ABSTRACT. This paper presents a protocol for achieving average consensus in networks with time-varying directed graphs using broadcast-only communication. The method extends techniques for computing the dominant left eigenvector of an infinite product of time-varying row-stochastic matrices. Using a local rescaling strategy, agents locally compute correction offsets to steer the consensus value toward the arithmetic average. The approach eliminates the need for point-to-point communication and out-degree knowledge, requiring only an upper bound on the total number of agents. Theoretical convergence guarantees are provided, along with numerical results to corroborate the effectiveness of the proposed protocol.

ABSTRACT. In recent years, security and resilience have become a feature of utmost importance in the context of industrial control systems. This submission aims to present novel methodologies to safeguard the integrity of information exchanged within the industrial network. First, an innovative Zero-Knowledge Proof scheme tailored for network control systems is presented, enabling a networked controller to prove its knowledge of the dynamical model and its ability to control a discrete-time LTI system to a sensor, without revealing the model. Then, we propose a two-level security framework: (i) the level of the interconnected devices in the field, that perform a distributed consensus algorithm robust to false data injection attacks; (ii) the upper level of a verifier that poses authentication challenges to the underlying agents. Thanks to the collaboration between the layers, it is possible to implement a recovery procedure that involves re-allocating the mass of the agent recognized as malicious starting from the last authenticated and legitimate value.

ABSTRACT. In this paper, we propose MAcroscopic Consensus and micRoscopic gradient-based OPTimization (MACROPT), a novel distributed method for a network of agents able to learn a probabilistic macroscopic model and concurrently optimize it by acting on the microscopic agents’ states. The macroscopic model is defined through the aggregation of local kernels each representing a probabilistic feature of a single agent (e.g., its local sensing model), while the optimization is done with respect to a given cost index, e.g., the Kullback-Leibler divergence with respect to a target distribution. MACROPT improves the macroscopic model by microscopically coordinating the agents according to a distributed gradient-based policy. Concurrently, it allows each agent to locally learn the macroscopic model through a consensus-based mechanism. We analyze the resulting interconnected method through the lens of system theory. We demonstrate that MACROPT asymptotically converges to the set of stationary points of the nonconvex cost function. The theoretical findings are supported by numerical simulations in sensor network event-detection scenarios.

ABSTRACT. The criticality of Cyber-Physical Systems in managing industrial processes from different critical domains, such as smart manufacturing, transportation systems, and robotics, require the adoption of effective and trustworthy anomaly detection techniques to make CPSs more resilient to faulty conditions. However, the distributed nature of CPSs opens significant challenges including the need for scalable solutions to address the growing complexity and dimensionality of these systems and the necessity to handle heterogenous data from multiple sources of information, each with different nature [3]. Regarding the first point, due to CPSs mostly managing hard real-time systems, there is a need to provide potential detections as fast as possible to help operators reacting against the anomaly without negatively affecting the industrial plant. Secondly, the diversity of information sources requires distributed management, involving the aggregation of groups which share similar characteristics, or monitor the same portion of the industrial process. Finally, due to the critical nature of the industrial processes managed by CPSs, it is essential to ensure that the anomaly detection solution is both trustworthy and reliable, offering clear insights into the reasons behind each specific detection [2].

In our previous work, we demonstrated the effectiveness of a trustworthy decision fusion technique for cyber-physical anomaly detection based on Time-Varying Dynamic Bayesian Networks (TV-DBNs) [2]. Specifically, the solution provides a System-Level Agreed Decision (SLAD) by merging the decisions of local detectors regarding potential ongoing cyber and physical anomalies. In more detail, the TV-DBN enhances the reliability of the overall decision by weighting the local predictions based on both validation performance and the run-time reputation assigned to each detector. This reputation value is dynamically updated based on each detector's level of agreement with the SLAD and determines whether a detector remains valuable for contributing to the SLAD or should be temporarily excluded. However, such a solution did not address either the aggregation of heterogeneous data in a distributed environment, or the scalability issue as the aggregation was performed in a single centralized network. This second issue is particularly challenging as Bayesian inference is an NP-hard problem, which causes the computation time for inference to grow exponentially with the number of nodes and connections within the network. Therefore, without distributing the decision fusion process, merging the decision of a high number of detectors becomes an intractable problem.

To address these issues, in this work, we propose d-TV-DBN, a scalable and distributed architecture to scale cyber-physical anomaly detection with the state-of-the-art TV-DBN-based decision fusion in industrial CPSs [1]. Fig. 1 provides an overview of the proposed architecture: at the bottom, are the industrial plants and operations which implement the industrial process; then the wireless sensor network provides local decisions based on the gathered cyber and physical data from the distributed plant. Following is the distributed decision fusion, where multiple TV-DBNs are employed to merge local decisions into a middle-level aggregated decision. Finally, the upper level provides the final SLAD by further processing the lower decisions.

The architecture of the distributed model consists of two types of tiers: the D tier, consisting of Q cyber-physical detectors, and the H_i,i ∈{1,…,α} tier grouping all the processes performing decision fusion. Specifically, the D tier is split into the partition P(D)={D_1,…,D_β}, where each element groups a set of detector nodes. Similarly, the H_i tier is split into partition P(H_i)={H_(i,1),..., H_(i,γ)}, where each element H_(i,j) groups a subset of hierarchical nodes at level i, and provides an overall SLAD (〖SLAD〗_(i,j)). If H_i=H_1, each element of the tier groups a subset of processes that perform decision fusion from the decisions of D_k∈P(D),k = {1,...,β}. Otherwise, each element of this partition groups a subset of processes that perform higher-level decision fusion from some of the SLADs of the hierarchical level below. Fig. 2 shows an example of the hierarchical architecture, where the D tier is split into partition P(D)={D_1,…,D_10}, for a total of 28 detectors, while H_1 and H_2 tiers into P(H_1)={H_1,1,...,H_1,3} and P(H_2)={H_1,1} partitions. In particular, H_2 is responsible for providing the final SLAD, by merging the H1-level SLADs.

The scalability property of the proposed architecture has been evaluated through a Proof-of-Concept consisting of 19 different configurations where Q is varied from 2 to 20. Such detectors are distributed in the D tier consisting of β detector groups, with β varying from 1 to 5. The maximum number of detectors per group is set to 4. Each detector simulates a cyber-physical anomaly detector providing evidence about the detection of a cyber, physical and cyber-physical anomalies. Results have been reported in terms of the overall accuracy, the Mean Inference Time (MIT) and the Maximum Inference Time (MaxIT) required to provide the SLAD for each PoC configuration. These results have been compared with those obtained using the standalone TV-DBN, in which all detectors are aggregated into a single group. While the TV-DBN encounters a Timeout Error (i.e., when the time required to construct the Bayesian Network exceeds the 2-hour threshold) with Q=11 detectors, our distributed and hierarchical solution keeps the MIT and MaxIT values significantly below the one-second threshold. We have demonstrated that the MIT and MaxIT values increase linearly with the number of detectors, thus surpassing the limitations of the Bayesian inference, which causes the computation time to scale exponentially with the number of detectors.

This work was supported by the Italian National Project INAIL BRIC 2023 ID 44 “Industrial Cyber Shield (ICS)” under CUP C83C22001460001.

[1] S. Guarino, F. Vitale, E. D. Prete, L. Faramondi, N. Mazzocca and R. Setola, "d-TV-DBN: A Hierarchical and Distributed Architecture for Scalable Cyber-Physical Anomaly Detection with Bayesian Networks," in IEEE Transactions on Emerging Topics in Computing, doi: 10.1109/TETC.2025.3573051. [2] S. Guarino, F. Vitale, F. Flammini, L. Faramondi, N. Mazzocca and R. Setola, "A Two-Level Fusion Framework for Cyber-Physical Anomaly Detection," in IEEE Transactions on Industrial Cyber-Physical Systems, vol. 2, pp. 1-13, 2024, doi: 10.1109/TICPS.2023.3336608. [3] W. Yu and C. Zhao, "Online Fault Diagnosis for Industrial Processes With Bayesian Network-Based Probabilistic Ensemble Learning Strategy," in IEEE Transactions on Automation Science and Engineering, vol. 16, no. 4, pp. 1922-1932, Oct. 2019, doi: 10.1109/TASE.2019.2915286.

ABSTRACT. We consider an optimization problem in which a planner aims to maximize the average transient opinion in the Friedkin-Johnsen model by intervening on the agents' innate opinions. Since the full network knowledge is often not accessible, we focus on intervention strategies that are based on statistical instead of exact knowledge of the network. We then conduct a numerical analysis suggesting that the implementation of optimal statistical interventions found by our approach may be close to optimal also on fixed networks.

ABSTRACT. We present a recent stability result for a class of multi-agent systems termed “open” (OMASs) where the agents have the flexibility to join or leave the network at any time. The usual notion of stability for dynamical systems is generalized to that of “open stability” w.r.t. the infinity norm, which makes the distance between two points in the state space independent from their dimension, thus obviating the necessity for distance normalization, as required by the standard Euclidean norm. The practical significance of this stability result is illustrated via numerical simulations that validate a novel protocol for dynamic max-consensus in OMASs.

ABSTRACT. We present a framework for designing optimal state-feedback control that uses a wireless actuation link with imperfect channel state information to transfer the current and future control inputs that actuators can apply if future control messages are lost. The dropout compensation strategy supports scaling inputs to actuators when necessary. We analytically solve finite- and infinite-horizon control problems and present a necessary and sufficient stability condition for any given infinite-horizon state-feedback control law. We validate the results using illustrative examples.

ABSTRACT. This work introduces a novel control framework for Heterogeneous Multi-Agent Systems (HMASs) that enables swarms of agents, which are completely decoupled and lack inter-swarm communication, to achieve a synchronized consensus. This phenomenon, termed blind matching, relies on the concept of fixed emergent output values, where the final steady-state configuration of agents within a swarm is an intrinsic property of the collective dynamics, independent of the agents’ initial conditions or their specific internal dynamics. The theoretical analysis is developed for single-input-single-output (SISO) systems.

ABSTRACT. Nowadays, assistive technologies and service robots have gained increasing prominence in healthcare, not only for supporting patients but also for optimizing clinical workflows and assisting in domestic environments. To take this step, we need technologies that are capable of operating within the real world in a natural and safe manner. This implies endowing robots with capabilities that are inherently human, as the world we aim to interact with has been designed by and for humans. For this reason, humanoid robots are increasingly being used, as their anthropomorphic design gives them the potential to replicate various human-level skills. In hospital or domestic settings, for instance, a robot may be asked to carry objects to a patient, or to push a wheelchair. To achieve effective and efficient results in this scenario, a close collaboration between robotics research centers and clinical laboratories is essential. This is precisely the approach promoted by the Fit4MedRob project. In this work, we present the matchmaking between the University of Pisa and Istituti Clinici Scientifici Maugeri IRCCS Milano. We outline our experience with the Alter-Ego humanoid robot for assistive tasks, highlighting potential applications in both domestic and healthcare settings, supported by preliminary results, and considerations on a specific task related to cartpushing.

ABSTRACT. Augmented Reality (AR) enhances interaction with the physical world by overlaying digital content onto real environments. While head-mounted displays offer immersive AR, they often cause discomfort, tracking issues, and cybersickness, limiting prolonged use. Mobile devices like smartphones and tablets are a practical alternative for extended AR sessions, but their handheld nature hinders use during manual tasks, such as manufacturing or surgery. This paper presents a robotic solution that automates the movement of a mobile device in AR settings. The system adjusts the device’s pose to maintain alignment with the user’s viewpoint and moves it away when not needed, enhancing task efficiency. Experiments with 20 participants show that the system reduces perceived workload and improves usability compared to handheld use, without affecting task completion time.

ABSTRACT. In human augmentation, researchers have successfully demonstrated the control of supernumerary limbs but none have investigated the human potential to control the maximum number of degrees of freedom (DoFs) from both kinematic and cognitive perspectives. Existing works impose both the body part to be used and the specific movement to perform, limiting user adaptability. Here, in order to characterize the human potential, we implemented a strategy enabling the control of multiple DoFs through user-defined movements. Subjects can efficiently manage up to two DoFs with little practice, good performance, and acceptable perceived workload. Consistent performance when controlling more than two DoFs requires additional training.

ABSTRACT. This abstract analyses the state of the art and future perspectives of the research field of Human-Robot Augmentation by supernumerary robotic limbs. We claim that the success of this new human-robot interaction paradigm largely depends on the accurate design of suitable sensorimotor interfaces connecting the human to the extra limbs.

ABSTRACT. This research aims at exploiting a novel control architecture for achieving a dynamically stable ultra-slow gait in quadruped robots, addressing challenges related to locomotion on unstructured and slippery terrains. Building on prior research in two-leg balancing, the method employs imitation learning (IL) to train two neural controllers that replicate a model-based expert while satisfying Lyapunov-based stability and safety constraints. A switching mechanism enables stable alternation between leg pairs, allowing slow, deliberate movements with continuous balance control. The neural networks are trained in simulation using a feedforward architecture with augmented loss functions incorporating Linear Matrix Inequality (LMI) conditions. Preliminary results show accurate center-of-mass tracking and reliability across gait phases, laying the groundworkfor real-world deployment on uneven terrains.

ABSTRACT. "Field Robotics" is the branch of robotics dedicated to designing, developing, and deploying autonomous or semi-autonomous robots for effective operation in unstructured, dynamic, and often challenging real-world environments—distinct from traditional, controlled factory settings. This includes diverse applications such as agriculture, construction, environmental monitoring, search and rescue, exploration, etc.

Several challenges affect such robots from the control theory perspective, standing out uncertainty (in the dynamic model parameters) and external disturbances, which complicate the design and implementation of model-based controllers. For instance, an autonomous aquatic vehicle, when it is deployed in an outdoors waterbody, it subjected to wind and water currents acting as external disturbances. Besides, its payload including the control box, batteries, straps and other hardware elements turn difficult to accurately find its dynamic parameters. A heavy-duty quadcopter, for example, is also affected by wind gusts when flying outdoors, and its inertial properties also change depending on the payload added accordingly to accomplish a determined mission.

Another example to be considered is a mobile manipulatobuilt for agriculture purposes. Its inertial properties vary according to its payload, thereby a model-based control approach is hard to tune. Hence considering the aforesaid scenarios, utilizing model-free approaches is reasonable to control the motion of field robots while effectively counteracting both external disturbances and uncertainties.

Among the diverse options available in the literature, Sliding Mode Control (SMC) excels in providing robustness to the controlled systems, and some recent advances have shown that it is possible to attenuate or eliminate the characteristic chattering effect that SMC approaches bring about. As a case in point, it has been proven that nested-continuous SMC is advantageous to robustly control robotic systems. In the literature, it has been shown that the fully nested Super-Twisting SMC algorithm, is effective to address trajectory tracking problems despite the external disturbances. However, it still depends on the accurate knowledge of robot dynamics parameters, which is difficult to achieve in practice, at least for the case of field robots, since they are employed to perform heavy-duty activities in harsh environments as it was mentioned before. Thus, the solution proposal is to take the advantage of the robustness supplied by SMC, but leaving aside the model dependency, in addition to nesting the discontinuous term of the Super-Twisting, so that a continuous control signal is supplied to the robotic system. By means of a thorough Lyapunov-based analysis, it can be shown that the position error dynamics trajectories converge uniformly asymptotically to zero, through the sliding surface, despite uncertainty.

The challenges for the implementation of this approach are the necessity of trustworthy sensor measurements and the gains tuning. While sensor signals can be filtered and complemented with estimators, gains tuning can be instead addressed using adaptive techniques, therefore, nested-continuous SMC is a promising model-free control approach for uncertain disturbed systems like field robots.

ABSTRACT. This research proposes a multi-layer navigation system for indoor mobile robots when they share space with vulnerable individuals. The primary objectives are increasing or maintaining safety measures and curtailing operational costs, emphasizing reducing reliance on intricate sensor technologies and computational resources. The developed system employs a three-tiered control approach, with each layer playing a pivotal role in the navigation process. The "online" control layer integrates a human-in-the-loop strategy, where the human operator detects missing obstacles or approaching danger through a user interface and sends a trigger to the robot's controller. This trigger enables the system to estimate the coordinates of the danger and update the robot's navigation path in real time, minimizing reliance on complex sensor systems.The "semi-online" control layer generates dynamic virtual barriers to restrict the robot's navigation in specific areas during specific times. This ensures the robot avoids hazardous zones that could pose temporary risks to the human or robot. For example, areas with temporary obstructions or potential danger, such as kids' play zones or during cleaning, are temporarily restricted from the robot's path, ensuring safe navigation without relying solely on real-time sensor data. The "offline" control layer centers around the use of semantic information to control the robot's behavior according to user-defined space management and safety requirements. By leveraging Building Information Models (BIM) as digital twins, this layer combines semantic and geometric data to comprehensively understand the environment. It enables the robot to navigate according to precise user requirements, utilizing the semantic context for path planning and behavior control. This layer obviates the need for a real-time sensor mapping process, making the system more efficient and adaptable to user needs. This Research represents a significant step forward in enhancing the navigational capabilities of robots within human-centric indoor environments. With a core focus on safety, adaptability, and cost-effectiveness.

ABSTRACT. This work presents a unified control architecture for semi-autonomous robots designed for precision agriculture, with a focus on collaborative human-robot harvesting tasks in vineyards. Unlike conventional industrial robotics, agricultural settings require adaptable, sensor-rich, and compliant robotic systems capable of operating in unstructured environments. The proposed system integrates a high-level task planner and a low-level control scheme based on Hierarchical Quadratic Programming (HQP) and Control Barrier Functions (CBFs). The architecture enables optimal task allocation between humans and robots through a Mixed Integer Linear Programming (MILP) approach that minimizes execution time, balances workload, and incorporates human preferences. At the control level, the robot adapts its behavior for safety, compliance, and configuration optimization. A perception module allows grape and peduncle detection, enabling accurate harvesting actions. The robot's autonomy level dynamically adjusts based on perception outcomes and task execution, supporting shared control and safe physical interaction. Field validation experiments demonstrate the system’s effectiveness in enhancing efficiency and safety in dynamic agricultural environments.

ABSTRACT. This study proposes a hybrid truck–robot approach to enhance last-mile delivery in urban centres. Trucks unload parcels at local hubs, and autonomous sidewalk robots complete the deliveries, reducing vehicle traffic in historic zones. An optimisation model is developed to maximise daily deliveries, considering parcel sizes and robot capacities. The method is tested through simulation in Bari (Italy) using the SUMO platform with real-world constraints. Results confirm improved time and operational efficiency compared to traditional delivery.

ABSTRACT. In this extended abstract, a detailed analysis of the sources of instability for proxy-based constrained admittance controllers in pHRI control schemes by considering parasitic effects such as transmission elasticity, motor velocity saturation, and actuation delay is conducted. Furthermore, a sensitivity analysis is also conducted.

ABSTRACT. This paper advocates for modularity in data-driven cognitive architectures to advance Human-Machine Systems (HMS). Inspired by cognitive science, this approach proposes distinct, specialized modules to interpret human behavior, enabling targeted solutions and independent optimization. Modularity enhances robustness, interpretability, and extensibility, allowing for formal guarantees on components and seamless integration of new capabilities. We present a toolkit of data-driven feedback control modules: a Perception Module for real-time state estimation, a Coordination Control Module for subconscious entrainment, and a Social Communication Module for emotion encoding in movement kinematics. The power of this framework lies in the synergistic interaction of these modules, leading to richer, context-aware behaviors. We frame learning problems as control problems, leveraging data-driven methods to achieve control-theoretic objectives like stability and performance in complex environments. This modular approach, exemplified by applications in remote assisted physiotherapy, offers a tangible path towards intelligent, trustworthy, and collaborative HMS, addressing key conference themes such as adaptive control, reinforcement learning, human-in-the-loop systems, and trust in AI.

ABSTRACT. Planning robot trajectories in human-populated environments is inherently challenging, as control algorithms cannot depend on deterministic predictions of human motion, and the environment itself is intrinsically uncertain. Nevertheless, even stochastic estimates of future trajectories can substantially enhance path planning. In this work, we introduce a hierarchical planning framework that integrates a macro and a micro planner for navigating a robot in crowded environments. The macro planner provides high-level guidance by identifying preferred regions of the environment based on macroscopic features. In contrast, the micro planner operates at a finer level, taking into account dynamic constraints and safety considerations to compute feasible and safe trajectories in the short term. Our framework enables the robot to act proactively by anticipating possible deviations from the predicted trajectories of obstacles. This anticipatory behavior allows the robot to avoid collisions through smoother and less abrupt motion adjustments.

ABSTRACT. Diabetes mellitus type 1 (T1D) is a chronic metabolic disorder marked by persistently high blood glucose levels, which, if not properly managed, can lead to progressive damage to vital organs including the heart, blood vessels, eyes, kidneys, and peripheral nerves. Automated Insulin Delivery (AID) systems—combining Continuous Glucose Monitors (CGMs), insulin pumps, and control algorithms—have revolutionized T1D management by enabling real-time, adaptive insulin dosing. However, the complexity and variability of glucose-insulin dynamics have spurred interest in machine learning, particularly Reinforcement Learning (RL), as a means to develop more adaptive and personalized control strategies.

In this paper, we present two recent contributions from our research group that advance the development of AID systems for T1D management.

The first study investigates the use of the Deep Deterministic Policy Gradient (DDPG) algorithm for insulin infusion control, benchmarking its performance against classical controllers such as Model Predictive Control (MPC) and Proportional-Integral-Derivative (PID). Results show that DDPG not only matches but outperforms these traditional strategies in minimizing hypo- and hyperglycemic episodes and stabilizing glucose levels.

The second study proposes a safety-aware RL framework based on Maskable Proximal Policy Optimization (Maskable PPO), which incorporates invalid action masking to enforce safety constraints during learning and deployment. This framework ensures safe and efficient insulin dosing across diverse patient cohorts—including adults, adolescents, and children—leading to notable gains in Time In Range (TIR), reductions in total insulin administration, and fewer emergency bolus interventions.

By consolidating these findings, we offer new insights and methodologies to the community, supporting the development of more robust, safe, and personalized AID systems for effective T1D management.

ABSTRACT. Brain Computer Interface (BCI) is a system that establishes a direct communication channel between brain neural activity and external systems, offering transformative assistive capabilities for individuals with motor impairments. Among non-invasive paradigms, Motor Imagery (MI) BCIs exploit sensorimotor rhythm modulations associated with imagined limb movements to yield continuous and intuitive control signals. This paper proposes a BCI that enables users to pilot the Alter-Ego robot via a menu-driven graphical interface by merely imagining left- or right-hand movements. Electroencephalographic signals are acquired with a 14-channel Emotiv X headset sampled at 256 Hz; only the six electrodes located over the sensorimotor cortex are fed to the processing pipeline. Performance are evaluated in three consecutive runs with a single participant, yielding a total of 23 cued commands. Accuracy, rejection rate, and F1-score attest to the robustness and real-time suitability of the proposed approach for intuitive robot manipulation.

ABSTRACT. Phase estimation is a classical problem in systems theory, where instantaneaous phase values are associated to (almost-)periodic signals, with applications in control systems involving, e.g., biological signals and human motion data. Existing methods are hindered by offline operation, dimensionality reduction, and noise sensitivity. To address these limitations, we propose ROPE (Recursive Online Phase Estimator), a novel algorithm for real-time phase estimation of multidimensional pseudo-periodic signals. Extensive validation on chaotic trajectories, human motion data, and ECG signals shows significantly superior performance, enabling advanced control and synchronization in human-in-the-loop systems.

ABSTRACT. Human-In-The-Loop (HITL) control of robot hands via surface electromyography (sEMG) commonly uses machine learning to decode grasping intentions. However, challenges remain due to complexities and inaccuracies that limit stable sEMG-based interfaces. Existing supervised and unsupervised methods show structural limits in regression precision and reliability of simultaneous multi-motion control. We propose minimally supervised machine learning to address these issues, exploiting minimal supervision to manage nonlinear fitting and imprecise labeling in sEMG signals. Two implementations are presented: one using a differentiable Dynamic TimeWarping (soft-DTW) loss for neural network training, and another combining Non-Negative Matrix Factorization (NMF) with neural networks for self-supervised regression. Offline tests with multiple subjects show promising results.

ABSTRACT. Virtual Coupling (VC) has gained increasing attention as a next-generation railway signaling solution. Its implementation, however, poses critical control challenges, as following trains must maintain small inter-train spacing under dynamic operating conditions. This paper proposes a decentralized control framework for VC followers, where safety is defined by the relative braking distance (RBD) principle and enforced through High Order Control Barrier Functions (HOCBFs). An adaptive switching heuristic is introduced to compute the reference control force while accounting for potential Train-to-Train (T2T) communication delays. The HOCBF constraints and reference control are embedded into a sequence of Quadratic Programs (QPs) solvable in real time. The proposed method guarantees safety for VC followers without requiring centralized coordination or full trajectory prediction of the preceding trains.

ABSTRACT. Modeling non-linear systems has always been a challenge in the field of control engineering. Linear Parameter Varying (LPV) models can be a valid choice to model complex systems. A common approach to identify a LPV system in an ARX form is the Least Squares Support Vector Machines (LS-SVM) method. However, due to its computational complexity, it is difficult to employ such algorithm in online applications, when a model must be identified each time a new datum is collected from the system. An efficient recursive update algorithm has been recently presented in the literature for such cases, where only the most informative data points are selected to update the model, thus generally reducing the required computational effort. However, in certain conditions such algorithm selects too many data points, still leading to an high computational time. In this work, a quadratic entropy based algorithm is proposed to overcome these limitations.

ABSTRACT. The current climate crisis necessitates a shift toward a more sustainable lifestyle, involving the adoption of sustainable technologies and services, and understanding how to foster the adoption of such green technologies in a systematic way is crucial to facilitate such a green transition. In light of this need, this work introduces an extension of the Friedkin-Johnsen opinion dynamics model to include the long-term effects of (structural) policies aimed at fostering sustainable behaviors. By incorporating mechanisms that capture the influence of external interventions on individuals’ opinions, our model enables a nuanced understanding of how attitudes toward green technologies evolve over time. Building on this framework, we propose an optimal nudging strategy designed to balance the goals of widespread adoption with investment limitations imposed by budget constraints. Through numerical tests, we demonstrate how targeted interventions can shape individuals’ inclinations toward sustainability.

ABSTRACT. We present an approach to identify a quasi Linear Parameter Varying (qLPV) model of a plant, with the qLPV model guaranteed to admit a robust control invariant (RCI) set. It builds upon the concurrent synthesis framework recently presented by the authors, in which the requirement of existence of an RCI set is modeled as a control-oriented regularization. Here, we reduce the conservativeness of the approach by bounding the qLPV system with an uncertain LTI system, which we derive using bound propagation approaches. The resulting regularization function is the optimal value of a nonlinear robust optimization problem that we solve via a differentiable algorithm. We numerically demonstrate the benefits of the proposed approach over two benchmark approaches.

ABSTRACT. The problem of estimating the reaction coefficient of a system governed by a reaction-diffusion partial differential equation is tackled. An estimator relying on boundary measurements only is proposed. The estimator is based upon a setpoint regulation strategy and leads to an asymptotically converging estimate of the unknown reaction coefficient. The proposed estimator is combined with a state observer and shown to provide an asymptotic estimate of the actual system state. A numerical example supports and illustrates the theoretical results.

ABSTRACT. In the context of non-convex mixed-integer robust scenario optimization problems, we show that a separable structure in the constraints can be exploited to bound the complexity associated to the solution. The evaluation of this bound can be easily computed by solving a linear program and, importantly, can be computed prior to the solution of the scenario problem itself. These bounds can be used, once scenarios are collected, to find suitable certificates to the solution's risk, as well as allowing to determine the sample size necessary to achieve a user-defined reliability level, both in a one-shot and multi-step approach.

ABSTRACT. This extended abstract focuses on the application of the Super-twisting (STW) Sliding Modes algorithm to the consensus problem in multi-agent systems with first-order perturbed dynamics. Here is developed a method to construct strict Lyapunov functions for the STW signed Laplacian flow protocol over directed weighted-balanced graphs, both with and without perturbations.

ABSTRACT. In this paper, we focus on two properties related to the privacy of partially observable discrete event systems: state-trajectory opacity and initial-state opacity. Both properties are trajectory-dependent, meaning that if a trajectory violates the property, all of its continuations will also violate it. Our contribution lies in demonstrating that the verification of these two properties can be addressed using the notion of fault diagnoser. Specifically, we define two ad hoc automata, one for each property, and show that, by analyzing their respective diagnosers, the properties can be verified. While the complexity of our method is the same as that of other state of the art procedures, our approach opens the door to tackling complex problems by combining approaches developed within two areas of opacity and diagnosis.

ABSTRACT. Offline training of Deep Reinforcement Learning agents offers a practical solution for autonomous control and robotics, particularly when real-world data is scarce and simulators are unavailable. However, its effectiveness is limited by sample inefficiency, especially with high-dimensional inputs like images. Counterfactual data augmentation mitigates this by generating plausible samples aligned with the environments’ dynamics, yet its use in vision-based control remains limited. This work presents a novel counterfactual augmentation scheme for vision-based tasks, using Deep Generative Models to learn the Structural Causal Model (SCM) of the environment and the reward function. The learned model is then used to enrich the training data and improve policy learning. We demonstrate the superiority of our approach over state-of-the-art baselines across various stochastic control tasks with both discrete and continuous action spaces.

ABSTRACT. Visual Odometry (VO) and Visual SLAM (V-SLAM) systems often struggle in low-light and dark environments due to the lack of robust visual features. In this paper, we propose a novel active illumination framework to enhance the performance of VO and V-SLAM algorithms in these challenging conditions. The developed approach dynamically controls a moving light source to illuminate highly textured areas, thereby improving feature extraction and tracking. Specifically, a detector block, which incorporates a deep learning-based enhancing network, identifies regions with relevant features. Then, a pan-tilt controller is responsible for guiding the light beam toward these areas, so that to provide information-rich images to the ego-motion estimation algorithm. Experimental results on a real robotic platform demonstrate the effectiveness of the proposed method, showing a reduction in the pose estimation error up to 75% with respect to a traditional fixed lighting technique.

ABSTRACT. In this paper, we propose a novel distributed data-driven optimization scheme. In particular, we focus on the so-called aggregative framework, namely, the scenario in which a set of agents aim to cooperatively minimize the sum of local costs, each depending on both local decision variables and an aggregation of all of them. We consider a data-driven setup in which each objective function is unknown and can be only sampled at a single point per iteration (thanks to, e.g., feedback from human users or physical sensors). We address this scenario through a distributed algorithm that combines three key components: (i) a learning part that leverages neural networks to learn the local cost functions descent direction, (ii) an optimization routine that steers the estimates according to the learned direction to minimize the global cost, and (iii) a tracking mechanism that locally reconstructs the unavailable global quantities. By using tools from system theory, i.e., timescale separation and averaging theory, we formally prove that, in strongly convex setups, the overall distributed strategy linearly converges in a neighborhood of the optimal solution whose radius depends on the given accuracy capabilities of the neural networks. Finally, we corroborate the theoretical results with numerical simulations.

ABSTRACT. Flow control in microchannels is a rapidly evolving field focused on the manipulation of small quantities of fluids, enabling highprecision microfluidic processes across a wide range of domains, including chemical synthesis and crystallization (Rasche et al., 2016), biological assays such as blood rheology (Del Giudice et al., 2022), and thermal management systems like microscale refrigeration (Ye, Yuxin, et al., 2022). Flow control becomes particularly challenging when multiple fluids (multiphase flow) interact in the same microchannel, necessitating the combination of the Navier–Stokes equations, which govern microscale fluid motion, with models to represent the change in space and time of the interface between fluids, surface tension, wettability, and interfacial effects (Tabeling, 2023). In such situations, model-based approaches are complex, computationally intensive, and often lack the adaptability required for real-time control applications.

This research presents a combined micro-optofluidic device, capable of in-situ monitoring, alongside a data-driven approach, to model nonlinear, distributed flow dynamics and to establish real-time controls guaranteeing system stability and reproducibility. In particular, as a test-bench for this approach, we consider two immiscible fluids interacting in a micro-channel, in two different configurations: liquid-liquid or liquid-gas. The objective is to maintain an interlaced sequence of the two fluids, known as slug flow, by controlling the frequency of their passage, their length, and velocity. The micro-optofluidic device—capable of non-invasive, real-time monitoring of slug velocity, length, and frequency—was originally presented in Stella et al., 2024, while the fully integrated sensing-and-control framework was introduced in Moscato et al., 2024.

The working principle, geometry, and a picture of the micro-optofluidic device are shown in Fig.1. The device features a T-junction with two inlets and one outlet to generate the slug flow and an optical setup consisting of one input fiber for light delivery and two output fibers for collecting the transmitted signal. Detection relies on absorption phenomena: as fluids with different refractive indices pass through the channel, they modulate the optical signal by selectively altering light transmission and reflection, as shown in Fig.2.

The micro-optofluidic device performance and optical methodology were initially validated using both liquid-liquid (HexadecaneWater, Fig.3(a)) and gas-liquid (Air-Water, Fig.3(b)) flow under a wide range of flow conditions and compared with the theoretical and numerical (Computational Fluid Dynamics, CFD) results. The comparison confirmed the reliability of the micro-optofluidic device and the consistency of the observed slug dynamics and validated the accuracy of the slug frequency and velocity estimation method. Some discrepancies found—especially in gas–liquid regimes—revealed limitations of relying solely on theoretical models and imposed flow rates, highlighting the need for in situ and continuous monitoring (Moscato et al., 2025).

The control framework used to regulate slug frequency was chosen to be based on a data-driven, constrained Model Predictive Control (MPC) scheme operating in closed-loop.

A second-order linear model was identified using data-driven system identification, based on experimental analysis of the system’s steady-state and transient responses to different input flow rate conditions for the liquid-liquid experiments, with the input flow rate as input variable, and the slug frequency as output variable to be controlled. To assess the most suitable and reliable controller for our application purposes, a comparison in simulation between three state-of-the-art controllers has been carried out, considering PID, LQR, and MPC. The obtained results, shown in Fig.4, demonstrate the MPC's ability to adjust and constrain the input variables while maintaining satisfactory performance. Indeed, the constrained MPC enables the generation of smooth control actions, which are particularly useful for preventing damage to the microfluidic device (see Fig.4(b)). This approach also allows for the definition of control input constraints, thereby maintaining the desired slug flow process without the need for parameter tuning.

Then the control architecture, shown in Fig.5, was implemented by the integration of a constrained MPC algorithm running on a Raspberry Pi, which computes optimal flow rates and imposes them on the syringe pumps connected to the micro-optofluidic device. The slug process is monitored via a laser-based optical setup, with the transmitted signal acquired by a photodiode and feedback to the controller, enabling real-time closed-loop regulation. To assess the constrained MPC performance and show the system's capabilities of tracking different reference values, a multi-reference experimental condition was evaluated, see Fig.6. Specifically, the imposed values have been chosen to highly solicit the control loop by interchanging high and low slug frequency reference values. Closed-loop experimental validation showed consistent performance in terms of steady-state errors, settling times, and overshoot across the range of setpoints and operating conditions considered, demonstrating the effectiveness of the approach and its applicability in real-time lab-on-chip devices. Although validated on liquid–liquid regimes, the approach is also extendable to gas–liquid configurations. Future work will be done in the direction of performing non-linear model identification to implement a nonlinear MPC and, by leveraging the efficacy of the proposed methodology, a slug velocity control will be envisaged and implemented, opening new ways of controlling bi-phase processes.

Acknowledgment This work is the result of the research project funded by PRIN 2022 Settore PE7 progetto "HOME4.0: Brain Signal Humanoid Integrated home assistant platform" (CUP: E53D23000510006)

ABSTRACT. This work studies security in a two-agent decentralized network, focusing on detecting and locating attacks. Both agents and the attacker are modeled as finite state machines, with the attacker acting on communication channels. An integrated model of network evolution under attack is developed, providing necessary and sufficient conditions for detecting and localizing the compromised channel.

ABSTRACT. Dynamic systems are commonly represented in local state space, emphasizing state evolution over time. However, this approach poses challenges in analysis, prediction, and control. A more structured alternative is the use of Ordinary Differential Equations (ODEs), which offer a compact and expressive mathematical formulation but are often computationally demanding and difficult to control. To mitigate these limitations, spectral decomposition techniques enable the transformation of nonlinear dynamics into a linear framework, simplifying analysis and control. In this study, we compare two prominent methodologies for achieving such decompositions: Machine Learning (ML) and Koopman Operator (KO) theory. Specifically, we investigate purely data-driven ML, Physics-Informed ML (IML), data-driven KO, and ODE-based KO, analyzing their respective strengths and limitations. Using the mass-spring-damper system as a case study, we explore the effectiveness of these approaches when either only observational data or explicit ODE formulations are available. To the best of our knowledge, no prior work has systematically compared ML and KO-based methods under these conditions. Our findings aim to provide insights into the advantages and trade-offs of data-driven and physics-informed techniques for modeling and controlling dynamical systems.

ABSTRACT. This abstract presents DOT-ADMM to solve online learning problems in a multi-agent setting offering the following set of features: (i) Convergence with a linear rate for a wide class of learning problems (e.g., linear and logistic regression); (ii) Applicability in an online scenario where the data sets available to the agents change over time; (iii) Robustness to asynchronous processing/updates; (iv) Robustness to inexact local computations; (v) Robustness to faulty and noisy communications. Numerical simulations reveals how DOT-ADMM outperforms other state-of-the-art algorithms, which is the only one that may deal with challenges (ii)–(v) at the same time.

ABSTRACT. Efficient traffic management in high-density ur- ban areas is a critical challenge, specifically in highway entry ramps and merging points. This paper deals with the problem of managing vehicles that have to enter highway ramps and intersections. The problem is formulated as an optimization task to minimize sudden braking, waiting times, and collision risks. A Deep Reinforcement Learning approach based on the Actor- Critic framework is proposed to train the intelligent traffic light agents managing the vehicle entries. The reward function is designed to dynamically balance safety and efficiency by reducing braking events and congestion. A simulation campaign is performed in a high-density merging scenario located in central-northern Italy at the intersection between the A1 Highway and the A14 Motorway. The results demonstrate the benefits obtained in reduced traffic flow and increased safety.

ABSTRACT. In the age of intelligent transportation systems and smart cities, teleoperated driving aims to bridge the gap between human and fully autonomous driving. However, the reliability of teleoperated driving is heavily dependent on the quality of the cellular networks, which could be overcome with 6G networks that aim to enhance ultralow latency and high reliability. This study proposes an architecture for designing an integrated simulator for teleoperated driving by utilizing Deep Reinforcement Learning (DRL) in a comprehensive framework of 6G and Network Digital Twin. The presented simulation framework combines different tools (i.e., SUMO, OMNeT++, and Simu5G) to model realistic traffic and network dynamics. In addition, the Random Forest algorithm is used for the coverage prediction system and maintaining stable connectivity, and a DRL model optimizes vehicle routing by balancing path length and signal coverage. A case study is simulated considering the city of Bari (Italy). The framework demonstrates robust communication between teleoperated vehicles and 6G Digital Twin infrastructure.

ABSTRACT. Whole-cell quantitative models are gaining interest to explain complex phenotypes and to accelerate biological discoveries. Here we present a model integrating metabolism, growth and cycle, the main cellular functions in living cells. Within the spirit of whole-cell models, it is conceived to be modular and hierarchical. Its coarse-grained backbone is able to reproduce different quantitative and qualitative experimental findings both in single cells and in populations. Besides, it offers a robust scaffold for plugging-in finer molecular details in each coarse module, so allowing to identify the molecular basis of mutant strains and to assign genotype-to-phenotype correlations.

ABSTRACT. Combining targeted therapeutics can significantly help address the dynamic changes in cancer biology abnormalities and thus improve the duration of response and outcome. However, the efficacy of such approaches is highly dependent on the combination, interactions, and timing between the administered drugs. Current clinical trials can test only a low number of schedules with fixed designs. Pharmacometric tools can assist in exploring and selecting the most effective drug dosages and schedules by modeling traits of patients with different clinical and biological characteristics. Methods: This study proposes a pharmacokinetic-pharmacodynamic model describing the networked system of tumor development and angiogenesis under the control of antiangiogenic and cytotoxic, i.e., Ramucirumab and Paclitaxel second-line combination therapy. A two-step scalable algorithm is proposed to calibrate model parameters and match virtual to real population therapy outcomes, followed by fine-tuning directly on the Progression-free Survival (PFS)-2 Kaplan-Meier curve. Two cohorts of advanced gastric cancer patients were considered: a calibration cohort from South Korea, and an external verification cohort from IRCCS "S. De Bellis", an Italian research hospital. These real-world patients had heterogeneous clinical starting conditions. We perform prospective evaluations of new combination regimens that adhere to pharmacological constraints that are paramount for clinical translation, in which the administration time of the cytotoxic agent is triggered by the normalization window opening, monitored by a tumor microenvironment digital biomarker. Results: The calibration procedure led to the discovery of a new mathematical biomarker describing the influence of intrinsic tumor growth and angiogenesis on treatment outcomes. The predictive value was assessed through the log-rank test between two PFS-2 groups, which exhibited different therapy response trends (p-value <0.0001). Our results showcase a new regimen that, by using 33% less cytotoxic drug, achieves indistinguishable PFS-2. Additionally, we present another regimen that extends PFS-2 from 49.2% to 60.9% after 121 days of therapy (p-value <0.0001), by using the same dosing as the standard protocol.

ABSTRACT. Despite the technological advances of the last years, glucose control in type 1 diabetes still remains a challenging task. The main features that complicate the control problems are the high inter- and intra-patient variability and the limitations in data collection due to safety concerns. This work presents the last achievements related to the development of an artificial pancreas able to adapt itself to the system changes over the time. Inter-patient variability is addressed through personalized models, exploring classical impulse-response technique and hybrid physics-informed neural networks. Soft-switched multi-model MPCs and periodic MPCs have been developed to manage intra-patient variability. Safety is enhanced via asymmetric cost functions based on the clinical index and predictive layers have been designed for early hypoglycemia detection/prevention and meal estimation. Simulations with the UVA/Padova T1D simulator show better glucose control in terms of time in target range and hypoglycemia reduction.

ABSTRACT. Monitoring industrial processes and understanding deviations is critical in ensuring product quality, process efficiency, and early detection of anomalies. Traditional methods for dimensionality reduction and anomaly detection, such as Principal Component Analysis (PCA) or Partial Least Squares (PLS), often struggle to capture the complex and dynamic nature of batch data. In this study, we propose a novel approach that combines an AutoEncoder (AE), based on Long Short-Term Memory (LSTM) layers, with a rolling threshold for anomaly evaluation. Unlike conventional threshold methods that rely on global statistical parameters, the applied threshold leverages rolling median and rolling Median Absolute Deviation (MAD) to adaptively detect deviations, making it more resilient to outliers and distribution shifts. The LSTM-AE demonstrates superior performance in anomaly detection with respect to PCA and more recent model approaches, specifically for the reference dataset, obtained from a GlaxoSmithKline (GSK) production plant. Additionally, an LSTM regression model is employed to forecast future data points, which are then fed into the LSTM-AE to enable a predictive approach. This framework leverages the temporal dependencies captured by LSTM layers and reconstruction efficiency of the AE, facilitating a predictive anomaly detection in real-world applications.

ABSTRACT. Type 1 diabetes mellitus is characterized by dangerously elevated blood glucose levels resulting from inadequate or absent insulin production. To maintain proper glucose regulation, patients with this condition require external insulin administration. This study compares four control strategies for the autonomous management of blood glucose: integer-order proportional-integral (IO-PI), fractional-order proportional-integral (FO-PI), integer-order sliding mode control (IO-SMC), and fractional-order sliding mode control (FO-SMC). The controllers are optimized using genetic algorithms and evaluated through numerical simulations based on the fractional Bergman glucoregulatory model, incorporating meal disturbances. Results indicate that FO-SMC outperforms the other methods, offering better glycemic control by maximizing time within the normal range and minimizing episodes of hypoglycemia, while also demonstrating robustness to variations in meal intake.

ABSTRACT. Type 1 diabetes mellitus is a chronic condition that requires insulin delivery to maintain blood glucose levels within a desired range. The artificial pancreas (AP), which integrates a continuous glucose sensor, an insulin pump, and a control algorithm, is a promising solution for automating insulin delivery. Designing optimal controllers for the AP is crucial to its effectiveness. Existing approaches often rely on advanced controllers based on models of the insulin-glucose system. However, this system is highly complex, nonlinear, and subject to time-varying dynamics and inter-patient variability, which pose significant challenges for model accuracy and control design. Hence, data-driven and machine learning-based models are emerging as powerful alternatives. This paper presents a novel data-driven modeling approach that combines two components: a linear model and a machine learning-based model. This latter is computed with the CHoKI learning method, to capture the nonlinear deviations of the actual system from the linear model, enabling the combined model to better represent the insulinglucose system. This hybrid modeling approach offers improved prediction accuracy compared to previously proposed models in the literature. The improved model accuracy can lead to better controllers for the AP. The proposed approach is validated using the virtual patients of the FDA-accepted UVA/Padova simulator. The results outperform state-of-the-art models in prediction errors, demonstrating its potential as a step forward in AP control system design.

ABSTRACT. General anesthesia is a crucial component of invasive medical procedures. Our primary focus is the optimization of the simultaneous infusion of propofol and remifentanil during induction. In particular, we compute a feedforward infusion rate that minimizes the time for bringing the Bispectral Index (BIS) value below an assigned threshold, while ensuring that the BIS remains above an assigned minimum value. We consider a two-drug model, that takes into account the synergistic effect of the two drugs. Due to the expression of the Hill function, the constraint of maintaining the BIS in a given interval is non-convex. We propose a convex tightening to formulate the overall optimal control task as a convex optimization problem. Through simulation experiments, we show that this hybrid feedforward-feedback strategy quickly achieves the BIS target, while maintaining safe and clinically acceptable infusion rates.

ABSTRACT. Extensive clinical evidence supports the beneficial role of physical activity in delaying the progression of type-2 diabetes. However, current recommendations on the individual amount of exercise to be performed lack a quantitative assessment, thus remaining merely qualitative, and do not account for real-time prescriptions based on the individual current condition. This work is devoted to the proposal of an original model-based approach to the control of diabetes progression via physical activity, leveraging a sampled-data observer-based model predictive control framework. Starting from a control theoretical formulation of the benefits of the exercise, the proposed approach relies on two steps: 1) the design the control law on a compact widespread model of diabetes evolution; 2) the validation of the proposed control strategy, tested in silico by closing the loop on a population of virtual subjects simulated by a different, higher-dimensional model of the long-term effect of exercise on diabetes progression, also considering additional non-idealities, such as quantized measurements. Results prove the potential of our approach and can set the basis for the development of a general, quantitatively assessed protocol for diabetes control via physical activity.

ABSTRACT. We introduce the Compartmental Recurrent Neural Network COMP-RNN, a novel method for modeling glucose-insulin dynamics in type 1 diabetes mellitus. By integrating physiological knowledge and topology into recurrent neural networks, COMP-RNN significantly improves predictive accuracy and parameter efficiency compared to traditional models and standard recurrent neural networks. Simulated patient data validate its superior performance and demonstrate that COMP-RNN’s internal states reflect key physiological patterns, underscoring its potential to improve artificial pancreas systems.