ABSTRACT. Efficient thermal management is critical in modern Systems on Chip to balance performance and hardware safety. This study formally models a Power Management System that integrates Dynamic Power Scaling with thermal regulation. The system is represented as a Hybrid Automaton driven by continuous workload inputs, and is validated through OpenModelica simulations and time-abstract reachability analysis. The results demonstrate successful spike filtering and identify critical environmental temperature thresholds that lead to irreversible thermal deadlocks. This approach provides a rigorous methodology to systematically verify cyber-physical thermal constraints.

ABSTRACT. Understanding the mechanisms of diffusion of behaviors in social groups represents a major challenge in social science. The Linear Threshold Model (LTM) on graphs provides a useful framework, modelling each individual as a node that adopts a behavior if the influence from active neighbours exceeds a personal threshold. In this thesis we address the problem faced by a social planner aiming to induce complete adoption through propagation while minimizing intervention costs. We analyze two intervention strategies: targeting, which selects individuals for direct activation, and partial incentive, which reduces thresholds to ease diffusion. Although both problems are NP-hard on arbitrary graphs, we prove that specific topologies admit polynomial-time exact algorithms. We propose novel efficient algorithms for complete graphs with heterogeneous costs and complete multipartite graphs, and extend existing results on paths, trees, and cycles. For the general problem, we propose a randomized algorithm based on Simulated Annealing, which applies to arbitrary instances and returns high-quality solutions on large-scale real-world networks.

ABSTRACT. The thesis addresses the high-resolution digitization of forest ecosystems to automate forest inventory operations within the Forestry 4.0 paradigm. Since traditional aerial remote sensing suffers from canopy occlusions, this project develops autonomous exploration algorithms for the Boston Dynamics Spot quadruped robot to conduct systematic ground-level mapping. Overcoming the limitations of a rigid serpentine path, the solution implements a Frontier-Based Exploration Algorithm utilizing an occupancy grid. To ensure comprehensive area coverage, cells initially detected as blocked are dynamically re-evaluated from new approach angles during exploration. The system processes data from onboard depth cameras and integrates deeply with Spot’s GraphNav system to navigate unstructured environments. Significant outdoor perceptual challenges—such as uncut grass being falsely classified as an impassable hazard—were resolved by shifting to a distance-based grid with specific volumetric height thresholds. Real-world hardware experiments demonstrated that the algorithm dynamically calculates alternative paths to bypass unexpected obstacles, ensuring safe, collision- free navigation while successfully generating comprehensive 3D point clouds of the targeted woodland environment.

ABSTRACT. We introduce CALM-RL, a context-aware model-based reinforcement learning framework for transfer and adaptation across related physical systems. A latent context, inferred from transition data through an autoencoding structure, conditions both a shared Gaussian process dynamics model and the policy, which is learned by means of particle-based optimization. On a simulated cartpole benchmark, the method achieves 100% success on unseen parameter realizations and reduces the number of trials required to adapt to previously unseen dynamics shifts.

ABSTRACT. In this work, we propose P-MC-PILCO, a policy gradient Model-Based Reinforcement Learning (MBRL) algorithm that exploits Gaussian Processes (GPs) to learn the system dynamics. Compared to state-of-the-art GP-based MBRL algorithms, P-MC-PILCO combines GP approximations with a new sampling approach to substantially reduce computational and memory requirements of policy optimization, from quadratic to linear complexity, thus improving scalability to higher-dimensional systems. We benchmark the proposed algorithm both in simulated and actual environments of increasing complexity, including a high-frequency trajectory tracking control with an actual 7-DOF Panda robot. Results demonstrate that the proposed methods effectively preserve the benefits of exact GP models while significantly reducing computational and memory requirements.

ABSTRACT. This paper addresses autonomous task-servicing missions for UAVs navigating graph-structured maps, where agents must balance mission efficiency and energy safety. We propose a Dynamic Battery Threshold (DBT) reward integrated into a Double Deep Q-Network (DDQN) framework. Unlike static threshold approaches, DBT computes a topology-aware safety margin using Dijkstra shortest-path distances to the nearest charging station, enabling context-sensitive recharge decisions. The mission is formulated as a Markov Decision Process, where the agent selects actions — moving, picking up tools, executing fixes, or recharging — based on its current task and energy state. Monte-Carlo evaluation over 10,000 episodes demonstrates 100% mission success with zero battery failures, convergence to near-optimal Monte-Carlo Tree Search performance, and superior results over fixed-threshold greedy heuristics, at significantly lower computational cost.

ABSTRACT. Trajectory tracking for compliant robots is difficult because of nonlinear dynamics and modeling uncertainties. Iterative Learning Control (ILC) offers high tracking precision and preserves passive compliance but needs a new learning phase for each trajectory. Reinforcement Learning (RL) generalizes across trajectories yet suffers from high sample complexity. This work presents Reinforced Iterative Learning Control (RILC), which couples ILC and RL concurrently to achieve fast adaptation with guaranteed convergence. A theoretical proof shows that the RL component preserves ILC stability and, under specific conditions, reduces the asymptotic error bound. Validations on a robotic arm with two elastic joints for different trajectories and modeling mismatches validate the framework.

ABSTRACT. In this paper, we propose a preliminary study for the design of an adaptive data-driven control law for solving the model-reference problem of an unknown switched system. The proposed strategy aims to regulate the switched system such that the closed-loop behavior matches the desired dynamics defined by a reference model. By treating the switching signal as an external disturbance with sufficiently long dwell times between consecutive switching instants, we develop a model-reference design procedure that provides stability guarantees. Finally, simulation results demonstrate the effectiveness of the proposed approach.

ABSTRACT. In this article, we address two practical challenges of distributed learning in multiagent network systems, namely, personalization and resilience. Personalization is the need of heterogeneous agents to learn local models tailored to their own data and tasks, while still generalizing well; on the other hand, the learning process must be resilient to cyberattacks or anomalous training data to avoid disruption. Motivated by a conceptual affinity between these two requirements, we devise a distributed learning algorithm that combines distributed gradient descent and the Friedkin–Johnsen model of opinion dynamics to fulfill both of them. We quantify its convergence speed and the neighborhood that contains the final learned models, which can be easily controlled by tuning the algorithm parameters to enforce a more personalized/resilient behavior. We numerically showcase the effectiveness of our algorithm on synthetic and real-world distributed learning tasks, where it achieves high global accuracy both for personalized models and with malicious agents compared to standard strategies.

ABSTRACT. Potential-based control is a well-established paradigm for multi-agent coordination. However, real-world deployment is constrained by the reliance on analytical expressions, continuous-time dynamics, and a high sensitivity to manual gain-tuning as the number of agents scales. This abstract aims to addresses these challenges. First, we introduce neural-based potentials to capture complex hardware-dependent interactions (e.g., via cameras or antennas) directly from sensory data. This data-driven approach still preserves the structural properties required for stable distributed coordination with formal guarantees. Second, we propose a discrete-time formulation that removes the gain-tuning issue. To handle real physical constraints, it employs a distributed consensus protocol to compute a valid control gain, while enforcing heterogeneous actuator saturation.

ABSTRACT. The scenario approach is a data-driven paradigm for control design and decision-making under uncertainty. A decision is produced from observations and accompanied by probabilistic guarantees on its future risk of failure, namely the probability that it will fail to meet the required specifications in future situations.

The theory of the scenario approach has revealed a tight relationship between risk and complexity, where complexity is defined in unambiguous mathematical terms. This presentation focuses on a sequential, risk-averse use of scenario decisions: at each round, a decision is computed from sliding-window data, its data-dependent complexity is inspected, and the decision is applied only if this complexity is sufficiently small.

We ask how many failures are incurred on average by a sequential scheme that may reject decisions judged too complex. We summarize a recent result providing valid distribution-free bounds for this average number of failures. As we will see, naive common-sense evaluations may be misleading, and computing the correct bounds requires in-depth analysis.

The technical results and the conceptual message emerging from them are relevant to applications including the optimization of control actions over shifting windows, as in receding-horizon control, investments with recourse, and sequential prediction.

ABSTRACT. Autoimmune Atrophic Gastritis (AAG) is a chronic autoimmune disease characterized by progressive gastric mucosal damage and increased risk of complications, including type I gastric neuroendocrine tumors (NETs). However, its temporal evolution and the determinants of NET occurrence remain poorly characterized. This study aimed to model AAG progression, identify distinct temporal phenotypes, stratify patients according to different NET risk levels, and investigate the role of age in NET occurrence. We applied a graph-based framework combining Topological Data Analysis, through the Mapper algorithm, and a Minimum Spanning Tree filter to a multicentric Italian cohort. The temporal feature was included in the topology construction to preserve disease progression information, and patients were assigned to trajectories using Jaccard similarity. The framework identified five disease trajectories with significant clinical and histological differences. Survival analyses using both follow-up time and patient age confirmed the presence of heterogeneous NET risk profiles, distinguishing earlier and more aggressive progression patterns from delayed ones.

ABSTRACT. Type 1 Diabetes Mellitus management aims to maintain Blood Glucose (BG) within the euglycemic range (70-180 mg/dL), avoiding hypoglycemic (BG<70 mg/dL) and hyperglycemic (BG>180 mg/dL) episodes, both associated with life-threatening complications. This work addresses multi-horizon BG forecasting and early hypo/hyperglycemia alarm generation across 45, 60, 75, and 90-minute horizons, evaluated in-silico on adult, adolescent, and pediatric populations from the UVA/Padova simulator. From a 60-minute history of BG, insulin, and meal data, the future BG trajectory is predicted, and the probability of hypoglycemic, hyperglycemic, or simultaneous events is classified at each horizon, triggering an alarm whenever a tuned threshold is exceeded. Two candidate architectures, a compartmental LSTM and a physiologically-constrained XGBoost, are considered for both tasks. The F2-score was used as metric for early hypoglycemia detection, prioritizing sensitivity to limit missed events, and the F1-score for hyperglycemia. Results show that hypoglycemia detection is on average more robust with the XGBoost model across all cohorts, with the largest margin observed in children, while hyperglycemia detection is comparable between the two approaches. These findings highlight the feasibility of reliable multi-horizon alarm generation across heterogeneous populations, laying the groundwork for integrating early alarms with corrective actions to maintain patients' BG within a safe range.

ABSTRACT. We present a nonlinear dynamical model to describe the within-session running fatigue of an athlete in endurance sport. In the proposed model, fatigue plays the role of the state, and speed relative to a critical threshold is the input. All model parameters are fully interpretable, and we identify them by exploiting both physiological information and data collected through wearable devices. Validation compares the resulting race prediction performances with commercial solutions showing accurate race finishing time predictions.

ABSTRACT. This work addresses the problem of linear system realizations by mammillary and mammillary-like models, providing necessary and sufficient conditions under which a given transfer function can be represented in this form. The corresponding realization is shown to be unique, and a constructive procedure is given to compute its parameters. The framework is then extended to the structured pharmacokinetic/pharmacodynamic model of propofol, for which uniqueness is generally lost but all admissible positive realizations can still be systematically computed.

ABSTRACT. This paper presents a Hybrid Model Predictive Control (HMPC) strategy for safe and informative closed-loop identification of pH dynamics in microalgae raceway bioreactors.

Raceway reactors are widely used for large-scale microalgae cultivation but face significant control challenges: pH dynamics are affected by rapidly changing solar irradiance, transport delays, and integrative behavior. The key problem is balancing the need for informative experimental data with operational safety.

The proposed three-step workflow comprises: (1) obtaining a coarse first-order model from standard relay operation, (2) designing control-relevant multisine excitation embedded in an HMPC formulation, and (3) identifying an MPC-compatible model from closed-loop data.

The HMPC framework uniquely integrates persistent excitation design with safety constraints in a single optimization layer. Logic policies enforce dead-bands on modulation updates and minimum actuation spacing, yielding physically meaningful input trajectories. A supervisory safety policy prevents excitation when system outputs approach critical thresholds.

Experimental validation at IFAPA-CIESOL facility during May 2025 confirmed the approach generates informative identification data while maintaining pH within safe bounds. The resulting model was successfully deployed for MPC control with satisfactory regulation performance, demonstrating practical applicability for industrial microalgae cultivation systems.

ABSTRACT. Microbial consortia are pervasive in nature, inhabiting diverse environments from the human body to aquatic ecosystems and soil. In biotechnology, designing and controlling such communities is key to enabling sustainable production of biofuels, pharmaceuticals, and other valuable compounds. Co-cultures comprising multiple microbial strains can outperform monocultures by leveraging division-of-labor strategies that reduce metabolic burden, improve pathway modularity, and enable processing of complex substrates. Despite these advantages, industrial deployment of co-cultures remains limited, predominantly due to the lack of robust and scalable architectures for ensuring coexistence and maintaining stable consortium compositions. Without adequate control mechanisms, competitive exclusion can lead to community collapse. Control strategies are therefore essential to regulate growth and ensure long-term stability and productivity.

ABSTRACT. Bacterial cellulose (BC) functionalized with ionic liquids and conductive polymer electrodes is a promising smart organic material for sustainable mechanoelectric sensing. Its response involves coupled mechanical deformation, ionic redistribution, electrical polarization, and material-dependent dynamics. This work summarizes three complementary modeling levels for BC-based smart sensors: a white-box multiphysics model for interpreting mechanoelectric transduction, a grey-box fractional-order model for compact electrical characterization as fractional order element, and a black-box data-driven model for prediction and transferability across material configurations.

ABSTRACT. Optimization of therapy rotation in individuals with mutating viral infections is critical. This paper applies a state-dependent switching framework for scheduling therapeutic rotations under Dwell-Time (DT) constraints, in a chronic infection scenario formalized as an autonomous switched positive linear system. The presence of multi-drug resistant genotypes renders viral suppression unachievable, shifting the control paradigm toward maximizing the postponement of viral escape. Building on a Lyapunov-Metzler framework proposed for switched linear systems with DT constraints, a state-feedback switching policy is synthesized, guaranteeing a minimum DT before each switch and selecting the next regimen via a direct comparison of Lyapunov energies. The efficacy of the proposed policy is validated through a comparative analysis against clinical protocols, including Virological-Failure-driven and Periodic switching, and advanced control approaches, including Model Predictive Control and the Best Sequence benchmark, over a 200-day simulation. Results show that none of the compared strategies achieves eradication, consistent with the non-stabilizable nature of the chronic scenario; however, the proposed policy matches the performance of the advanced scheduling references, while outperforming the clinical baselines, at lower online computational cost. These findings indicate that state-dependent switching under DT constraints offers an efficient alternative to optimization-based scheduling for postponing viral escape in chronic infections.

ABSTRACT. La tecnologia non rappresenta più, da sola, il principale fattore competitivo della manifattura. Il vantaggio competitivo nasce oggi dalla capacità di governare la convergenza tra tecnologie, dati, persone e processi, trasformando la fabbrica in un ecosistema cognitivo in cui la qualità delle decisioni diventa centrale. L’integrazione tra intelligenza artificiale, robotica, sistemi cyber-fisici, cloud e nuovi quadri normativi europei sta ridefinendo modelli operativi, responsabilità organizzative e competenze richieste alle imprese. In questo scenario, il Comitato Scientifico di SPS Italia, attraverso il Position Paper 2026, propone una lettura condivisa delle principali trasformazioni in atto. Il documento non offre soluzioni prescrittive, ma criteri interpretativi utili a supportare decisioni strategiche più consapevoli, valorizzando il confronto tra imprese, università ed esperti e trasformandolo in conoscenza accessibile e trasferibile per l’intero sistema manifatturiero. L’articolo presenta la visione che ispira il Position Paper e si sviluppa lungo quattro direttrici fondamentali: l’evoluzione del quadro normativo, l’intelligenza decisionale, la convergenza tra robotica e intelligenza artificiale e l’integrazione delle architetture digitali dal campo al cloud. La tesi centrale è che la sfida non consista più nell’adottare nuove tecnologie, ma nel governarne l’integrazione all’interno di sistemi coerenti, affidabili e sostenibili, favorendo il passaggio dalla fabbrica automatizzata alla fabbrica cognitiva.

ABSTRACT. The XPlanar system from Beckhoff provides free 2D motion with up to six degrees of freedom, combining product transport and processing in a single platform. The movers levitate and travel over modular tiles, following programmable paths at speeds of up to 3 m/s with a repeat positioning accuracy of ≤ 5 µm. Each product can take an individual route, enabling economical lot size 1 production. The system requires only tiles, movers, a Beckhoff industrial PC with TwinCAT, and power/EtherCAT G connections, minimizing installation effort and simplifying integration with other machine components. The tiles integrate power electronics, position detection, and communication, and can be protected for hygienic environments. The XPlanar software, fully integrated into TwinCAT, manages graphic configuration, real-time monitoring, track planning, and precise position control, supporting complex 6D movements and collision avoidance. This compact, centralized architecture makes XPlanar ideal for future machine concepts, offering high flexibility and easy software updates for new functionalities.

ABSTRACT. As part of a collaboration between Krones’ R&D department and the Italian – based Krones Innovation Hub, an investigation of planar motor technology started at the beginning of 2025 and is still active today. The primary objective of the investigation is to evaluate the technology itself as well as its applicability to the bottling industry. The present research is conducted on APS9001, a starter kit from Beckhoff composed of twelve 240 x 240 mm stators and four 155 x 155 movers.

ABSTRACT. This paper presents CoboREEL, a collaborative robotic solution for automated label-reel handling in packaging and labelling systems. The application addresses a recurrent industrial problem: reel replacement is simple but repetitive, ergonomically unfavourable, and disruptive when performed manually during production. The proposed architecture combines a collaborative robot, dedicated reel tooling, autosplicer integration, barcode traceability, and PLC/HMI supervision in a compact layout suitable for existing machines. The paper discusses the design rationale, system architecture, and expected industrial impact, emphasizing why collaborative robotics is particularly relevant where automation must coexist with operator access, format changes, and production flexibility.

ABSTRACT. Three Decades of Autonomous Driving: From Pioneering Innovation to Scalable Solutions Three Decades of Autonomous Driving: From Pioneering Innovation to Scalable Solutions

ABSTRACT. In recent years, the deployment of social robots in human-populated environments has introduced critical challenges for efficient and safe trajectory planning. Motion planners must balance multiple objectives, such as target-reaching, effort minimization, and collision avoidance, while continuously predicting the movements of nearby pedestrians. Traditionally, collision avoidance is formulated as a hard constraint within optimization frameworks such as Model Predictive Control (MPC). However, this can yield non-smooth, action-reaction behaviors over short horizons or overly conservative maneuvers over multi-step horizons. Conversely, probabilistic chance-constrained approaches account for trajectory uncertainties but remain highly sensitive to manually tuned safety thresholds, risking collisions if improperly calibrated.

A gap in the existing literature is the lack of mechanisms to dynamically modulate the conservativeness of a robot based on the perceived contextual awareness, specifically, the level of attention of surrounding humans. This paper addresses this gap by proposing a human attention-aware motion planning framework. Rather than relying on rigid constraints or inferred trust models, this approach introduces an adaptive Control Barrier Function (CBF) within a receding-horizon MPC. The framework continuously estimates pedestrian attentiveness and optimally regulates constraint relaxation, enabling context-adaptive navigation that behaves cautiously around inattentive pedestrians while allowing smoother, less conservative maneuvers around attentive ones.

ABSTRACT. Opacity is a fundamental confidentiality property in discrete-event systems, ensuring that an external, possibly malicious, observer cannot determine whether the system is in a secret state. A secret state generally represents a system vulnerability condition to be preserved. A suitable approach consists in opacity enforcement, mainly via supervisory control. While extensively studied under passive observation, opacity enforcement under active adversaries remains less explored, due to its inherent implementation complexity.

This paper investigates current-state opacity for deterministic DES subject to actuator-enablement (AE) attacks, where an adversary may force disabled controllable events to occur. We propose a unified framework that combines supervisory opacity enforcement with a mitigation mechanism for defendable events, based on an observer-driven characterization of admissible control decisions.

The approach is preliminarily validated on a water distribution laboratory testbed.

ABSTRACT. Real-world assistive robotics demands both high level semantic understanding for human-robot interaction and low-level deterministic control for safe physical execution. Tradi tional architectures often rely on cloud-dependent models or rigid pre-programmed state machines, raising concerns about privacy, latency, and environmental adaptability. This paper proposes a novel, generalized layered cognitive-control architecture that decouples heavy, non-deterministic cognitive reasoning driven by local Visual Language Models (VLMs) from time-critical, deterministic physical execution loops. Our architecture consists of: (1) a Cognitive Layer that evaluates scene context and generates behavior sequences using local VLM inference; (2) a Physical Execution Layer running standard robotic navigation and manipulation stack primitives; and (3) a high-throughput, multi-threaded Communication Bridge that coordinates telemetry, image compression, and action dispatch. We also present a distance-based spatial semantic memory clustering algorithm to group and ground environmental objects without hallucinations. Evaluation of this layered model demonstrates that local VLM reasoning can be seamlessly integrated into real-world assistive applications, ensuring low latency, data privacy, and adaptive mission execution.

ABSTRACT. Eco-evolutionary models for biological systems couple population dynamics with phenotypic evolution, making intervention design inherently anticipatory: each input steers the ecological state while reshaping the system’s traits, including those that determine the input’s future effectiveness. In cancer, for instance, sustained therapy selects for traits that reduce drug efficacy. Here, we summarize our eco-evolutionary modeling backbone and its control applications, then discuss recent results on hybrid control of targeted therapy, where continuous population-trait dynamics couple with discrete transitions from reversible tolerance to irreversible resistance. Within this framework, supervised hybrid model predictive control steers tumor dynamics toward eradication while counteracting therapy-induced loss of control authority, a failure mode that arises whenever an input selects for its own attenuation.

ABSTRACT. Modern industrial plants increasingly rely on fleets of autonomous vehicles operating in complex and highly dynamic environments, where efficient coordination is essential to maximize productivity, improve operational efficiency, and support reliable autonomous operation. This talk presents recent research results developed in collaboration with E80 Group S.p.A. and Gruppo TecnoFerrari S.p.A. con socio unico, focusing on scalable path planning and traffic management algorithms for industrial AGV systems. The presentation begins with a decentralized predictive coordination framework for multi-AGV systems operating in realistic, non-standardized warehouse environments, enabling congestion-aware navigation and proactive conflict resolution. It then introduces a traffic management architecture based on Lifelong Multi-Agent Path Finding, designed for large and heterogeneous AGV fleets operating in high-density industrial environments, achieving improved throughput and efficient fleet coordination. Finally, the talk briefly outlines recent advances in hybrid planning architectures, integrating structured roadmaps and dynamically changing environments through hierarchical multi-objective optimization to support adaptive and scalable autonomous navigation. By bringing together predictive path planning, intelligent traffic management, and hybrid navigation strategies, the presented research outlines a unified framework for the next generation of autonomous industrial transportation systems, demonstrating how close collaboration between academia and industry can drive innovative solutions for real-world applications.

ABSTRACT. Cyber-physical-human systems place people inside the control loop. This changes any control problem in a way that standard control design does not account for. A human is indeed an independent system with objectives of its own, whose decision rule we can observe but cannot design. This talk revisits control as an action-selection problem and asks what happens when only one of the decision maps in the loop belongs to the designer. The key issue is that accurate individual models do not compose. A collection of well-fitted response maps may admit no collective equilibrium at all, since every classical existence result rests on convexity or on a Lipschitz bound that unconstrained learning does not provide. Two complementary directions are discussed. The first constrains the learning itself, so that structural properties hold by construction and equilibrium guarantees are recovered at the price of expressiveness. The second weakens the conclusion, showing that when physical coupling destroys convexity, local guarantees survive under some suitable regularity conditions.

ABSTRACT. The paper is aimed to show the advantages and the constraints in the use of a fully homomorphic encryption system in remote control scenarios. The main advantage is that in this encryption scheme the controller performs the computations in the encrypted domain, i.e. without knowing the meaning of the processed data. On the other hand, one of the main drawbacks is that non-polynimial operations cannot be performed directly and must be obtained through polynomial approximation.

ABSTRACT. Ensuring consensus while preventing the leakage of sensitive information to external observers is a fundamental challenge in networked multi-agent systems. This submission presents a communication-layer masking strategy to protect exchanged state information from passive external eavesdroppers while preserving the nominal consensus dynamics among legitimate agents. The mechanism applies reversible Lie-group transformations on each directed edge through private keys shared only by endpoint nodes. The analysis shows that intercepted traffic is compatible with an uncountable family of plaintext trajectories, and that uniquely inferable information is limited to invariants of the selected group action. With time-varying i.i.d. Haar-distributed keys, ciphertexts produced by plaintexts in the same orbit are indistinguishable in distribution at each time instant. At steady state, the consensus value remains identifiable only up to orbit constraints induced by intercepted links.

ABSTRACT. Control Barrier Functions (CBFs) have emerged as a powerful tool in the design of safety-critical controllers for nonlinear systems. In modern applications, complex systems often involve the feedback interconnection of subsystems evolving at different timescales, e.g., two parts from different physical domains (e.g., the electrical and mechanical parts of robotic systems) or a physical plant and an (optimization or control) algorithm. In these scenarios, safety constraints often involve only a portion of the overall system. Inspired by singular perturbations for stability analysis, we develop a formal procedure to lift a safety certificate designed on a reduced-order model to the overall feedback-interconnected system. Specifically, we show that under a sufficient timescale separation between slow and fast dynamics, a composite CBF can be designed to certify the forward invariance of the safe set for the interconnected system. As a result, the online safety filter only needs to be solved for the lower-dimensional, reduced-order model. We numerically test the proposed approach on: (i) a robotic arm with joint motor dynamics, and (ii) a physical plant driven by an optimization algorithm.

ABSTRACT. This paper proposes a robust data-driven policyupdate method for discrete-time linear quadratic regulation of unknown linear systems affected by bounded data disturbances. Input–state measurements define the entire set of dynamics compatible with the data, without identifying a nominal model. A left-inverse representation eliminates the unknown system matrices from the closed-loop dynamics, while Petersen’s lemma yields linear matrix inequality (LMI) conditions certifying a common quadratic Lyapunov function over the data-consistent model set. Consecutive controllers are linked by a proximal redesign term, and every candidate is independently certified before acceptance. Tests on ten unstable systems show a 100% final certification rate and a 30.0% median cost reduction, while robust closed-loop stability remains the formal guarantee.

ABSTRACT. Safe autonomous navigation in shared environments requires the ability to anticipate and react to the latent behaviors of surrounding robots. In this paper, we propose SOWL-MPC, a safe learning-based predictive control strategy for a novel scenario, which we name ego-world robotic framework. In this setting, the control policy of the world robot is unknown and the ego exploits data to learn it and perform safe maneuvers. The proposed architecture combines an online learning mechanism based on Sparse Variational Gaussian Processes (SVGPs) with a receding-horizon control scheme. Relying solely on noisy state measurements, our approach infers a posterior distribution over the latent world policy, which is updated on streaming data via Online Variational Conditioning (OVC). The learned policy is propagated through the nonlinear world dynamics using an approximate moment propagation scheme, and fed to an uncertainty-aware Model Predictive Control (MPC), thus enabling safe maneuvering of the ego robot. The real-time feasibility and safety guarantees of SOWL-MPC are demonstrated through extensive Monte Carlo virtual experiments in ROS 2, and validated on real-world robotic hardware in an indoor arena.