ABSTRACT. Assistive robots such as smart wheelchairs must do more than act intelligently in their environment: they must also interpret the commands of users who operate them through very limited interfaces. Head or chin joysticks, switches and brain-machine interfaces deliver sparse, quantized, delayed or even wrong commands. This extended abstract summarizes two recent contributions of the IAS-Lab of the University of Padova that treat the user command as uncertain evidence of intent rather than as ground truth. The first is a multi-layer navigation system that combines footprint-aware reactive avoidance, model predictive control and open-vocabulary perception to infer targets when commands are faulty. The second is GUARD, a planner-independent, geometry-aware safety layer that arbitrates user commands in real time. Both were validated on a real wheelchair with healthy participants under artificially degraded input, showing safe navigation, fewer corrective inputs on average (MLNS) and higher usability under degraded input (GUARD).

ABSTRACT. Multi-hypothesis cooperative localisation systems that fuse inertial and ranging measurements must repeatedly propagate high-rate IMU data across many candidate state hypotheses, which is computationally and communication-wise expensive. We present a preintegration framework that decouples inertial motion and covariance propagation from the initial state, with an exact decoupling of the accelerometer bias. The resulting fixed-size representation can be applied to an arbitrary number of state hypotheses at a cost independent of the integration window, making it well suited to particle-based estimation of multi-agent systems. We outline the approach, its use in a multi-hypothesis filtering, and the hardware platform on which an experimental campaign is currently underway to evaluate the framework.

ABSTRACT. Robot-assisted rehabilitation systems provide quantitative measures of task performance, although comparable outcomes may be achieved through different motor, compensatory, and visual strategies. This work proposes a proof-of-concept multimodal framework integrating an end-effector robotic device for upper limb rehabilitation, wearable inertial sensors and an eye-tracking system with the aim of combining the robotic parameters, the upper-body kinematics and gaze information. These complementary domains are integrated into a novel evaluation metric accounting for upper limb motor performance, possible postural compensations, and visual behaviour. By characterizing not only the task outcome but also the strategies adopted during its execution, the proposed approach may provide a more comprehensive interpretation of patient–robot interaction and the future development of patient-aware and adaptive rehabilitation systems.

ABSTRACT. FRONTIER is a novel soft exoskeleton for coordinated upper-limb assistance in people with neurological motor impairments. The platform combines soft pneumatic modules for shoulder and elbow assistance with a cable-driven hand exoskeleton. The modular hardware is combined with dedicated low-level controllers and a coordinated supervisory architecture. Pressure sensing supports local regulation of the proximal actuators, whereas inertial sensing provides limb kinematics for active assistance and event-based sequencing. A task-oriented strategy organizes proximal-to-distal recruitment and uses completion of a sensed forearm supination phase to trigger hand closure, allowing distal assistance to follow the user timing. Configurable operating modes, a clinical operator interface, data logging, and layered safety functions complete the architecture. The resulting platform provides an expandable technological basis for coordinated multi-joint assistance and subsequent clinical translation.

ABSTRACT. Wrist-worn wearables are the dominant paradigm for step counting, but their performance degrades noticeably when used by people with motor impairments. We instrument a standard cane with a low-cost sensing framework to address this and, at the same time, to lay the foundation for a self-supervised gait-quality assessment in which a convolutional autoencoder trained only on normal walking uses reconstruction error as an anomaly score. A downward-facing time-of-flight sensor measures the distance of the cane tip from the ground and provides the reference against which step counting is validated, reaching 95.2% accuracy, while on the same recordings the anomaly detector reaches 99.3% AUPRC, also ordering the types of gait problems by how strongly they deform the signal.

ABSTRACT. High-resolution, real-time tactile sensing is essential for advanced robotic tasks such as grasping and manipulation of delicate or irregularly shaped objects. Conventional taxel-based tactile sensors suffer from limitations in spatial resolution, due to taxel size, wiring, and cross-talks, whereas vision-based tactile sensors typically incur computational latency. In this work, we present a novel type of vision-based tactile sensor, which exploits a mechanochromic effect to directly encode mechanical strain into spatially resolved structural colors. The sensor consists of a stretchable mechanochromic Bragg reflector sandwiched between a black outer elastomer backing and a transparent elastomeric layer. This design enables high-resolution tactile mapping without deep learning data enhancement or computational latency.

ABSTRACT. Continuous and objective monitoring of upper limb recovery after stroke could complement conventional clinical assessments and support more personalized rehabilitation pathways. Robotic rehabilitation systems, inertial sensors, and wearable devices provide heterogeneous measurements collected during supervised rehabilitation, structured motor assessments, and daily life, but the extent to which they capture longitudinal changes in functional motor performance remains unclear. This work proposes a multimodal sensing framework, not yet applied to experimental data, to investigate whether longitudinal changes in sensor-derived measures are associated with changes in upper limb functional performance, assessed through the Box and Block Test (B&B) as a proxy of clinical scales. The three sensing modalities are assigned three distinct roles: robot-derived measures describe performance and autonomy during supervised therapy; inertial measures are intended to quantify movement quality and inter-limb asymmetry, including outside the hospital; measures from wearable sensors describe patient compliance and general well-being. The integration of such measurements into the proposed multimodal sensing framework would enable home monitoring and support data-driven personalised precision rehabilitation treatments.

ABSTRACT. Conventional computerized cognitive assessments typically rely on discrete responses, providing limited information about how a response develops over time. This pilot study investigates whether directional forces applied to a robotic end-effector can reveal cognitive interference while providing a continuous description of response execution. Ten healthy participants responded to spatially congruent and incongruent visual stimuli by exerting forces along four Cartesian directions. Force-onset reaction times were significantly longer in incongruent than congruent trials (499.40 vs. 442.08 ms), demonstrating that interference from task-irrelevant spatial information can be detected through physical human–robot interaction. Force trajectories additionally retained information about response direction, magnitude, temporal development, and across-trial variability. These preliminary findings support force-based interaction as a promising approach for moving beyond discrete behavioral outcomes toward richer, process-oriented robot-mediated cognitive assessment.

ABSTRACT. Physical exercises can elicit substantially different physiological responses across individuals under the same external stimulus, motivating adaptive strategies able to account for the user’s internal response during training. This work presents a closed-loop approach for adapting physical exercise plan execution to target workload dose in robotic physical coaching scenario. Physiological monitoring is used to estimate workload and dynamically regulate exercises and recovery duration in order to deliver a prescribed workload dose. The approach preserves the planned exercise sequence and nominal intensity profile while adapting execution time according to the individual physiological response. A preliminary evaluation with five healthy participants showed that our approach entails suitable subject-specific temporal adaptations with respect to fixed-time execution, supporting the feasibility of physiology-driven workload dose regulation during training. Although initial, this result seems promising and worth to be further investigated and assessed.

ABSTRACT. Additive manufacturing of Soft Pneumatic Actuators (SPAs) via Fused Deposition Modeling (FDM) offers great flexibility for custom robotics, allowing the fabrication of complex geometries with a wide pool of different materials. This paper presents the preliminary design, manufacturing, and experimental evaluation of a Thermoplastic Polyurethane (TPU) three- finger soft gripper. Each finger consists of an antagonistic pair of air-bellows actuators, interconnected by S-shaped hinges and obtained from the repetition of 3D closed-wall re-entrant auxetic S-hinged unit cells. By maintaining a ≈1.40 mm wall thickness, complete pneumatic airtightness is achieved without post- processing. A differential pressure control scheme is introduced by mimicking the antagonistic actuation of human biomechanics: a pre-set co-contraction pressure P0 regulates the mechanical stiffness of the fingers, while differential pressure inputs ±∆P, relative to the baseline P0, drive fast and controlled bidirectional bending, thereby improving bending, gripping performance and conformability of the gripper to delicate objects.

ABSTRACT. This work presents a switched-systems framework for dynamic quadrupedal jumping based on a planar single-rigid-body model with explicit foot placement. By jointly optimizing the base trajectory and phase durations over a predefined contact sequence, the proposed method continuously adapts contact timings, avoiding the combinatorial complexity of discrete mode-selection variables. To mitigate systematic modeling discrepancies and hardware tracking errors, the optimized trajectory is executed using Iterative Learning Control (ILC) coupled with PD feedback. Hardware experiments on the Solo-12 and Otto quadruped platforms demonstrate dynamic step-jumping maneuvers, with consistent reduction in joint-space tracking errors over successive trials.

ABSTRACT. Series Elastic Actuators (SEA) give robots the adaptability to store and exploit energy and to absorb impacts from the environment, but their compliant dynamics hinder accurate trajectory tracking. Iterative Learning Control (ILC) achieves high precision on a single reference while preserving passivity, yet it must be retrained whenever the trajectory changes. Reinforcement Learning (RL) instead generalizes to new references, although it demands many samples and offers coarse precision. Reinforced ILC (RILC) combines the strengths of both approaches. ILC synthesizes accurate feedforward commands and an RL policy extends them to new trajectories. We evaluate RILC on a robotic arm with two elastic joints. RILC reduces the initial tracking error by a large margin and reaches convergence within a few iterations, with an asymptotic error tighter than standalone ILC.

ABSTRACT. A prosthetic limb must establish a stable and comfortable biomechanical coupling with the user’s residual limb, while accommodating individual anatomical differences and, increasingly, sensing technologies capable of detecting user intentions. In this work, a digital workflow is proposed for the design and additive manufacturing of customized silicone liners, i.e., soft socks worn between the residual limb and the prosthetic socket to establish the physical interface of limb prostheses. Specifically, patient-specific liner geometries with dedicated accommodations for electromyography (EMG) electrodes are developed. Residual-limb morphology is acquired by 3D scanning and converted into a personalized liner geometry, in which local thickness variations and electrode accommodations are directly incorporated during the digital design phase. Upper- and lower-limb liners are fabricated through liquid additive manufacturing of silicone material. Preliminary prototyping confirms the feasibility of the proposed approach, while future work focuses on the systematic assessment of fit, usability, and sensing performance.

ABSTRACT. Robotic cockpit testing requires localization of small control elements without permanently modifying the environment, such as placing fiducial markers. This work presents a coarse-to-fine perception framework for markerless robotic cockpit testing. The approach combines global 6-DoF panel pose estimation with local image-based visual servoing, so that manipulation-level accuracy is not required from the global estimator alone. A sparse RGB-D panel representation, initialized from manually selected keypoints, is used to automatically generate a real annotated dataset for PVNet fine-tuning without requiring a complete CAD model. Experiments show that the pose estimator provides a usable coarse localization, while IBVS has been independently validated with fiducial features as a surrogate for the final markerless alignment stage. Full PVNet--IBVS integration and validation on real cockpit panels are ongoing.

ABSTRACT. Vision–Language–Action (VLA) models can be adapted from teleoperated demonstrations and refined with human-in-the-loop (HIL) interventions, but the interface used to collect both datasets is decisive. This paper compares GELLO, a low-cost joint-space leader–follower kit, with a Meta Quest 3 VR headset on a UR5e cell with four lab manipulation tasks. Separate policies were fine-tuned from 100 expert demonstrations per interface. In a mixed-design study (N = 20), GELLO was faster, more successful, more usable and less demanding for direct teleoperation, whereas VR was preferred for takeover during VLA rollout. The results show an interaction between interface and human role, not a universally better device.

ABSTRACT. Agriculture is facing a significant crisis due to a combination of factors like climate change and workforce shortages. Robotic platforms are potential solutions to these problems, and perception subsystems are crucial for effective and fast phenotyping. We propose a multi-sensor perception framework for detecting individual lettuce plants and estimating phenotypic traits, including NDVI and geometric indicators.

ABSTRACT. Surgical skill assessment still relies largely on subjective evaluations, limiting the objective characterization of technical performance and the development of standardized training protocols. We present a quantitative analysis of simulated surgical gestures by integrating Motion Capture and surface electromyography (EMG) during suturing and incision tasks performed by participants with different levels of surgical experience. Kinematic and neuromuscular features were extracted to investigate movement efficiency, joint coordination, muscle activation and patterns of motor organization. The results show differences in kinematic and muscle activation parameters across experience groups, with task-dependent trends and partial overlap. These findings highlight the potential of multimodal biomechanical analysis to support objective surgical skill characterization and provide a methodological basis for future feedback systems and technology-assisted surgical training.

ABSTRACT. The ability to interact directly with the environment can extend the capabilities of quadruped robots in exploration activities, such as search-and-rescue missions. This work presents a teleoperated loco-pedipulation framework that reassigns any locomotion leg to non-prehensile manipulation without additional hardware. Gamepad inputs select locomotion gaits, the active leg, and mode-dependent base or Cartesian foot references. A leg mode encoding role and contact state explicitly schedules locomotion-pedipulation transitions within a Model Predictive Control formulation, coupled with hierarchical whole-body control and state estimation. The complete architecture was evaluated in simulation, while experiments on a Unitree Go1 tested teleoperated transitions and free-space Cartesian positioning with each leg. The results confirm the correct generation of contact-mode and leg-role sequences. In a representative real-robot experiment, free-motion foot tracking achieved an RMSE of 0.001 meters relative to the position reconstructed through forward kinematics.

ABSTRACT. This proof-of-concept study investigated the feasibility of wearable EEG monitoring during robotic upper-limb training and explored EEG changes associated with cognitive load and fatigue. Healthy participants performed three tasks using the MOTORE® robotic system while EEG was recorded with the four-channel Muse 2 device. Preliminary data from two participants showed that EEG acquisition during robotic training was technically feasible, although minor adjustments were required to ensure stable sensor contact. Post-task recordings showed increased delta and reduced alpha power, while no consistent increase in theta activity was observed. Subjective mental fatigue was low, suggesting that the current motor paradigm may not provide sufficient cognitive demand to elicit robust fatigue-related EEG changes. These proof-of-concept findings support the integration of wearable EEG into robotic upper-limb training and provide a basis for refining the experimental paradigm before its application in individuals with stroke.

ABSTRACT. Vision-language-action (VLA) models are typically pretrained on broad, diverse datasets and subsequently fine-tuned on a comparatively small dataset tailored to a specific target robot. Such adaptation recipe is now widely established, achieving its most mature results on fixed-base arms due to their static base dynamics and consistent reference frame. Mounting the same arm on a quadruped extends the workspace to the entire floor, but fundamentally changes the embodiment: rather than controlling the arm in isolation, the policy must jointly coordinate base navigation, body posture, and end-effector motion. In this work, we study the practical demands of this embodiment change across data collection, model adaptation, and inference design to achieve closed-loop loco-manipulation on a Unitree Go2 with an AgileX Piper arm. We report the choices that made this real-robot deployment feasible, namely model variant, action horizon, and demonstration conventions, together with a study tailored to a pick-and-place task, separating data coverage from model capacity. We show that expanding demonstration coverage reduces held-out end-effector error by 30\%, and incorporating secondary skills (such as pushing rather than grasping) introduces new behavioral modes without degrading baseline performance. Finally, closed-loop real-world experiments show how dataset composition shapes success and failure, validating the pipeline for more complex loco-manipulation.

ABSTRACT. This work presents an innovative control framework integrating robot hand guidance and high-speed burr navigation to improve precision and efficiency in spinal surgery. A task space control algorithm is implemented for accurate positioning of the tool center point (TCP) within the surgical navigation frame. Our approach is based on exploiting the redundancies of the robot, ensuring constrained motion within the safe surgical workspace. Experimental validation, implemented on a KUKA LBR Med robot, shows improved accuracy and reduced surgical procedure time.

ABSTRACT. Affinity (similarity) bias, where perceived resemblance between an evaluator and a candidate skews evaluation, is a known but hard-to-detect factor in recruitment interviews. Investigating this in an AI-assisted setting requires a way to estimate human-perceived facial similarity, which conventional face-recognition models do not provide, as they are designed to determine identity rather than perceived resemblance between different individuals. This work fine-tunes FaceNet512 on human-ranked face-comparison data to better capture perceptual similarity. Progressive fine-tuning improved test accuracy from a baseline of 57.84% (unmodified FaceNet512) to 66.19%, providing initial proof of concept. This model is intended as the first component of a human-in-the-loop framework for bias-aware interview support.

ABSTRACT. This work presents a circular electromechanical reservoir comprising a rigid enclosure, a flexible fluid pouch and a motor-driven rotating lever. Lever rotation varies the space available to the pouch, accommodating filling and progressively compressing it during active discharge. The reservoir geometry is designed for a chamber capacity of approximately 500 mL, with provisions for integrating sensing and control within the reservoir structure. Preliminary benchtop experiments demonstrate fluid discharge within the specified time, pressure and residual-volume targets. A separate force-sensing demonstration illustrates sensor-dependent modulation of motor commands, providing a basis for adaptive discharge control. These findings support further development through pouch optimization, improved interaction sensing and component integration, towards compact fluid-management systems with potential wearable and implantable applications.

ABSTRACT. AI-enabled remote patient monitoring increasingly combines sensors, longitudinal records, and decision support. Most systems primarily deliver monitoring, alerts, or human-facing recommendations. Agentic robotics can extend this pipeline into situated assistance by acquiring missing information, selecting bounded actions, and revising behavior from feedback. In healthcare, closing the loop from observation to physical action requires evidence provenance, separation between reported state and permission to act, and retained human clinical authority. We present a work-in-progress architecture in which shared contextual memory connects patient, caregiver, clinician, sensor, and robot events to a goal-directed perception-decision-action loop. Given a clinician-authorized session goal, a constrained planner selects among evidence retrieval, contextual questions, repeated measurements, authorized exercise explanation or demonstration, deferral, termination, and escalation. A deterministic governance gate checks identity, evidence, authorization, consent, sensor quality, and predefined stop conditions before symbolic execution. Observations, rejections, and execution outcomes update memory to support context reconstruction and replanning. Implemented components comprise a clinical retrieval pipeline, typed multi-actor event and action schemas, a constrained planner interface, deterministic validation, an evidence-bearing action ledger, and a partially observable synthetic environment. Ongoing work integrates these components and evaluates their mechanism-level contributions. The architecture defines agentic robotics as governed, feedback-driven adaptation from longitudinal monitoring to bounded assistive action under human clinical authority.

ABSTRACT. Robotic interaction with soft biological materials requires reliable contact initiation, especially when surface position is unknown. We present a multisensory event-driven landing strategy combining force and ultrasound sensing. Experiments on agarose phantoms and fixed human colonic tissue showed ultrasound-dominated triggering in agarose and mixed ultrasound/load-cell triggering in tissue. These results support multisensory event detection for position-independent contact initiation on heterogeneous soft targets.

ABSTRACT. Preoperative surgical planning is a crucial step for defining patient-specific surgical strategies and anticipating possible intraoperative challenges, especially for complex interventions. Although planning traditionally relies on two-dimensional medical imaging, patient-specific three-dimensional models are increasingly used due to their advantages over diagnostic imaging alone. 3D models can be visualized through screen-based systems, immersive virtual reality headsets, or 3D-printed to obtain a physical representation. However, there is still limited evidence on which visualization modality is most appropriate for different surgical contexts. As the effectiveness of a visualization modality may depend on the anatomical district and the surgical task, we propose a patient-specific 3D visualization framework for preoperative planning. The framework includes clinical case selection, image segmentation or 3D surface acquisition, model preparation, visualization through physical or virtual modalities, and experimental assessment. The ultimate goal is to establish standardized, evidence-based guidelines for the effective use of 3D models in preoperative planning.

ABSTRACT. Anaerobic digestion converts organic residues into methane-rich biogas, but full-scale operation remains largely empirical because the process is slow, nonlinear, partially observable, and sensitive to feedstock variability. This paper presents the ongoing AMOBio project, which develops a decision framework for safe, data-efficient optimization of anaerobic digestion plants. The framework combines Gaussian process surrogates with constrained, cost-aware Bayesian optimization to select informative operating points under uncertainty. Transfer learning uses laboratory data to initialize plant-level models and reduce the number of full-scale trials. A coordinated agentic layer manages heterogeneous data, checks candidate set points against biological and operational constraints, schedules laboratory analyses, and communicates recommendations while preserving operator approval. We describe the architecture, problem formulation, validation protocol for two industrial use cases, and predefined evaluation criteria.

ABSTRACT. Motor rehabilitation at home requires assistive systems that can support patient mobility while providing objective and repeatable performance measurements. This work presents a ROS 2-based simulation framework for a smart walker operating in a domestic environment. The framework integrates autonomous navigation, sensor-fusion-based localization, and parametric rehabilitation paths for the quantitative assessment of path-tracking performance. A digital representation of a domestic rehabilitation room was developed in Gazebo, while the navigation stack was configured using Nav2. Different global planners and localization strategies were evaluated under simulated operational conditions. The results show that sensor fusion improves pose-estimation accuracy and temporal continuity compared with particle-filter localization, supporting more reliable trajectory recording during training sessions. The proposed framework provides a preliminary basis for the future integration and validation of an assistive walker in home-rehabilitation scenarios.

ABSTRACT. Assisting post-stroke neurorehabilitation with a mobile manipulator requires safe navigation and arm motion near the patient, and hence a volumetric perception of the environment that a 2D laser map cannot offer. This work presents a 3D mapping and autonomous-navigation pipeline for the collaborative robot TIAGo in a clinical setting. RGB-D images and 2D laser scans are fused with RTAB-Map to build an OctoMap, an octree-based probabilistic volumetric map that distinguishes free, occupied and unknown space and captures elevated obstacles that a base laser scanner alone would miss. Autonomous exploration is implemented in ROS 2 Nav2 through a frontier-based strategy, a local cost map augmented with depth measurements, and a custom behavior tree that regulates the planning behavior. The pipeline was validated quantitatively against a simulated ground truth, reaching a median cloud-to-mesh error of 0.30 cm (mean 1.08 cm), and qualitatively on the real robot in a real living lab. Results show the feasibility of combining 3D mapping and autonomous navigation for this class of clinical environments.

ABSTRACT. Spatial neglect is a frequent consequence of stroke and severely limits the exploration of the contralesional space. Smooth Pursuit Training is one of the visual rehabilitation strategies employed to reduce this deficit, but its administration is generally qualitative and operator dependent. This work presents a pipeline that combines a wearable eye tracker, Tobii Pro Glasses 3, with the collaborative robot TIAGo in order to obtain quantitative measures of visual pursuit. The robot moves a physical target along predefined trajectories while the eye tracker records the gaze of the subject. A ninepoint calibration grid, observed at 85 cm, provides two linear regressions that convert gaze coordinates expressed in pixels into horizontal and vertical ocular angles. A YOLO network, trained on datasets built on recordings from the scene camera of the eye tracker, localizes the target in every frame. After removal of false positives, reconstruction of missing detections and temporal synchronization between gaze samples and video frames, the pipeline returns for each frame the angular position of gaze and target, their angular distance and a classification of the frame as IN or OUT of the target region. The system was validated on a healthy subject simulating a reduced visual exploration.

ABSTRACT. This study presents a ROS 2-based software architecture for the TIAGo mobile manipulator robot, designed to perform Pick and Place task to support the rehabilitation of post-stroke patients suffering from unilateral spatial neglect, a condition that impairs perception and interaction in the contralesional space. The system features an advanced visual perception module implemented in Python that processes RGB-D sensor data using YOLO deep learning models to segment target objects, filtering the associated Point Cloud to dynamically calculate their physical dimensions and 3D centroid. These spatial coordinates are then transformed into the robot's global reference frame and fed into the MoveIt 2 framework, which handles inverse kinematics, obstacle avoidance, and safe trajectory planning. The proposed architecture was extensively validated first in a simulated Gazebo environment to assess the accuracy of the vision node and the success rate of the grasping phases and subsequently tested on the physical TIAGo robot in a real laboratory setting, where it demonstrated high reliability and successfully outperformed the simulated results.

ABSTRACT. Current robotic systems for pediatric upper-limb rehabilitation are mostly adult adaptations that overlook children's specific motor, cognitive, and behavioral features. The SUPER-ARM project (Fit4MedRob) bridges this gap by integrating pediatric neurorehabilitation and motor control principles. Phase one (T0) assesses 14 typically developing children (TD) and 24 pediatric patients (PN) with neurological conditions and upper limb functional disability using the MOTORE end-effector robot alongside comprehensive neuromotor and neuropsychological assessments. Expected results highlight deficits in visuomotor precision and visual attention, correlating with robotic performance metrics like execution time and force control. Phase two (T1) evaluates an advanced prototype integrating Tobii ProFusion eye-tracking and Xsens wearable inertial sensors for high-resolution gaze localization and kinematic reconstruction. Ultimately, SUPER-ARM establishes clinical and industrial recommendations to optimize workspaces, reduce visual crowding, and advance paediatric robotic rehabilitation.

ABSTRACT. Robotic surgical skill assessment requires a combined analysis of instrument motion and of the surgeon's psychophysical state. This work proposes a protocol for building a multimodal dataset on the da Vinci Research Kit, synchronizing robot kinematics, multi-source video, and the operator's physiological and biomechanical signals (EMG, IMU, ECG and EDA). Based on suturing tasks of increasing complexity performed by surgeons of different expertise, the protocol provides a foundation for data-driven skill assessment and Guardian AI systems for intraoperative assistance.