global cost functions or gradients. Magnetic tunnel junctions (MTJs) offer adjustable, tunable stability and rapid switching between binary states, making them promising building blocks for energy-efficient cognitive computing architectures [1]. In this work, we investigate stochastic MTJs (sMTJs) for level-encoded approaches and Spin-Torque Oscillators (STOs) for phase-encoded computing (Fig. 1).

We introduce a new family of abstractions based on a data structure that we call labeled union-find , an extension of the classic efficient union-find data structure where edges carry labels. These labels have a composition operation that obey the group axioms. Like union-find, the labeled version can efficiently compute the transitive closure of a relation, but it is not limited to equivalence relations; it can represent any injective transformation between equivalence classes, which includes two-variables per equality (TVPE) constraints of the form y = a × x + b . Using abstract interpretation theory, we study the properties deriving from the use of abstract relations as labels, and the combination of labeled union-find with other representations of constraints, allowing both improvements in precision and simplification of existing constraints. Due to its efficiency, the labeled union-find abstractions could find many uses; we use it in two use cases, program analysis based on abstract interpretation and constraint solving for SMT, with encouraging preliminary results.

In System of Systems (SoS) applications leveraging 5G communications, maintaining end-to-end (E2E) latency constraints is critical for performance and reliability. However, enforcing strict latency constraints at all times can lead to inefficient resource utilization, network congestion, and increased operational costs. This paper proposes a Dynamic Relaxation Approach using Deep Reinforcement Learning (DRL), where each autonomous subsystem dynamically adjusts its latency relaxation factor based on network conditions, task criticality, and system load. The proposed framework is evaluated under three congestion scenarios using metrics of reward convergence and relaxation state dynamics, demonstrating rapid adaptation and significant improvements in balancing performance with resource consumption. Additionally, despite employing reinforcement learning, our approach remains lightweight due to its binary, threshold-based decision-making process, which determines whether to relax constraints based solely on network parameters.

Health is "a state of complete physical, mental and social well-being, and not merely the absence of disease or infirmity". Health must therefore be considered in its entirety. To begin with, however, it is vital to move swiftly from the concept of holistic health to its practical implementation. This is what the first Digital Health Network symposium proposed on May 30 and 31, 2024. The Digital Health Network was originally set up by a group of players from the healthcare sector, industry, research and elected representatives, with the aim of pooling feedback from experience as close as possible to local realities. The following points were addressed: i) rebalancing the curative approach with a greater emphasis on prevention and health promotion; ii) rebalancing the genetic approach with a greater emphasis on exposomics; iii) insufficient use of digital technology to improve the health of local populations; iv) governance and organizational challenges from the point of view of local decision-makers and observers, with a focus on research; and v) the ambitions of the Digital Health Network. The symposium provided an opportunity to share case studies of public health challenges in local areas, and possible responses in terms of infrastructure and stakeholder participation: citizens, patients, healthcare professionals, service providers, experts and other stakeholders. The challenges of ageing and geriatric frailty underline the importance of a global, preventive and multidisciplinary approach to promote successful, independent ageing.

This paper introduces a sparse beamforming algorithm designed to compensate for guided wave dispersion while addressing phase skewing and energy focusing in highly anisotropic composite plates. Developed for structural health monitoring, this algorithm targets impact-induced defects, such as barely visible impact damages (BVIDs), using a limited number of arbitrarily positioned transducers. Validation is first performed on finite element simulated data with the A0 mode in an eight-ply unidirectional composite with significant anisotropy and a delamination defect, highlighting the critical role of phase skewing compensation. Additionally, energy focusing compensation strategies are introduced, improving image contrast in highly anisotropic materials. Experimental validation is carried out on a real BVID caused by a 6.3 J impact on an aerospace-grade composite plate and compared to a simulation with a similar size delamination. Results show good agreement between experiments and simulations, with significant enhancements in imaging resolution when considering dispersion compensation compared to the commonly used delay-and-sum algorithm. The proposed algorithm improves image dynamics and reduces focal spot area by a factor of more than 100. Finally, the study quantifies errors introduced by neglecting phase skewing in dispersion compensation, demonstrating coherence with observed imaging results for both composite materials studied in this paper.