Experimental fusion research facilities, such as the International Fusion Materials Irradiation Facility-DEMO Oriented Neutron Source (IFMIF-DONES), require advanced remote handling (RH) systems to perform maintenance and inspection tasks in a safe and reliable manner, due to their intrinsic high-radiation nature. The mixed-criticality requirements of the data streams used in these systems force the deployment of separate networks and communication technologies. Commonly, it includes fieldbuses for traffic control, standard Ethernet for video and general-purpose traffic, and dedicated networks for the most critical safety-related signals. This fragmentation leads to complex and costly deployments and also prevents the application of models for predictive maintenance or advanced monitoring. The time-sensitive networking (TSN) technology stack aims to provide deterministic behaviour for data transmission over standard Ethernet, allowing for convergence on a single network and ensuring bounded latencies for critical traffic. In this work, we propose a design and validate the TSN-based communication architecture for the RH system of IFMIF-DONES. The design ensures bounded delivery times for safety-critical interlock signals, achieving a worst-case delay under 30 mu s, even under high network load. The proposed network is also validated in a real robotic teleoperation task, where artificial intelligence is applied for object detection and tracking, using mixed-criticality video streams. Our results show that TSN traffic shapers are essential in providing the necessary latency and bandwidth guarantees for such teleoperation tasks, enabling network convergence in this kind of deployments.

Learning to control the body is a fundamental process in human development. Before acquiring goal-directed skills such as walking or reaching, infants undergo developmental phases characterised by spontaneous movements with no apparent objective. These motions are believed to shape the sensorimotor system by facilitating body-environment interaction. However, how this exploration contributes to sensorimotor structuring remains an open question. A major challenge in studying spontaneous movements has been the lack of appropriate comparative data. To address this, we introduce a synthetic data-driven approach to analyse infant motion. We analysed 20 infants comprising 270 spontaneous movement units from 12 RGB-D infant recordings (22.5 ± 5.96 movement units per infant) together with 206 units from 8 RGB YouTube recordings (25.75 ± 6.34 movement units per infant), and compared these empirical datasets against two synthetic datasets. Our analysis revealed that spontaneous movements, both at the infant and cluster level, engaged arm dynamics more extensively than reaching-like motions and displayed acceleration distributions skewed towards trajectories optimised for maximal dynamic excitation. Furthermore, the kinematic space explored by infants exhibited significantly higher variability. These findings demonstrate that spontaneous movements are dynamically rich, providing emergent features potentially helpful for infants to explore movement possibilities and develop coordination and control.

Access to real-world data in robotics domains is often challenging due to restrictions on data sharing and limited availability. Although privacy and intellectual property concerns are the main barriers, ensuring data access is crucial for advancing data-driven models. Specifically, machine-learning-based inverse dynamic models show promising results for nonrigid robot identification, but the data used to train them are often kept private due to intellectual property protections. Federated learning proposes a methodology to access such data without centralizing them in a single repository, thus avoiding intellectual property limitations. We propose a solution that uses federated learning to train a model from distributed data to develop a robust robotic arm inverse dynamic model. Our approach demonstrates the feasibility of using a machine learning method in which local robots train on their own data while collaborating without sharing raw information. Furthermore, we propose a novel custom aggregation method that integrates locally learned solutions from different workspaces into a single global model without requiring raw data sharing. This method improves accuracy in our federated solution by approximately 20% for the learned inverse dynamic model.

Time transfer accuracy in packet-switched networks is fundamentally dictated by the infrastructure’s timing awareness, shifting from a jitter-limited bottleneck in PTP-unaware systems to an asymmetry- and hardware-limited regime in PTP-aware deployments. This paper presents a structured, context-aware methodology designed to isolate and characterize the diverse network-induced impairments that degrade time transfer accuracy under controlled conditions. This framework organizes the analysis across five aligned stages: (A) analytical infrastructure context, (B) traffic-dynamics modeling, (C) preexperimental calibration, (D) controlled experimentation, and (E) an integrative analysis. As a primary validation of the framework’s baseline capabilities, the methodology is applied to a high-noise scenario in a controlled laboratory testbed: a static, PTP-unaware Spine–Leaf Ethernet fabric where PDV constitutes the dominant driver of offset evolution. The evaluation encompasses one server and three clients with distinct clock characteristics (quartz-based, high-precision oscillator, and high-precision with DPLL) across various traffic profiles and protocol settings. The experimental results reveal that: the DPLL client requires nearly 1 h to acquire lock, whereas the other oscillators lock within seconds; the optimal observation interval scales with variability (5 min under load, 15 min at 0% load)—behaviors uncaptured by current ITU-T recommendations; and the DPLL client yields the lowest time offset despite its locking latency. By successfully isolating these baseline impairments, the method proves time-efficient and establishes a rigorous foundation for extending this statistical pipeline to PTP-aware infrastructures.

This data article describes SLATE-PTP, a high-fidelity IEEE 1588 telemetry dataset captured within a physical two-Spine four-Leaf datacenter fabric engineered with PTP-unaware switches. The data collection architecture isolates network-induced impairments from local oscillator drift. While a GNSS-disciplined Grandmaster clock provided a physical PPS reference to stabilize the hardware clock of the capturing network interface card, the LinuxPTP (ptp4l) daemon operated in a non-adjusting, free-running state. Consequently, the recorded four-way timestamps (t1, t2, t3, and t4) capture pure network dynamics, including queuing delays, PDV, and structural path asymmetries. To replicate realistic datacenter environments, a hardware traffic generator injected background loads using four distinct packet-size distributions combined with three temporal profiles (constant, stepped, and ramped loads). The dataset comprises structured CSV files containing raw packet telemetry. For reuse potential, the data is coupled with an encapsulated, object-oriented Jupyter Notebook execution pipeline. By leveraging a decoupled interface, this pipeline provides an open-access evaluation sandbox. Researchers can directly ingest these data tracks to train, validate, and benchmark custom statistical packet filters, ML-based delay classifiers, or alternative time offset estimation algorithms without deploying a high-precision physical testbed.

The striatum plays a central role in action selection and reinforcement learning, integrating cortical inputs with dopaminergic signals encoding reward prediction errors. While dopamine modulates synaptic plasticity underlying value learning, the mechanisms that enable selective reinforcement of behaviorally relevant stimulus-action associations-the structural credit assignment problem-remain poorly understood, especially in environments with multiple competing stimuli and actions. Here, we present a computational model in which acetylcholine (ACh), released by striatal cholinergic interneurons, acts as a channel-specific gating signal that restricts plasticity to brief temporal windows following action execution. The model implements a biologically plausible three-factor learning rule requiring presynaptic activity, postsynaptic depolarization, and phasic dopamine, with plasticity gated by cholinergic pauses that temporally align with behaviorally relevant events. This mechanism ensures that only synapses involved in the selected behavior are eligible for modification. Through systematic evaluation across tasks with distractors and contingency reversals, we show that ACh-gated learning promotes synaptic specificity, suppresses cross-channel interference, and yields increasingly competitive performance relative to Q-learning in complex tasks, reflecting the scalability of the proposed learning mechanism. Moreover, the model reveals distinct roles for striatal pathways: direct pathway (D1) neurons maintain stimulus-specific responses, while indirect pathway (D2) neurons are progressively recruited to suppress outdated associations during policy adaptation. These findings provide a mechanistic account of how coordinated cholinergic and dopaminergic signaling can support scalable and efficient reinforcement learning in the striatum, consistent with experimental observations of pathway-specific plasticity.

The IEEE 1588 Precision Time Protocol (PTP) is capable of achieving nanosecond-level accuracy over standard Ethernet, making it widely used in industrial and scientific facilities. However, its time offset measurement process is highly sensitive to asymmetries and dynamic variation in the packet delays, hindering its applicability in wide-area use cases. In this paper, we demonstrate that PTP can be enhanced and used for Internet point-to-point time transfer with equal or better performance than the traditional Network Time Protocol (NTP). For this purpose, we propose a technique called offset measurement outlier filtering (OMOF), which leverages the short-term stability of the local clock to discard offset measurements affected by packet delay variation. We implemented OMOF in an open-source PTP client and validated it on a production network connection between two sites located more than 300km apart. Both sites were fitted with high-performance reference clocks synchronized to less than a nanosecond of error using the GNSS-based common-view technique, enabling accurate measurement of our solution’s performance with state-of-the-art metrology techniques. Results show that our approach achieves a 95% improvement in peak-to-peak time error over the most popular open-source PTP implementation and 71% over NTP synchronization.

The cerebellar granular layer plays a central role in sensory processing and pattern separation through its distinctive feedforward architecture. Here, we present a biologically realistic computational model of the granular layer designed to explore the functional impact of synaptic inhibition mediated by Golgi cells. The model integrates anatomical and physiological constraints to simulate realistic mossy fiber activity patterns, including spatial correlations and varying activation levels. We validate the model by replicating key findings from recent in vivo experiments, such as the role of inhibition in shaping granule cell responsiveness and the emergence of nonlinear suppression during multisensory integration. Beyond validation, the model provides a robust computational tool for studying how inhibition contributes to energy-efficient and noise-resilient sensory encoding. Mechanistic analyses revealed that moderate inhibition levels optimize pattern separation performance, with feedforward and feedback inhibitory circuits exerting distinct effects on coding expansion and decorrelation. All model code and simulation scripts are openly available, offering a framework for generating testable hypotheses and further investigating cerebellar computation and learning mechanisms in divergent feedforward networks.

Robots have to adjust their motor behavior to changing environments and variable task requirements to successfully operate in the real world and physically interact with humans. Thus, robotics strives to enable a broad spectrum of adjustable motor behavior, aiming to mimic the human ability to function in unstructured scenarios. In humans, motor behavior arises from the integrative action of the central nervous system and body biomechanics; motion must be understood from a neuromechanics perspective. Nervous regions such as the cerebellum facilitate learning, adaptation, and coordination of our motor responses, ultimately driven by muscle activation. Muscles, in turn, self-stabilize motion through mechanical viscoelasticity. In addition, the agonist-antagonist arrangement of muscles surrounding joints enables cocontraction, which can be regulated to enhance motion accuracy and adapt joint stiffness, thereby providing impedance modulation and broadening the motor repertoire. Here, we propose a control solution that harnesses neuromechanics to enable adjustable robot motor behavior. Our solution integrates a muscle model that replicates mechanical viscoelasticity and cocontraction together with a cerebellar network providing motor adaptation. The resulting cerebello-muscular controller drives the robot through torque commands in a feedback control loop. Changes in cocontraction modify the muscle dynamics, and the cerebellum provides motor adaptation without relying on prior analytical solutions, driving the robot in different motor tasks, including payload perturbations and operation across unknown terrains. Experimental results show that cocontraction modulates robot stiffness, performance accuracy, and robustness against external perturbations. Through cocontraction modulation, our cerebello-muscular torque controller enables a broad spectrum of robot motor behavior.

Controlling collaborative robots (cobots) is a new and challenging paradigm within the field of robot motion control and safe human-robot interaction (HRI). The safety measures needed for a reliable interaction between the robot and its environment hinder the use of classical position control methods, pushing researchers to explore alternative motor control techniques, with a strong focus on those rooted in machine learning (ML). While reinforcement learning has emerged as the predominant approach for creating intelligent controllers for cobots, supervised learning represents a promising alternative in developing data-driven model-based ML controllers in a faster and safer way. In this work, we study several aspects of the methodology needed to create a dataset for learning the dynamics of a robot. To this aim, we fine-tune several PD controllers across different benchmark trajectories using multi-objective evolutionary algorithms (MOEAs) that take into account controller accuracy, and compliance in terms of low torques in the framework of safe HRI. We delve into various aspects of the data extraction methodology including the selection and calibration of the MOEAs. We also demonstrate the need to tune controllers individually for each trajectory and how the speed of a trajectory influences both the tuning process and the resulting dynamics of the robot. Finally, we create a novel dataset and validate its use by feeding all the extracted dynamic data into an inverse dynamic robot model and integrating it into a feedforward control loop. Our approach significantly outperforms individual standard PD controllers previously tuned, thus illustrating the effectiveness of the proposed methodology.

Cuckoo Hashing is a well-known scheme for maintaining a hash-based data structure with worst-case constant search time. The search can be easily pipelined in FPGAs to obtain a response every clock cycle. However, inserting new elements may require multiple iterations to reallocate elements due to collisions, a challenge that intensifies with increasing table occupancy (load). Here, we pursue an efficient implementation of Cuckoo Hashing in FPGAs that exploits the inherent parallel processing capabilities of the technology to minimize the insertion time and enhance the performance of real-time applications like communications networks. Our analysis shows that the candidate implementation alternatives can be grouped into four different categories, considering the number of iterations required to perform new element insertions relative to the hash table’s load. Within each category, the performance is similar, providing flexibility for implementation. Among these methods, one exhibits the best results under all load conditions, without incurring a large complexity penalty, reducing the number of iterations for the insertion by more than 60% for loads beyond 95% compared to the most recent works. A Cuckoo Hashing architecture for networking applications is presented and used to evaluate and verify all insertion methods through hardware evaluation. Our implementation improves on existing architectures by at least 0.2 operations per clock cycle.

The rapid growth of the aging population poses serious challenges for our society e.g. the increase of the healthcare costs. Smart-Health Cyber-Physical Systems (CPSs) offer innovative solutions to ease this burden. This work proposes a general framework adapted in run-time to optimize the system's overall performance, continuously monitoring system working qualities such as response time, accuracy, or energy consumption. Adaptation is achieved through the automatic deployment of different artificial intelligence (AI) based models on local edges (particularly deep learning models (DL)). Local processing is performed in embedded devices that provide short latency and real-time processing despite their limited computation capacity compared to high-end cloud servers. The paper validates this reconfigurable CPS in a challenging scenario: indoor ambient assisted living for the elderly. Our system collects lifestyle user data in a non-invasive manner to promote healthy habits and triggers alarms in case of emergency. Local edge video processing nodes identify indoor activities powered by state-of-the-art deep-learning action recognition models. The optimized embedded nodes locally reduce cost and power consumption, but the system still needs to maximize the overall performance in a changing environment. To that end, our solution enables run-time reconfiguration to adapt in terms of functionality or resource availability, offloading computation when required. The experimental section shows a real setup performing run-time adaptation with different reconfiguration policies considering average times for different daily activities. For that example, the adaptation extends the working time in more than 60% and achieves a 3x confidence in recognition for critical actions.

Complex interactions between brain regions and the spinal cord (SC) govern body motion, which is ultimately driven by muscle activation. Motor planning or learning are mainly conducted at higher brain regions, whilst the SC acts as a brain-muscle gateway and as a motor control centre providing fast reflexes and muscle activity regulation. Thus, higher brain areas need to cope with the SC as an inherent and evolutionary older part of the body dynamics. Here, we address the question of how SC dynamics affects motor learning within the cerebellum; in particular, does the SC facilitate cerebellar motor learning or constitute a biological constraint? We provide an exploratory framework by integrating biologically plausible cerebellar and SC computational models in a musculoskeletal upper limb control loop. The cerebellar model, equipped with the main form of cerebellar plasticity, provides motor adaptation; whilst the SC model implements stretch reflex and reciprocal inhibition between antagonist muscles. The resulting spino-cerebellar model is tested performing a set of upper limb motor tasks, including external perturbation studies. A cerebellar model, lacking the implemented SC model and directly controlling the simulated muscles, was also tested in the same. The performances of the spino-cerebellar and cerebellar models were then compared, thus allowing directly addressing the SC influence on cerebellar motor adaptation and learning, and on handling external motor perturbations. Performance was assessed in both joint and muscle space, and compared with kinematic and EMG recordings from healthy participants. The differences in cerebellar synaptic adaptation between both models were also studied. We conclude that the SC facilitates cerebellar motor learning; when the SC circuits are in the loop, faster convergence in motor learning is achieved with simpler cerebellar synaptic weight distributions. The SC is also found to improve robustness against external perturbations, by better reproducing and modulating muscle cocontraction patterns.

In nature, intelligent living beings have developed emotions to modulate their behavior as a fundamental evolutionary advantage. However, researchers seeking to endow machines with this advantage lack a clear theory from cognitive neuroscience describing emotional elicitation from first principles, namely, from raw observations to specific affects. As a result, they often rely on case-specific solutions and arbitrary or hard-coded models that fail to generalize well to other agents and tasks. Here we propose that emotions correspond to distinct temporal patterns perceived in crucial values for living beings in their environment (like recent rewards, expected future rewards or anticipated world states) and introduce a fully self-learning emotional framework for Artificial Intelligence agents convincingly associating them with documented natural emotions. Applied in a case study, an artificial neural network trained on unlabeled agent’s experiences successfully learned and identified eight basic emotional patterns that are situationally coherent and reproduce natural emotional dynamics. Validation through an emotional attribution survey, where human observers rated their pleasure-arousal-dominance dimensions, showed high statistical agreement, distinguishability, and strong alignment with experimental psychology accounts. We believe that the framework’s generality and cross-disciplinary language defined, grounded on first principles from Reinforcement Learning, may lay the foundations for further research and applications, leading us toward emotional machines that think and act more like us.

Abstract Particle accelerators have high-radiation parts requiring remote maintenance tasks, which involve using cameras generally placed by human experts. This work addresses camera placement through black-box global optimization. This approach allows opting for automated solutions and relieving experts of that responsibility. The formulation of the objective function hides the environment-specific placement constraints and shows domain ones only. The proposed methodology allows the use of regular meta-heuristics for standard black-box box-constrained problems, as it is only necessary to honor the bounds of variables. Accordingly, the optimization algorithm used is the genetic algorithm provided by the Global Optimization Toolbox of MATLAB. The objective function computation relies on the simulation capabilities of the widespread Unity game engine. It brings us an integrated framework for defining the target process to consider and for checking the different proposals visually. Based on this framework, this work reconstructs the virtual reality simulation of a maintenance task in a particle accelerator. The results obtained show that the camera positioning achieved by the proposal for four cameras and one target process outperforms the arrangement defined by a human expert.

The function of the olivary nucleus is key to cerebellar adaptation as it modulates long term synaptic plasticity between parallel fibres and Purkinje cells. Here, we posit that the neural dynamics of the inferior olive (IO) network, and in particular the phase of subthreshold oscillations with respect to afferent excitatory inputs, plays a role in cerebellar sensorimotor adaptation. To test this hypothesis, we first modelled a network of 200 multi-compartment Hodgkin-Huxley IO cells, electrically coupled via anisotropic gap junctions. The model IO neural dynamics captured the properties of real olivary activity in terms of subthreshold oscillations and spike burst responses to dendritic input currents. Then, we integrated the IO network into a large-scale olivo-cerebellar model to study vestibular ocular reflex (VOR) adaptation. VOR produces eye movements contralateral to head motion to stabilise the image on the retina. Hence, studying cerebellar-dependent VOR adaptation provided insights into the functional interplay between olivary subthreshold oscillations and responses to retinal slips (i.e., image movements triggering optokinetic adaptation). Our results showed that the phase-locking of IO subthreshold oscillations to retina slip signals is a necessary condition for cerebellar VOR learning. We also found that phase-locking makes the transmission of IO spike bursts to Purkinje cells more informative with respect to the variable amplitude of retina slip errors. Finally, our results showed that the joint action of IO phase-locking and cerebellar nuclei GABAergic modulation of IO cells’ electrical coupling is crucial to increase the state variability of the IO network, which significantly improves cerebellar adaptation. Author summary This study aims to elucidate the dual functionality of the inferior olive (IO) in cerebellar motor control, reconciling hypotheses regarding its role as either a timing or instructive signal. Specifically, we explore the role of subthreshold oscillations (STOs) within the IO, investigating their potential influence on the climbing fibres-to-Purkinje cell spike pattern responses and subsequent cerebellar adaptation, notably during the vestibulo ocular reflex. Aiming these objectives, we constructed a detailed olivary network model within a cerebellar neural network, enabling a mechanistic analysis of the functional relevance of STOs in spike burst generation, propagation, and modulation within target Purkinje cells. Our findings reveal the intricate nature of complex spike bursts triggered by climbing fibres—IO axons—into Purkinje cell dendrites, demonstrating a hybrid nature involving binary clock-like signals and graded spikelet components acting as an instructive signal. ### Competing Interest Statement The authors have declared no competing interest.

Pre-configured virtual reality (VR) simulations of the logistics and maintenance processes have proven to be useful for identifying potential design issues as well as planning operations during an early design phase of facility. But VR simulations can also be used to deeply explore the feasibility of these procedures in a more interactive manner, so that we can identify potential risks and difficulty levels from early stages and study different maintenance strategies to assist the maintenance worker during these procedures. This article presents a framework to design and validate logistics and maintenance procedures in complex facilities, such as the International Fusion Materials Irradiation Facility DEMO Oriented Neutron Source (IFMIF-DONES). Our framework begins with a preparatory phase where essential information about the procedures and Computer-Aided Design (CAD) models is compiled into a comprehensive Virtualization Task Document (VTD). Differently from previous work, this VTD allows representation of parallel tasks. We implement the interactive version of the virtual environment, where the different maintenance and logistics equipment, as well as a virtual maintenance worker is controlled by the user (the person executing the interactive simulation). We have validated this interactive framework with two simulations for the installation process of the Superconducting Radio Frequency Linear accelerator (SRF Linac) modules in IFMIF-DONES. In one simulation (the automatic one), the procedures are reproduced as they are planned, while in the second simulation (the interactive one) the user freely controls the movements of the moving parts of the crane, grab and release plant equipment, move platforms, etc. Based on our simulations, the interactive version allows easier detection of potential points of collisions as well as more precise assessment of the difficulty of the tasks to be performed.

Driven by the increasing care needs of residents in long-term care facilities, Ambient Assisted Living paradigms have become very popular, offering new solutions to alleviate this burden. This work proposes an efficient edge-cloud system for indoor activity monitoring in long-term care institutions. Action recognition from video streams is implemented via Deep Learning networks running at edge nodes. Edge Computing stands out for its power efficiency, reduction in data transmission bandwidth, and inherent protection of residents’ sensitive data. To implement Artificial Intelligence models on these resource-limited edge nodes, complex Deep Learning networks are first distilled. Knowledge distillation allows for more accurate and efficient neural networks, boosting recognition performance of the solution by up to 8% without impacting resource usage. Finally, the central server runs a Quality and Resource Management (QRM) tool that monitors hardware qualities and recognition performance. This QRM tool performs runtime resource load balancing among the local processing devices ensuring real-time operation and optimized energy consumption. Also, the QRM module conducts runtime reconfiguration switching the running neural network to optimize the use of resources at the node and to improve the overall recognition, especially for critical situations such as falls. As part of our contributions, we also release the manually curated Indoor Action Dataset.

The integration of a time synchronization and time transfer system in data center networks is an increasingly common demand. Since moving the entire infrastructure of a data center network to switches with time-dedicated hardware is nowadays a considerable expense of resources, it is convenient to analyze and study it. In this paper we work on real hardware, presenting a characterization of the problem for a Spine-Leaf network by means of load variation tests proposed by the ITU, using the PTP protocol and common non-PTP aware switches. Through this study we manage to show, thanks to real measurements, the problem of dynamic asymmetry and how it affects the time distribution. Additionally, it is shown that if the network load is quadrupled, the PTP pk-pk offset through the network does not even double. Finally, thanks to long-term tests, we can conclude that time synchronization in a Spine-Leaf network shows considerable resilience to load changes.