ICINCO 2026 Abstracts

Area 1 - Industrial Informatics
Full Papers Short Papers
Area 2 - Intelligent Control Systems and Optimization
Full Papers Short Papers
Area 3 - Robotics and Automation
Full Papers Short Papers
Area 4 - Signal Processing, Sensors, Systems Modelling and Control
Full Papers Short Papers

Area 1 - Industrial Informatics

Full Papers
Paper Nr: 90
Title:

Symbiotic Fluidic Navigation of Robot Swarms for Distributed Monitoring in Drilling Fluids
Authors:

João Henrique Campos Soares, Ronnier Frates Rohrich and André Schneider de Oliveira
Abstract: This paper proposes a new architecture for sensing downhole oil well drilling activities using environmentally driven robotics. Unlike traditionally actuated robotics, which frequently fail in extreme downhole conditions, this work introduces the concept of Symbiotic Fluidic Navigation (SFN). The SFN approach uses a swarm of environmentally driven robotics to exploit the hydrodynamic power inherent in drilling fluids for actuation. Each robot in the swarm contains a high-frequency IMU and barometric pressure sensors, enabling it to continuously sample the kinematic and thermodynamic states of the surrounding fluid. The wellbore fluid dynamics provide natural hydrodynamic forces to propel the agents. Consequently, the locomotion energy demands are externalized to the subterranean environment. This mechanism enables the actuatorless robot swarm to achieve collective spatial awareness. A 6-meter vertical flow loop was used to evaluate downhole hydrodynamic conditions and to validate the sensor architecture and SFN approach. Kinematic and barometric data are processed through a collaborative architecture to achieve redundant collaborative validation. The results demonstrate that an environment-driven swarm can provide distributed monitoring in wellbore drilling environments.
Download
Short Papers
Paper Nr: 31
Title:

An Adaptive Vision-Based Robotic System for Flexible Surface Polishing
Authors:

Abderraouf Maoudj, Ahmed Kharidege, Brahim Brahmi, Azhar Aulia Saputra and Mourad Nouioua
Abstract: Current industrial automation paradigms increasingly demand intelligent systems that augment human capabilities by offloading perceptual and cognitive tasks to automated vision pipelines, thereby enabling operators to supervise high-mix,low-volume manufacturing processes more effectively. This paper presents a human-centered cell that performs these perception steps automatically for surface-polishing tasks, integrating a fixed eye-to-hand CMOS camera, a Siemens S7-300 PLC with a WinCC human–machine interface for robotic surface-polishing demonstrations, and a 6-DOF FANUC LR Mate 200iD manipulator. A HALCON pipeline conditions the image by median filtering, segments the part using a fixed threshold taken from a bimodal intensity histogram, extracts its contour with the Canny operator, and classifies its geometry by anisotropic shape-model matching. A planar homography maps the detected centroid to workspace coordinates, and the recognized class selects a stored, CAD-derived polishing program. The cell was demonstrated on six PLA specimens-spherical and ellipsoidal geometries in white, red, and blue under fixed LED lighting. In these trials the displayed class matched each specimen and the pick–inspect–polish–place cycle completed for every part; the recorded shape-matching times were 0.45 ms and 0.73 ms, and a 60-mm calibration check gave a 0.33% scale error. The integrated architecture significantly augments the operator’s situational awareness through real-time visual feedback, enabling seamless identification, grasping, and polishing of assorted workpieces while reducing processing latency and the cognitive burden on human workers.
Download
Paper Nr: 34
Title:

The Impact of Industrial Metaverse on IT/OT Integration in the Sphere of Data Sharing in Resilient Manufacturing
Authors:

David Acev, Mohsin Raza, Casper Schou and Ole Madsen
Abstract: With the introduction of Industry 4.0, Information Technology / Operational Technology (IT/OT) integration appears as a novel foundation. At the same time, industrial metaverse emerges as an advanced application that benefits from IT/OT integration. However, the introduction of industrial metaverse acquaints practitioners with the need to enhance the “traditional” IT and OT domains, thus shifting the boundaries stated by the industrial automation pyramid. Furthermore, Industry 5.0 initiates resilience as one of its main pillars, which refers to dealing with, estimating, and blocking internal and/or external disruptions from undermining the manufacturing performance. Our study employs a Systematic Literature Review (SLR) as part of an echeloned Design Science Research (eDSR) to extract information from current studies and build upon that knowledge to conceptualize how IT/OT integration is modified in the sphere of data sharing to overcome its identified challenges linked to predominantly technical prerequisites of the industrial metaverse while successfully dealing with disruptive scenarios. Our demonstration supports practitioners in understanding the novel division and the overlap of IT and OT tasks to develop an empirical fieldwork in resilient manufacturing.
Download
Paper Nr: 48
Title:

A Modular IoT Architecture for Laboratory Monitoring and Automation
Authors:

Manojkumar Krishnaramanujam Baskaran, Mohammed Faeik Ruzaij Al-Okby, Thomas Roddelkopf and Kerstin Thurow
Abstract: This paper presents a modular IoT architecture for heterogeneous laboratory environments. Modern laboratories rely on a mix of environmental sensors, mobile robots, instrument controllers, and asset trackers with different schemas, cadences, and reliability profiles. Existing IoT solutions are often tailored to single domains, with limited support for integrating diverse payloads, reasoning across domains, and propagating critical conditions through organizational hierarchies. We propose a modular architecture combining authenticated MQTT-based ingestion, schema-flexible payload handling, an event- and rule-based evaluation engine, and a role-aware multi-level escalation model with a pluggable notification layer. All data are persisted in a combined time-series and document-storage backend and exposed through a unified REST and WebSocket interface. The architecture supports diverse laboratory subsystems by providing a common monitoring, event-processing, and decision-support infrastructure for laboratory automation workflows. The architecture is evaluated on a real-time environmental monitoring deployment. Experiments with up to 500 simulated devices showed end-to-end latencies below 45 ms (P99) at 488 messages per second, stable multi-user dashboards, and reliable event-driven escalation. The paper contributes (i) a modular IoT architecture for multi-source laboratory data, (ii) a dynamic payload ingestion and event-evaluation model evaluated using environmental, mobile-robot, asset-tracker, and UWB payload structures, and (iii) an experimental validation using a safety-critical environmental monitoring scenario.
Download
Paper Nr: 51
Title:

Defending Cyber-Physical Systems against Watermark Nullification
Authors:

C. M. Krishna
Abstract: Watermarking has been widely studied as a means to detect attacks on cyber-physical systems. It consists of adding unpredictable control inputs and checking whether sensors report the expected change in controlled-plant state. In this paper, we consider how sensor reports can be used by a well-resourced adversary to build increasingly refined estimates of the applied watermark and its implications for the effectiveness of watermarking as a security measure. Such estimates can then be used by the adversary to spoof sensor reports which are close, statistically, to what the system expects to receive, based on its control inputs. Countering such an adversary increases the watermarking overhead. This paper evaluates the impact of such an adversary on the cost to the system of having to defend against such an attack.
Download
Paper Nr: 52
Title:

Dynamic, Physics-Based Wear Prediction Using Multimodal Sensor Data for Industrial Applications
Authors:

Johann Wiens and Christoph Reich
Abstract: In autonomous manufacturing, the industrial deployment of deep learning models for tool wear prediction frequently fails due to concept drift. This paper presents a dynamic, physics-informed online learning framework for the tool wear documented in the PHM 2010 dataset, built around a comparative study of Temporal Convolutional Network (TCN) and Long Short-Term Memory (LSTM) architectures, optionally regularized by Elastic Weight Consolidation (EWC). Physical domain knowledge is embedded twice: sparse measurement points are interpolated via the Paris-Erdoğan crack-growth law combined with an exponential function, and the predictions are constrained to monotonicity. An asymmetric Huber loss penalizes safety-critical underestimations more heavily than overestimations. Using leave-one-cutter-out cross-validation (c1, c4, c6), static offline baselines are compared with adaptive online approaches and related to recently reported PHM 2010 methods, evaluated by predictive accuracy (MAE, MSE, RMSE, MAPE, Pearson-r, bias) and efficiency (inference time, model size). As the validation rests on the three labeled cutters of a single machine under fixed cutting parameters, the results are reported as a proof of concept under these conditions.
Download
Paper Nr: 78
Title:

Control-Oriented Analysis of Machine Reliability and AGV Dispatching in Automated Manufacturing and Logistics Cells
Authors:

Wa-Muzemba Anselm Tshibangu
Abstract: Automated guided vehicles (AGVs) and autonomous mobile robots are increasingly deployed in automated manufacturing and logistics cells. Control strategies for these systems must coordinate machine operations, vehicle movements, and buffer usage while handling machine failures, variable fleet sizes, and congestion. These challenges are amplified in a global context where equipment and maintenance resources can be disrupted. This paper presents a control-oriented evaluation of machine scheduling and AGV dispatching policies in a flexible manufacturing system (FMS) served by AGVs. Using discrete-event simulation, we study combinations of local machine scheduling rules (shortest processing time versus first-in-first-out) and AGV dispatching policies (shortest travel distance versus first-come-first-served) under different levels of machine reliability and AGV fleet sizes. Performance is measured in terms of throughput, mean flow time, machine utilization, and work-in-process (WIP). Results show that control policies that exploit job processing time and travel distance information (SPT/STD) significantly outperform FIFO-based policies, especially when machines are failure-prone and fleets are constrained. We discuss how these findings can guide the design of real-time control algorithms for AGV systems and their integration with manufacturing execution systems (MES) and warehouse management systems (WMS) in cyber-physical production and logistics environments.
Download
Paper Nr: 99
Title:

Latent Association Discovery in Noisy Multi-Output Industrial Processes: A Glass Container Forming Use Case
Authors:

David Peña-Mangas, Carlos Cernuda and Daniel Reguera-Bakhache
Abstract: Many industrial processes exhibit strongly correlated inputs and high-dimensional, noisy outputs naturally structured into repeated units and profiles. Glass container manufacturing is one such case: upstream thermal conditions affect forming, while downstream glass-distribution measurements are noisy and span multiple production units. We study a production line in which nine feeder temperatures, arranged in a 3x3 grid, are related to infrared intensities measured over 64 vertical zones in each of 36 mold cavities. The objective is not full prediction of the profiles, but identification of the output component most strongly linked to the feeder thermal state. Four structured approaches are compared: a manual Center-of-Gravity baseline, a simultaneous automated latent model, and two variants of a sequential automated latent model. Across the four analyzed time intervals, full score correlations are consistently strongest for the automated interval-specific methods, above 0.65 for those approaches, compared with 0.47 for the manual Center-of-Gravity baseline. The results show that a thermal-linked latent mode can be detected in the vertical glass distribution, and that preserving the cavity × vertical structure is important both for robustness and for physical interpretation.
Download
Paper Nr: 120
Title:

VapeSense-IoT: Privacy-Preserving Low-Cost Intelligent Indoor Aerosol Monitoring
Authors:

Mohammad Zabad, Abbass Rammal and Mageda A.A. Sharafeddin
Abstract: Indoor air-quality monitoring is increasingly important in schools, universities, public buildings, and industrial facilities, where aerosol events such as vaping, smoking, and chemical sprays can affect health, safety, and policy enforcement. This paper presents a low-cost IoT-based air-quality monitoring system for real-time multi-class aerosol detection. The proposed system integrates a Raspberry Pi Pico W, a PMS5003 laser particulate matter sensor, and a BME688 environmental and gas sensor with a FastAPI-based inference server. Sensor readings are collected every 10 seconds and transformed into a 15-dimensional feature vector that includes particulate matter measurements, environmental variables, logarithmic transformations, temporal delta features, and a particulate matter (PM) PM2.5/PM10 ratio designed to improve discrimination between vape aerosol and combustion smoke. A Gradient Boosting classifier is trained and evaluated on an expanded two-room dataset containing 2,005 real sensor samples collected from the deployed hardware across four classes: normal air, vape, smoke, and spray. Using an 80/20 train-test split, with 20% of the data reserved for unseen held-out testing, the final model achieves 97.76% test accuracy and 97.75% weighted F1-score, with a mean single-sample inference latency of 0.947 ms. When vape or smoke is detected with sufficient confidence, the system generates a real-time Telegram alert while logging all readings for continuous dataset growth and future retraining. The results show that accurate indoor aerosol classification can be achieved using commodity hardware costing approximately $36, offering a privacy-preserving and scalable alternative to expensive commercial vape-detection systems.
Download
Paper Nr: 131
Title:

Adaptable Human-Robot Task Scheduling Applicable in Diverse Operational Contexts
Authors:

Michail Maniadakis, Kapil Deshpande, Georgios Alexakis, Stefan Fixl, Konstantinos Papoutsakis, Sharath Chandra Akkaladevi and Manolis Lourakis
Abstract: Widespread adoption of collaborative robots across industrial sectors is hindered by the complexity of configuring them to operate alongside human operators, as current deployment practices require extensive reprogramming that limits scalability and increases development time and cost. To address this challenge, this work proposes a structured human–robot collaboration framework that decouples collaboration logic from robot-specific implementations, enabling portability across diverse robotic platforms and industrial domains. The framework comprises an application-independent collaboration layer with reusable multi-agent coordination mechanisms, supported by a Behaviour Tree-based execution system that dynamically instantiates parameterized robot skills and provides a flexible interface between workflow-level task scheduling and platform-specific control. In addition, a human observation module continuously monitors the operator's location and task progress, providing real-time situational awareness that enables informed scheduling decisions and responsive robot behaviour. The framework is validated in an industrial vehicle door disassembly scenario involving two heterogeneous robotic systems, demonstrating its flexibility and portability across distinct collaboration paradigms, interaction modalities, and robot capabilities.
Paper Nr: 133
Title:

An Integrated IoT Architecture for Real-Time Forest Fire Risk Monitoring and Early Warning Using LoRa, Cloud Computing and Machine Learning
Authors:

Raimundo Geanderson Amador dos Santos and Marlon Marcon
Abstract: This position paper presents an integrated, end-to-end Internet of Things (IoT) architecture for forest fire risk monitoring and early warning in remote regions lacking Wi-Fi or cellular connectivity. Autonomous sensor nodes (LilyGo TTGO T-Beam, ESP32-S3 with LoRa SX1262) communicate via point-to-point LoRa (868/915 MHz) to a cloud pipeline on Amazon Web Services (AWS) (IoT Core, Lambda, DynamoDB, S3), provisioned via Terraform. A containerised Random Forest model, deployed as an AWS Lambda function, predicts fire risk on a 0–5 scale from temperature, humidity, and consecutive dry days, visualised through a Streamlit Cloud dashboard. Five validation phases are complete: virtual prototyping on Wokwi, a multinode simulated LoRa network, an end-to-end AWS pipeline, machine learning validation on 627 INMET stations (F1 = 0.748 / AUC-ROC = 0.801, 2024 validation; F1 = 0.710 / AUC-ROC = 0.759, 2025 test), and a dashboard validated with 10,533 readings from three Parana stations. Physical hardware has been acquired ´ and field validation in Portugal is ongoing. Rather than proposing new machine learning methods, this work’s principal contribution is integrating autonomous sensing, long-range LoRa communication, cloud processing, and predictive analytics into a single, reproducible architecture, presented as a partially validated prototype pending field evidence and comparison against established fire-risk models.
Download
Paper Nr: 27
Title:

Adaptive Thresholding for Unsupervised Deep Learning-Based End-of-Line Quality Inspection in Electric Motor Production
Authors:

Jernej Mlinarič and Gregor Dolanc
Abstract: Reliable end-of-line (EoL) quality inspection is crucial in electric motor production. Deep learning autoencoders trained on healthy motors enable effective unsupervised anomaly detection. Unfortunately, conventional fixed thresholds often fail due to variations in manufacturing processes or noise. This paper suggests an adaptive thresholding strategy applied to a hybrid CNN–BiGRU autoencoder using Mel-frequency spectrograms (MFSs) of acoustic and vibration signals. Thresholds are dynamically estimated from the distribution of reconstruction errors, improving robustness without additional labelled data. Experiments on over 2,400 motors show reduced false positives while maintaining high detection accuracy. The method enables flexible, automated, and reliable EoL inspection in industrial environments.
Download
Paper Nr: 151
Title:

A Physics-Informed Neural Network Digital Twin for Thermal Control in Biomedical Robots
Authors:

Carlos Rangel-Cascajosa, Plácido Fernández-Cuevas, Juan M. Montes-Sánchez, Saturnino Vicente-Díaz and Ángel Jiménez-Fernández
Abstract: Precise thermal regulation in biomedical analysis robots is a critical requirement for maintaining the efficacy of chemical reactions in biological samples. While purely data-driven Recurrent Neural Networks (RNNs) are often used to model complex thermal inertia, they tend to accumulate significant error (drift) during autonomous closed-loop inference, especially when facing unobserved environments. To address this safety concern, we propose a Physics-Informed Neural Network (PINN) Digital Twin (DT). By integrating thermodynamic ordinary differential equations (ODE) into the network’s loss function, the model learns the physical constraints of the system, effectively blocking non-physical predictions. We evaluated the framework using a self-recorded large dataset of 4,106 experimental step-response signals from a real biomedical robot that was made available to the public. This dataset was carefully partitioned to create a strict Out-Of-Distribution (OOD) test set containing only extreme high-temperature operations (≥100◦C). The results are clear: while the standard purely empirical (STD) model struggles with extrapolation-reaching peak instantaneous errors of up to 46.99◦C-the PINN uses thermodynamic principles to keep extreme deviations below 4.13◦C. Overall, our physics-guided approach reduces the Mean Absolute Error (MAE) for the critical surface sensor by 62.3% and successfully mitigates recursive cascade drift. This paves the way for reliable real-time predictive maintenance in clinical environments.
Download

Area 2 - Intelligent Control Systems and Optimization

Full Papers
Paper Nr: 21
Title:

Learning Changes the World and Then What Was Learned Has Changed: Training and Deploying Neural Networks in the Control Loop
Authors:

Vittorio Lippi
Abstract: Neural networks offer critical adaptability for controlling complex systems where nominal modeling proves intractable. However, when offline-trained neural networks are deployed within closed-loop control systems, their actions fundamentally alter the state distribution upon which they were trained. This paper introduces a unified taxonomy to categorize these sequential failure modes: the initial Closed-Loop Identification Error, the deployment-induced Action Error, and the continuous Tracking Error. Through discrete-time sensitivity analysis, this work formalizes how offline-learned parameters locally guarantee a trajectory divergence upon deployment, degrading performance due to local loss geometry. To mitigate this distribution shift, we propose actionable design policies centered on time-scale separated online adaptation. These mechanisms are validated through ablation studies on a nonlinear Duffing oscillator, demonstrating that bounded continuous learning successfully mitigates trajectory divergence whereas static deployments fail, offering concrete guidelines for practitioners deploying machine learning components in safety-critical feedback systems.
Download
Paper Nr: 69
Title:

Planning Mobile Project Resources as Integrated Systems: A Research Roadmap for the Assignment, Scheduling, and Routing Problem
Authors:

Ayman R. Mohammed, Majed Hadid and Roberto Baldacci
Abstract: Mobile project resources such as crews, specialized equipment, and traveling technical teams create value only when the right capability reaches the right work at the right time. This requirement makes assignment, scheduling, and routing inseparable in construction and related mobile-service settings. The Assignment, Scheduling, and Routing Problem (ASRP) captures that integrated planning structure, but field deployment depends on more than solver strength. This paper develops a research Roadmap for ASRP-based planning of mobile resources that execute physical tasks. The Roadmap is grounded in three evidence streams: the structural coupling of assignment, scheduling, and routing; a systematic review of 123 mathematical-programming studies in construction optimization; and adjacent operations-research and construction-management literatures. The resulting agenda is organized around eight interacting levers: optimization and algorithms; data, benchmarks, and standards; technology and digitalization; sustainability and carbon; workforce, skills, and safety; systems and network design; uncertainty and resilience; and policy, contracts, and economics. A mind map summarizes the levers, their dependencies, and the cross-cutting barriers. This paper does not propose or experimentally benchmark a new solution method. Its contribution is a literature-grounded Roadmap that positions the ASRP relative to adjacent problem classes and defines the methodological, data, and validation requirements for future algorithmic and empirical studies.
Download
Paper Nr: 71
Title:

Adaptive Stochastic Model Predictive Control for Grid-Connected Hybrid Energy Storage Management
Authors:

Taranjitsingh Singh, Boulaid Boulkroune, Maxime Monsieur and Chris Ganseman
Abstract: This paper presents an adaptive scenario-based stochastic model predictive control framework for energy management in a grid-connected hybrid energy storage system. The considered system combines high-energy and high-power batteries, renewable generation, local demand, and utility-grid interaction. Deterministic MPC uses a nominal forecast, while conventional scenario-based stochastic MPC uses a fixed uncertainty distribution and equal scenario probabilities. The proposed adaptive stochastic MPC updates its uncertainty representation online using a causal rolling window of measured net-load forecast residuals. A residual predictor shifts the near-term probabilistic demand forecast, while residual-dependent spread corrections adapt the P10–P50–P90 envelope. Candidate Monte Carlo scenarios are assigned Gaussian-kernel scores, normalized into adaptive probabilities, and pruned before optimization. All controllers use the same multi-term scenario cost, and the proposed adaptive scenario-based stochastic Model Predictive Control (ASMPC) minimizes the adaptive probability-weighted expected scenario cost. Closed-loop performance is evaluated over a persistent unexpected load-shock interval. ASMPC performs better than the deterministic MPC, Robust MPC and Stochastic MPC, while satisfying battery state-of-charge constraints.
Download
Paper Nr: 100
Title:

An Explainable Neuro-Fuzzy Adaptive Graph KAN–TCN Framework for Short-Term Urban Parking Occupancy Forecasting
Authors:

Mohammad Reza Mohebbi, Farzaneh Keighobadi and Mario Döller
Abstract: Short-term urban parking occupancy forecasting is a challenging spatiotemporal problem because future parking availability depends on recent demand dynamics, spatial interactions among parking zones, and heterogeneous urban context. Reliable and interpretable forecasts are needed to support intelligent transportation decisions, since black-box deep learning models provide limited insight into when, where, and why their predictions should be trusted. This paper proposes an explainable neuro-fuzzy adaptive graph framework that integrates Kolmogorov-Arnold Network (KAN)-based nonlinear spatial representation with a Temporal Con-volutional Network (TCN) for short-term urban parking occupancy forecasting. The forecasting component combines hybrid adaptive graph construction, Chebyshev graph filtering, KAN-based nonlinear transformations, dilated causal temporal convolution, and temporal attention pooling. To improve transparency, an Adaptive Neuro-Fuzzy Inference System (ANFIS) is trained as a post-hoc surrogate to approximate the trained forecaster and extract interpretable feature importance patterns, temporal regime diagnostics, and human-readable fuzzy rules. Experiments on Melbourne on-street parking data across 15-, 30-, and 60-minute horizons show that the proposed model improves over the Adaptive Graph Convolutional Recurrent Unit (AGCRU) benchmark, with the strongest gain at the 15-minute horizon. The framework supports accurate forecasting and explainable decision support for intelligent urban parking management.
Download
Paper Nr: 106
Title:

An Evaluation of Deep Learning Models for DBH Estimation Using UAV-LiDAR Data in Eucalyptus Inventory
Authors:

Felipe M. Salles, Adimara B. Colturato, Cristiano R. Reis, Luis Pelin, Marcos G. Quiles and Kalinka Castelo Branco
Abstract: Forest inventory traditionally relies on manual field measurements, which are time-consuming and affected by human error. In this context, Light Detection and Ranging (LiDAR) technology, when integrated with Unmanned Aerial Vehicles (UAVs), has emerged as an alternative for acquiring high-resolution 3D data of forest structures. However, traditional LiDAR-based approaches still depend on empirical models calibrated with field data. To address this limitation, the main contribution of this study is the investigation of deep learning (DL) techniques for estimating DBH from UAV-LiDAR data in Eucalyptus forests using only LiDAR points. In this work, we proposed a Convolutional Neural Network (CNN), a PointNet model, and a PointTransformer model across three scenarios: using the same number of input points, adding Gaussian noise, and dropping 50% of points. The results showed that CNN achieved the best performance in the baseline scenario (MAE = 1.705 cm and RMSE = 1.897 cm), while PointNet was the most robust model under data degradation, obtaining the lowest errors in both the Gaussian noise (MAE = 1.481 cm, RMSE = 1.656 cm) and point-loss (MAE = 1.508 cm, RMSE = 1.845 cm) scenarios.
Download
Paper Nr: 119
Title:

A Symmetry-Reduced Optimal Control Formulation for Quadrotor Dynamics Based on Noether's Theorem
Authors:

Tanzeela Tasneem, Enzo Avelino, George Kamel, Joseph Chazbani and Daniela Nikolova
Abstract: Trajectory optimization plays a central role in enabling quadrotor unmanned aerial vehicles (UAVs) to execute feasible and energy-efficient maneuvers. Existing nonlinear optimal control methods solve the full-order system dynamics, with current research efforts primarily focused on improving algorithms, discretization techniques, and solver performance rather than simplifying the underlying mathematical model. In this paper, we present a symmetry-reduced optimal control framework that exploits the inherent rotational symmetry of quadrotor dynamics to reduce the dimensionality of the dynamic model before optimization. The quadrotor dynamics are first formulated using Lagrangian mechanics, and Noether's theorem is then applied to identify the conserved generalized momentum associated with yaw symmetry. This conserved quantity replaces the yaw differential equation with an algebraic constraint, yielding a reduced-order model that preserves the essential nonlinear dynamics. Both the full-order and symmetry-reduced formulations were implemented within the CasADi–acados optimal control framework and evaluated under identical optimization settings. Compared with the full-order formulation, the symmetry-reduced model achieved a 23% reduction in solver time and a 1.28% reduction in maneuver time, with only a 0.015% increase in the final objective cost. These results demonstrate that symmetry reduction can improve the computational efficiency of nonlinear optimal control problems while maintaining comparable trajectory quality.
Download
Paper Nr: 134
Title:

Q-Learning in Action: Explainable and Verified Renewable Energy Source Selection
Authors:

Afef Gafsi and Imen Ben Hafaiedh
Abstract: The transition to renewable energy requires smart systems choosing between sources (solar, wind, hydro, etc.) based on minimizing GHG emissions. We propose a framework integrating a Timed Automata model of a Q-learning agent to optimize selection by reducing both GHG and transition costs, while ensuring explainability and verifiability of convergence and stability. The Timed Automata represents the agent’s interaction with the energy system via states, transitions, and costs. Trained within this formal structure, the Q-learning agent learns optimal alternatives. Model checking provides verification and validation capabilities to analyze the agent’s behavior. Beyond traditional optimization, this work formally verifies convergence and stability under different scenarios. Model checking further provides explainable insights and verifiable performance guarantees. The integration offers a robust, verifiable approach to renewable management, with advantages in explainability and trustworthiness.
Download
Paper Nr: 158
Title:

Combining Large Language Models with Symbolic Reasoning for Robust Aspect-Based Sentiment Analysis
Authors:

Moemen Said, Nadia Smairi, Houda Abadlia and Olfa Belkahla Driss
Abstract: Aspect-based Sentiment Analysis (ABSA) needs accurate aspect term extraction and reliable sentiment classification, yet existing neural approaches remain susceptible to linguistic phenomena such as negation, comparative expressions and contrastive discourse. This paper proposes a hybrid neuro-symbolic three-stage framework combining LoRA-adapted Qwen2.5-7B for aspect extraction, DeBERTa-v3-large for aspect sentiment classification and Answer Set Programming (ASP) for post-inference symbolic refinement. The framework uses self-consistency decoding for robust aspect extraction and applies declarative linguistic rules to correct systematic sentiment prediction errors without retraining the neural models. Experiments on four SemEval benchmark datasets (Laptops 2014 and Restaurants 2014, 2015, and 2016) achieve joint Macro F1 scores of 80.21, 85.08, 84.56, and 79.76, respectively, consistently outperforming the compared baselines without performance regression. The results show that integrating symbolic reasoning with modern neural architectures improves the robustness and consistency of aspect-level sentiment analysis.
Short Papers
Paper Nr: 17
Title:

MIMO and Multi-Group Comparison Problem in Process Control: A Multivariate Statistical Framework for Systems Represented with Frequency Response Functions
Authors:

Vittorio Lippi, Leonard Johard and Gabriele Landucci
Abstract: Characterizing complex Multi-Input Multi-Output (MIMO) systems presents two issues: feedback controllers mask fault variance, rendering single-variable monitoring ineffective, and repeated testing across multiple conditions inflates false positive rates due to the multiple comparisons problem. This study proposes a multivariate statistical framework to resolve these limitations. We extend a statistical library that identifies system dynamics via Frequency Response Functions (FRFs) by transforming them into time-domain Pseudo-Impulse Responses (PIRs). This functional representation captures the complete dynamic signature of the system, offering a richer diagnostic profile than traditional static scalar metrics. The framework, originally developed for SISO systems, is extended to the MIMO case by introducing supervectors that aggregate multiple PIRs, evaluated using Permutational Multivariate Analysis of Variance (PERMANOVA). This non-parametric approach handles the high dimensionality and complex correlation structures inherent in MIMO functional data. We demonstrate that the PIR Supervector is more powerful than single-variable analysis, and the MIMO PERMANOVA approach outperforms traditional SISO anomaly detection. By distinguishing between normal operation, external thermal disturbance effectively compensated within the investigated operating range, and severe parametric faults that standard univariate methods miss, the framework provides a single, rigorous metric for specific fault diagnosis (p < 0.001).
Download
Paper Nr: 30
Title:

Analysis of Adhesion Effect on Grasp Stability with Robotic Gripper
Authors:

Caroline Pascal and Mathieu Grossard
Abstract: This paper investigates the benefits of active adhesion induced effects on grasp stability. In the framework of object grasping with multifingered robot gripper, a finger-object contact model is first introduced in that respect. It takes into account reactions effects coming from both an usual friction-based contact wrench and an active adhesion-induced wrench, in a way that allows an accommodation to the usual grasp formalism. We then formally demonstrate the advantages of this model when applied to force-closure like stability criteria. In particular, we prove that incorporating adhesion effects alongside purely friction-based actions theoretically leads to a set of optimal contact forces with reduced norms. Finally, these contributions are validated through several illustrative examples, which are analytically evaluated to highlight the positive impact of adhesion, especially in the context of non-prehensile grasps.
Download
Paper Nr: 50
Title:

Design of Neurocontrollers for Nonlinear Dynamic Systems with Summation and Difference Neurons
Authors:

Ivan Sekaj, Aleš Melichar, Danil Zlobin, Miroslav Halás, Jaromír Skirkanič and Ivan Kenický
Abstract: We propose a new neural network architecture for controllers of nonlinear dynamic systems. The network is based on a feedforward architecture with additional layers containing dynamic-behavior neurons, and it is trained using neuro-evolution. This is used by a genetic algorithm, which optimizes the neural network parameters by minimizing a chosen criterion function that represents the control performance measure. Next, we compared the proposed network with a conventional feedforward network with two hidden layers and input signal preprocessing. Additionally, we compared the neuro-evolution learning approach with another machine-learning approach based on the PPO-type reinforcement learning. All proposed machine-learning methods and architectures are experimentally compared with a conventional PID controller. Finally, a practical stability evaluation method is proposed and discussed.
Download
Paper Nr: 55
Title:

An Epipolar Geometry-Based Approach for Distributed Formation Control Using Vision
Authors:

Patricia M. Tavares, Edgar I. Chavez-Aparicio, Jean-Bernard Hayet and Hector M. Becerra
Abstract: In this work, we formulate a composition of consensus-based formation control approaches to solve the problem of reaching a formation with a group of unmanned aerial vehicles (UAVs). The originality of these proposals is that they do not use estimates of the UAVs global positions and employ onboard vision as the unique sensory modality, relying essentially on the geometry of two perspective views, i.e. the epipolar geometry. We validate our proposals in different simulation scenarios and show their robustness to image noise.
Download
Paper Nr: 60
Title:

Integrated Multi-Objective Optimization for AMR Fleet Operations in Automotive Manufacturing
Authors:

Ahmed Tolba
Abstract: Existing multi-objective formulations for AMR fleet scheduling in automotive just-in-time manufacturing optimize makespan or total tardiness: objectives that fail to capture worst-case delivery risk when any single violation causes a line stoppage. We introduce a bi-objective formulation pairing fleet size minimization with minimum slack maximization, directly quantifying operational buffer at the system’s most critical task. Because tasks arrive dynamically and full deliveries are released only after paired empties complete, offline population-based re-optimization is structurally incompatible with operational replanning; we therefore employ event-driven deterministic scheduling with irrevocable assignments, achieving O(|T|log|T|) complexity and processing 12,524 tasks in 1.2 seconds. Systematic evaluation across 48 configurations (four task priorities × two robot selection strategies × six fleet sizes) on an industrial-scale instance of 284 part types reveals that robot selection strategy dominates performance by over two orders of magnitude: distance-based (Nearest) selection achieves zero violations at 65 robots, while time-based (FIFO) selection produces a persistent 2.9% violation rate irrespective of fleet size. Task priority policy (EDF, LFT, SPT, Slack) has negligible impact under Nearest selection. The Pareto-optimal configuration, Nearest-EDF at 65 robots, delivers Smin = +70.3 s with 52% robot utilization (the necessary buffer for charging, traffic, and failure absorption). Deterministic scheduling thus suffices for algorithm selection; stochastic simulation is then required only for final fleet sizing of the already-identified optimal policy.
Paper Nr: 61
Title:

STRIDE: Investigating Efficient Sampling Strategies for Neural Approximations of Optimal Control
Authors:

Andrea Giusti, Maxime Monsieur and Branimir Mrak
Abstract: While performant, Model Predictive Control (MPC) is often not used for fast mechatronic applications with non-linear dynamics or complex tasks, due to its computational cost. To address this issue, neural network (NN) based approximations of MPC have been proposed. These generally follow an imitation learning approach, where example MPC solutions are approximated with NNs. This leads to several new challenges: efficiently generating training examples, which requires solving a large number of MPC optimization problems, and ensuring that the resulting dataset is informative enough to train a stable and robust NN controller. To address these challenges, we propose a new dataset aggregation strategy, combining the robustness improvement provided by DAGGER, with exploration and performance evaluation over the entire task space. We demonstrate the approach on a parallel SCARA robot, both in simulation and experimentally. Simulation results show that our approach reduces required data by a factor of 2, compared to equidistant gridding benchmark, while simultaneously reducing the positioning error by 71%.
Download
Paper Nr: 98
Title:

Spiking Neural Networks for Object Detection: A Review of 2024/2025 Advancements
Authors:

Mikołaj Waśniewski, Jakub Caputa, Wojciech Chmiel and Maciej Wielgosz
Abstract: Spiking Neural Networks (SNNs) have emerged as a promising paradigm for low-latency and energy-efficient visual perception, particularly in resource-constrained systems such as autonomous vehicles and unmanned aerial vehicles. Unlike conventional deep neural networks that depend on dense frame-based computation, SNNs process information through sparse, event-driven spike trains, naturally exploiting data from neuromorphic sensors. Despite their energy efficiency, leveraging SNNs for tasks such as object detection, which combines classification and bounding-box regression remains challenging due to the non-differentiable nature of spiking functions and information loss during network conversion. This review examines recent advancements in SNN-based object detection architectures introduced in 2024 and 2025. We discuss how these models incorporate novel mechanisms to address key challenges in neuromorphic vision. By synthesizing their performance across benchmarks such as MS COCO and Automotive GEN1, this paper highlights the narrowing performance gap between SNNs and artificial neural networks (ANNs) and outlines future directions for real-time neuromorphic object detection.
Download
Paper Nr: 136
Title:

Smooth Optimal Tracking-Control Approximating Maximum Control Energy Usage for a Class of Linear Systems with a Nonlinear Output Function and Control Limitations
Authors:

Mir Mamunuzzaman and Jörg Mareczek
Abstract: We consider a class of linear systems with nonlinear output function representing important applications like speed-controlled robot-manipulators. A desired trajectory for the output shall be such that the corresponding control trajectory is running into constraints so that tracking-error arises. We seek for the optimal control trajectory that minimizes the integral absolute tracking-error while exploiting maximum control energy. Thus, we face the optimal control problem with the time integral of the tracking-error squares as cost functional and inequality constraints for control limitations. This leads to Pontryagin’s Maximum Principle (PMP). However, except for simple example systems, PMP is running into huge numerical calculation costs. This can be avoided by taking care of the control constraints by an additional punishing function in the cost functional, known as regularization function; the inequality constraints become obsolete. However, the original optimization objective has to be abandoned. Indeed, it is current practice in optimal control to punish also control energy, e.g. in LQR. It reduces the amount of control energy but does not use maximum control energy. Beside that also barrier-functions are commonly applied. They tend to infinity as the control approaches its limitation. However, like punishing methods, it is not possible to generate solutions that approach the optimal PMPsolution. The contribution of this paper is to present a general design method for barrier-functions that fills this gap. With these specially designed barrier-functions we are able to approximate infinitely close the optimal PMP-solution thus using maximum control energy. Beside mathematical derivations of our novel design method we present an example for which PMP can be calculated analytically. We compare the effects of different regularization functions and show that our novel design method indeed approaches to PMP and thus yields best results.
Download
Paper Nr: 142
Title:

Reinforcement Learning Based Distributed Tracking Control of Multiple Uncertain Linear Systems
Authors:

S. Hernandez and Wenjie Dong
Abstract: This paper considers distributed tracking control of multiple uncertain linear systems with a leader. To solve the tracking control problem, a two-step controller design approach is applied. First, a distributed estimator is proposed for each system such that the leader’s information is estimated in finite time. Then, a tracking controller is proposed for each system with the aid of reinforcement learning technique. Due to uncertainty in each system, a direct adaptive reinforcement learning algorithm is proposed. Simulation study shows the effectiveness of the proposed controllers.
Download
Paper Nr: 25
Title:

A Biped Walking Consisting of Single-Support and Double-Support Phases
Authors:

Yannick Aoustin and Alexander Formal'Skii
Abstract: A walking with a single support and a double support phases is proposed for a seven-link biped with massless feet. In single support only the ankle of the supporting leg is actuated. The double support phase occurs in finite time. During the phase the rotation on the ground of the massless feet are synchronized, similarly to the human gait. The double support phase has a propulsive action during the walking. Two boundary value problems are solved to design the single support phase and the double support phase, and ensure the periodicity.
Download
Paper Nr: 44
Title:

Robust Bell-State Preparation and Observation in Two-Qubit Systems Using Sliding Mode Methods
Authors:

Kaydian Quintero, Bereket Berhane and Sergey V. Drakunov
Abstract: This work studies the robust preparation of entangled states in a coupled two-qubit system by means of sliding mode control. The two-qubit Schrödinger dynamics are first written as a bilinear control system, and then mapped into a set of real quadratic coordinates that describe selected two-qubit correlations. In these variables, the first two coordinates evolve as a second-order subsystem whose acceleration depends affinely on the available control fields. A discontinuous feedback law is designed to force the trajectories onto a prescribed sliding manifold and to regulate the correlation coordinates to Bell-state reference values. Numerical simulations illustrate fast convergence of the selected correlation variables and decay of the sliding surfaces for the four standard Bell-state references. The results support sliding mode methods as a useful robust-control framework for entanglement generation in bilinear quantum systems.
Download
Paper Nr: 79
Title:

Runway-Referenced Deep Reinforcement Learning Control for Autonomous Landing of a 6-DoF Aircraft Model
Authors:

Melisa Nisa Aşlayan Üzer and Barış Başpınar
Abstract: This study presents a runway-referenced autonomous landing control framework for a high-performance 6-DoF aircraft model using deep reinforcement learning. The flight dynamics are simulated in JSBSim, and a Deep Deterministic Policy Gradient (DDPG) agent is trained to generate continuous throttle, aileron, elevator, and rudder commands during the approach and flare phases. The landing task is formulated in a runway-aligned coordinate system based on the runway threshold, heading, along-track position, and cross-track deviation, enabling the controller to directly account for runway alignment and touchdown-location requirements. The reward function combines dense approach-shaping terms with terminal touchdown logic, including glide-path tracking, lateral corridor regulation, heading alignment, bank-angle and angle-of-attack constraints, touchdown-zone scoring, and control-effort penalties. A two-stage Bayesian hyperparameter optimization procedure is used to improve task-level landing performance. Under randomized initial approach conditions, the final policy achieves 89% touchdown detection, 67% touchdown-zone success, and 57% safe-touchdown rate when touchdown-zone, bank-angle, angle-of-attack, and heading constraints are jointly evaluated. Additional runway-specific simulations on the Akinci runway configuration show repeatable descending approach profiles and touchdown clustering near the intended runway region. These results demonstrate the preliminary feasibility of a runway-aware DRL-based control framework for autonomous landing in simulation.
Download
Paper Nr: 82
Title:

Evolutionary Fixed-Structure H ∞ Controller Synthesis with Performance Robustness against IQCs
Authors:

Philippe Feyel
Abstract: This paper deals with the computation of performance robust H∞ structured controllers subject to Integral Quadratic Constraints (IQC). Although the fixed-structure H∞ controller synthesis problem for linear systems is well addressed using either non-smooth optimization or evolutionary computation, the resulting controllers may suffer from a lack of stability and/or performance robustness against nonlinearities (such as dead-zone, hysteresis, etc.). Coupling the IQC framework a priori with the “classical” fixed-structure controller synthesis problem allows computing structured controllers with a certain degree of performance despite non-linearities, if they can be described by IQCs. Due to the complexity of the problem, we use the perturbed differential evolution optimization algorithm, which has proved to be quite performing in fixed-structure robust and optimal controller computation and for high dimensional problems more generally.
Paper Nr: 127
Title:

A Genetic Algorithm Approach for LPG Transportation Scheduling under Loading and Unloading Constraints
Authors:

Yadira Bravo, Pablo Proaño, Danilo Chavez and Alan Cuenca-Sánchez
Abstract: This paper presents a genetic algorithm-based scheduling approach for reducing excessive stoppage times in liquefied petroleum gas (LPG) tanker truck transportation. The study considers an anonymized logistics operation in Ecuador comprising two loading stations, two transportation routes, and one common unloading station. Since the loading and unloading facilities are modeled as single-server resources, simultaneous or closely spaced truck arrivals generate queues, idle time, and opportunity cost losses. The dispatching task is formulated as a discrete optimization problem in which each truck is assigned to one feasible arrival time within a predefined scheduling window. Each candidate solution is evaluated by simulating the loading queues, route-dependent travel times, and the unloading queue at the destination facility. The objective function minimizes the opportunity cost associated with waiting times throughout the transportation process. The case study considers 13 tanker trucks and six feasible arrival time slots. The results show that the proposed approach reduces the objective function value from 6555.41 to 4712.44, corresponding to a relative improvement of 28.11%. In addition, the cumulative waiting time at the common unloading station decreases from 158.90 h to 112.07 h. These findings indicate that queue-aware scheduling can reduce operational congestion, improve fleet utilization, and support daily planning decisions in LPG transportation systems.
Download
Paper Nr: 150
Title:

Feature-Level Explainability for Hyperdimensional Computing-Based Anomaly Detection
Authors:

Víctor Ortega, Soledad Escolar, Fernando Rincón, Jesús Barba, Julián Caba and Juan Carlos López
Abstract: Existing Hyperdimensional Computing (HDC) frameworks accurately detect anomalous samples but provide limited insight into the feature values responsible for triggering the detected anomaly. In this work, we define explainability as the capability of explicitly identifying the subset of feature values responsible for a detected anomaly, rather than merely estimating the relative contribution of each feature to the classification outcome. To this end, we propose a feature-level HDC framework that constructs prototype hypervectors for individual features, enabling the encoded hypervector of each feature to be independently compared against its corresponding feature prototype hypervector. The proposed approach is evaluated through a real-world anomaly detection case study based on emergency lighting devices, where controlled perturbations are injected into selected feature values to assess the explainability mechanism. Experimental results demonstrate a high Anomalous Feature Identification Rate (AFIR) under different evaluation scenarios, showing that the proposed framework accurately identifies the feature values responsible for detected anomalies while preserving the simplicity and computational efficiency of the HDC pipeline.
Download

Area 3 - Robotics and Automation

Full Papers
Paper Nr: 76
Title:

Model-Free Dynamic Compensation Based on Ultra-Local Modeling for Trajectory Tracking of an Aerial Manipulator
Authors:

Gabriela M. Andaluz, Paulo Leica, Oscar Camacho and Angélica Quito
Abstract: Aerial manipulators are highly nonlinear and strongly coupled robotic systems whose performance often depends on the availability of accurate dynamic models. However, model-based compensation strategies may suffer performance degradation in the presence of parameter uncertainties, payload variations, and external disturbances.This paper presents a Model-Free PID (MF-PID) dynamic compensation strategy based on ultra-local modeling for robust trajectory tracking of an aerial manipulator. The cascade architecture comprises an outer-loop PID for trajectory reference generation and an inner-loop MF-PID for online estimation of unknown dynamics using an input–output ultra-local model. Unlike model-based approaches, the proposed method obviates the need for an accurate dynamic model, enhancing robustness against uncertainties and disturbances. The controller is evaluated under severe conditions: a 60% increase in inertial parameters and a 3.5-fold amplification of Coriolis terms, representing payload variations and wind-induced disturbances. Performance is benchmarked against conventional PID and inverse dynamics (ID) controllers using ISE, IAE, and TVU indices. Simulation results show that the MF-PID achieves a 28% reduction in ISE and a 21% reduction in IAE relative to the ID controller, while also outperforming the conventional PID. These results confirm that ultra-local model-free dynamic compensation is an effective alternative to model-based control for aerial manipulators under severe uncertainties.
Download
Paper Nr: 77
Title:

Enhanced Indoor and Parking Navigation for UGV: A Multi-Layered Algorithmic Overlay for Patchwork++
Authors:

Kevin Helvig, Julien Marzat and Martial Sanfourche
Abstract: Robust ground segmentation is a fundamental prerequisite for safe autonomous navigation. While state-of-the-art algorithms like Patchwork++ developed for autonomous driving exhibit high performance in standard urban environments, they often face critical limitations in constrained indoor scenarios, e.g. underground parking lots. Common issues include false positives generated by ceiling artifacts and a lack of granularity in hazard assessment. In this paper, we propose a geometric multi-layered algorithmic overlay that directly processes Patchwork++ outputs to optimize indoor navigation. Our approach introduces three key contributions: (1) a spatial filtering mechanism to systematically eliminate ceiling and upper-layer noise; (2) a fine-grained obstacle classification module that geometrically distinguishes between traversable elements and non-traversable structures; and (3) a dedicated negative obstacle detection layer to identify ditches, drop-offs, and open manholes despite sparse LiDAR returns. Experimental evaluations conducted on real-world parking tests demonstrate that the grid map produced by our overlay provides an accurate traversability representation that can be reliably exploited in real time for autonomous navigation.
Download
Paper Nr: 81
Title:

Visual Servoing under Scalar Depth Uncertainty: Classical, Predictive, and Takagi–Sugeno Control
Authors:

Boucherika Mohamed Abdelmalek, Jason Chevrie, Laurent Vermeiren, Benyamine Allouche and Antoine Dequidt
Abstract: In image-based visual servoing (IBVS), the control error is defined directly in the image plane, but the point-feature interaction matrix still requires a depth premise. In this work, all compared controllers use the same runtime scalar depth estimate, so that the comparison isolates the effect of bounded scalar variation in 1/Z. Three controllers are evaluated on a UR5 eye-in-hand platform: classical proportional IBVS with an online scalar depth estimate, a Uniform Design Tensor Product model predictive controller (UD–TP MPC) restricted to the 1/Z axis, and the proposed Takagi–Sugeno Parallel Distributed Compensation (TS–PDC) controller. The TS–PDC design represents 1/Z through a two-rule sector nonlinearity and precomputes the feedback gains off-line from linear matrix inequalities (LMIs), which certify quadratic stability of the reduced frozen-feature design model over the specified depth range. Real-robot results show that the proposed TS–PDC achieves faster transient error decay with negligible online computation, while avoiding both the online pseudo-inverse update of classical IBVS and the online semidefinite program (SDP) solved by UD–TP MPC.
Download
Paper Nr: 89
Title:

A Dual-Stream Transformer Architecture for Illumination-Invariant TIR-LiDAR Person Tracking
Authors:

Yuki Minase, Kanji Tanaka, Changjun Wang and Tomoe Hiroki
Abstract: Robust person tracking is a critical capability for autonomous mobile robots operating in diverse and unpredictable environments. While RGB-D tracking has shown high precision, its performance severely degrades under challenging illumination conditions, such as total darkness or intense backlighting. To achieve all-weather robustness, this paper proposes a novel Thermal-Infrared and Depth (TIR-D) tracking architecture that leverages the standard sensor suite of SLAM-capable robots, namely LiDAR and TIR cameras. A major challenge in TIR-D tracking is the scarcity of annotated multi-modal datasets. To address this, we introduce a sequential knowledge transfer strategy that evolves structural priors from a large-scale thermal-trained model into the TIR-D domain. By employing a differential learning rate strategy-referred to as “Fine-grained Differential Learning Rate Strategy”-we effectively preserve pre-trained feature extraction capabilities while enabling rapid adaptation to geometric depth cues. Experimental results demonstrate that our proposed TIR-D tracker achieves superior performance, with an Average Overlap (AO) of 0.700 and a Success Rate (SR) of 58.7%, significantly outperforming conventional RGB-transfer and single-modality baselines. Our approach provides a practical and resource-efficient solution for robust human-following in all-weather robotics applications.
Download
Paper Nr: 94
Title:

Model-Based Switching Manifold Embedded Terminal Second-Order Sliding Mode Control for Rehabilitation Exoskeletons
Authors:

Brahim Brahmi, Wissem Dilmi, Jawhar Ghommam and Mohammad H. Rahman
Abstract: This paper presents a nonlinear terminal second-order sliding mode control (SOSM–NTSM) strategy for precise trajectory tracking of a 2-DoF upper-limb exoskeleton. By embedding the robot dynamic model into the switching manifold, the proposed approach simplifies the torque control structure and mitigates inter-joint chattering effects. The nonlinear terminal term ensures finite-time convergence of tracking errors while improving transient response and disturbance rejection. Compared to classical sliding mode control, the proposed scheme generates smoother control inputs, significantly reducing chattering and actuator stress under modeling uncertainties. Validation on the iTbot exoskeleton demonstrates improved tracking accuracy, reduced torque chattering, and enhanced robustness, making the method suitable for safe and reliable rehabilitation applications.
Download
Paper Nr: 97
Title:

Practical Feature-Based 3D Object Recognition for Complex-Shaped Sand Cores in Casting
Authors:

Cong Huy Nguyen and Yoshiyuki Noda
Abstract: This study presents a practical 3D object recognition system for automating the core-setting operation in the casting process. The system targets high-mix, low-volume (HMLV) environments, where products change frequently. Feature-based matching is well suited to such requirements, as it operates using only a target model without retraining. The pose of a sand core is estimated by comparing geometric features computed from the object surface. In actual industrial environments, however, this approach suffers from matching ambiguity. This ambiguity arises from differences in the visible region between the reference model and the scene obtained from a vision sensor. To address this ambiguity, two approaches are proposed and integrated into a baseline feature matching pipeline. In addition, a practical normal reorientation method is introduced for robust feature matching. Experiments using 3D-printed objects show that the proposed method achieves a high recognition rate with reduced runtime. These results indicate its practicality for deployment in actual production environments.
Download
Paper Nr: 107
Title:

LoRa-Based Large-Scale Indoor Localization Using Probabilistic Models of Strongest Received Signals
Authors:

J. Ranz-Abad, S. Maldonado-Bascón, S. Lafuente-Arrroyo and R. Flor-Rabadan
Abstract: LoRa communication technology has been employed for indoor localization in a complex large-scale environment covering an area of approximately 100×100 m2 . The proposed method estimates the robot position using a probabilistic model built from the emitters with the strongest received signal strength indicator (RSSI) measurements. An offline sampling phase is first performed to estimate mean distance and variance pair associated with each emitter for a given RSSI measurement. A localization probability distribution is estimated for each of the strongest emmiters. These resulting distributions are then combined to obtain a global probability map, whose centroid is used as the position estimate. Two transmitter deployment configurations were evaluated in the experiments: the first featured a dense emitter distribution within the working area, while the second provided a more uniform spatial distribution. A third scenario is analyzed from a reference work. The experimental results show that emitters located at larger distances contribute negligibly to the localization process and can therefore be excluded from the estimation procedure. The proposed approach achieved a mean localization error of approximately 3.7 m in the high-density configuration and 9.5 m in the low-density configuration. These results demonstrate the feasibility of exploiting existing LoRa-based IoT infrastructures to provide coarse localization in large-scale indoor environments
Download
Paper Nr: 111
Title:

Design and Experimental Validation of a 3-DOF Delta Robot with Trajectory Tracking Control
Authors:

Pablo Rivadeneira, Viviana Moya, Juan Pablo Vásconez, Luis A. Córdova-Tandazo, Angélica Quito and William Chamorro
Abstract: This work presents the development and implementation of a three-degree-of-freedom Delta-type parallel robot. The project includes the mechatronic design of the manipulator, where the geometric parameters, links, and joints are dimensioned according to the required workspace and the robot kinematic model. The workspace is validated through the analysis of geometric and kinematic singularities to define a safe and achievable operating region. The control architecture is implemented in ROS2, integrating the actuators and control system through dedicated nodes for kinematic control and trajectory tracking in three-dimensional space. The system uses the direct kinematic model and Jacobian formulation to evaluate the performance of proportional-integral and sliding mode controllers. This work provides the foundation for future applications of the platform in handling tasks such as pick-and-place operations.
Download
Paper Nr: 113
Title:

Comparative Evaluation of Path Planning Algorithms and Control Strategies for Omnidirectional Mobile Robot Navigation
Authors:

Mateo García, Viviana Moya, Andrea Pilco, Juan Pablo Vásconez, Álvaro Prado-Romo and William Chamorro
Abstract: This paper presents a comparative evaluation of three local path planning algorithms: Rapidly-exploring Random Tree (RRT), Dijkstra, and D* Lite, combined with three kinematic trajectory controllers: PI Classic (PI-C), Proportional-Lyapunov (P-LY), and Proportional Null-Space (P-NS) for reactive obstacle avoidance on an omnidirectional mecanum-wheeled robot. All nine planner controller combinations are deployed on a real JetAuto platform using a 2D LiDAR sensor for obstacle detection, with a static global map and a reactive local layer activated when unknown obstacles appear at runtime. Experiments in a 3×3 m2 indoor environment with two previously unmapped obstacles introduced at runtime show that the local planner is the dominant factor in tracking accuracy: RRT and Dijkstra achieve RMSE of 0.11–0.16 m while D* Lite yields 0.35–0.59 m under the same controllers. PI-C achieves the lowest RMSE, and Dijkstra+P-LY the lowest mean control effort among the accurate configurations.
Download
Paper Nr: 118
Title:

Comparative Analysis of Bio-Inspired Optimization Algorithms for Nonlinear PID Tuning of a Mobile Manipulator
Authors:

Pablo Proaño, Paulo Leica, Gabriela M. Andaluz, William Chamorro and Luis Morales
Abstract: This work presents a comparative study of optimization methods for tuning a nonlinear PID controller applied to trajectory tracking of a mobile manipulator. The controller combines a conventional PID structure with an error-dependent nonlinear gain, increasing tuning complexity through an additional parameter. Three bioinspired population-based algorithms, Particle Swarm Optimization, Genetic Algorithm, and Artificial Bee Colony, were compared with a deterministic SQP-based method implemented through MATLAB’s fmincon. The tuning process used a composite objective function combining the Integral of Absolute Error, Integral of Squared Error, and Total Variation of the control signal. A three-dimensional reference trajectory, composed of a square horizontal path and sinusoidal vertical motion, was used to evaluate tracking performance. All optimized controllers followed the desired trajectory. PSO obtained the lowest objective value (Φopt = 0.955), closely followed by GA (0.958). SQP was the fastest method and achieved the lowest IAE (0.293), but the highest TVu (0.531). Compared with SQP, PSO reduced TVu by 21.3% and Φopt by 7.0%. These results highlight the trade-off between tracking accuracy, control smoothness, and computational cost.
Download
Paper Nr: 122
Title:

Continuous-Time Set-Transformers for Asynchronous WiFi-IMU Indoor Localization
Authors:

Mohamed Bachar, Ilyass Abouelaziz and Yuehua Ding
Abstract: Indoor localization commonly combines WiFi received signal strength indicator (RSSI) fingerprints with inertial measurements. These streams are sampled asynchronously at different rates, so observations rarely align in time and may be missing or stale. Most methods resample onto a shared clock and process each modality in a separate branch. We instead present a continuous-time transformer that represents observations as an unordered set of timestamped tokens. Each token encodes its measurement, modality, and elapsed time since acquisition. A single attention block fuses information across modalities and time, avoiding resampling and modality-specific branches while remaining invariant to token order. During training, we randomly mask entire modalities and individual observations, so it operates under missing or outdated inputs. We validate each encoder on its own benchmark and evaluate the fused model on simulated and real datasets, including a cross-session dataset recorded on unseen days. Under real cross-session drift, the approach cuts the error of an open WiFi fingerprinting baseline by 57%, matches or exceeds pedestrian dead reckoning where a learned LSTM fusion baseline fails to transfer, and degrades gracefully under missing or stale observations. A single continuous-time transformer thus provides a practical framework for indoor localization from asynchronous sensors.
Download
Paper Nr: 122
Title:

Continuous-Time Set-Transformers for Asynchronous WiFi-IMU Indoor Localization
Authors:

Mohamed Bachar, Ilyass Abouelaziz and Yuehua Ding
Abstract: Indoor localization commonly combines WiFi received signal strength indicator (RSSI) fingerprints with inertial measurements. These streams are sampled asynchronously at different rates, so observations rarely align in time and may be missing or stale. Most methods resample onto a shared clock and process each modality in a separate branch. We instead present a continuous-time transformer that represents observations as an unordered set of timestamped tokens. Each token encodes its measurement, modality, and elapsed time since acquisition. A single attention block fuses information across modalities and time, avoiding resampling and modality-specific branches while remaining invariant to token order. During training, we randomly mask entire modalities and individual observations, so it operates under missing or outdated inputs. We validate each encoder on its own benchmark and evaluate the fused model on simulated and real datasets, including a cross-session dataset recorded on unseen days. Under real cross-session drift, the approach cuts the error of an open WiFi fingerprinting baseline by 57%, matches or exceeds pedestrian dead reckoning where a learned LSTM fusion baseline fails to transfer, and degrades gracefully under missing or stale observations. A single continuous-time transformer thus provides a practical framework for indoor localization from asynchronous sensors.
Download
Paper Nr: 130
Title:

Data-Efficient Pedestrian Multi-Object Tracking via Zero-Shot Detection and Lightweight DINOv3 Re-Identification
Authors:

Manu Gupta, Oliver Zeilerbauer, Mohammad Reza Mohebbi and Mario Doeller
Abstract: Pedestrian Multi-Object Tracking (MOT) remains challenging in crowded video because missed detections, camera motion, occlusion, and identity ambiguity jointly affect online association. Most high-performing MOT systems rely on pedestrian-tuned detectors and supervised Re-Identification (ReID) models, which increases annotation requirements and obscures whether detection, motion, association, or appearance is the dominant bottleneck. We investigate a data-efficient alternative based on zero-shot detection and lightweight self-supervised appearance adaptation. We combine a D-FINE detector pretrained on Objects365 and COCO, without pedestrian-specific fine-tuning, with DINOv3-based ReID in a modular tracking-by-detection framework. The tracker integrates Kalman prediction, affine Camera Motion Compensation (CMC), cascaded Hungarian association, and track lifecycle management, with each component separately measurable. On the MOT17 half-validation split, controlled ablations show that detection quality sets the main performance ceiling: replacing zero-shot D-FINE with a MOT17-finetuned YOLOX-X raises Higher Order Tracking Accuracy (HOTA) from 47.2 to 69.9 under the same tracker with raw DINOv3 features. Appearance cues are useful only when identity-discriminative: raw DINOv3 can degrade association, whereas a 99K-parameter attention head on frozen DINOv3 tokens recovers most of the supervised ReID gap. The results indicate that data-efficient pedestrian MOT should prioritize robust detection and use adapted appearance features as a complement to motion and geometry.
Download
Paper Nr: 156
Title:

Safe Trajectory Planning and Control via Flatness-Based MPC with Convex Polytopic Constraints
Authors:

Solal Wargon, Harendra Rangaradjou, Christophe Louembet, Martin Mujica, Adrien Durand-Petiteville and Viviane Cadenat
Abstract: This paper proposes a safe trajectory planning and control framework for an Ackermann-type mobile robot navigating in cluttered static environments represented by polygonal obstacles. The approach combines differential flatness with model predictive control (MPC) and geometric convexification of the free space. By exploiting flatness, the predictive control problem is formulated directly in the flat-output space, allowing dynamic feasibility to be enforced through an inverse mapping while maintaining computational efficiency. To handle obstacle avoidance, convex safe polygons are constructed analytically from polygonal obstacles using an ellipse-based expansion procedure. The resulting connected convex regions are embedded into the MPC constraints, ensuring consistency between geometric safety and dynamic limits. A polynomial replanning step is introduced to maintain sufficient clearance within each region before tracking via an MPC controller. Simulation results demonstrate that polygon generation and MPC resolution times remain compatible with real-time operation under disturbances, while experiments on a real Ackermann-type mobile robot further validate the feasibility and tracking performance of the proposed framework.
Download
Paper Nr: 161
Title:

Indoor SLAM and Waypoint Navigation for a Low-Cost Mobile Robot: A Heterogeneous FPGA and Embedded GPU Architecture with BLE-Assisted Localization
Authors:

Tudor Mihai Crihălmeanu, Mircea Paul Mureşan and Sergiu Nedevschi
Abstract: Reliable indoor localization, mapping and navigation are often achieved with LiDAR or stereo systems, whose performance comes with additional cost, power and integration requirements. This paper presents a low-cost indoor Simultaneous Localization and Mapping (SLAM) and waypoint-navigation platform divided between a Cora Z7 and a Jetson Orin Nano. The contribution is an integrated heterogeneous architecture: deterministic motion execution and neural ranging operate in the Zynq programmable logic, BLE management, confidence fusion and Extended Kalman Filter (EKF) localization execute in the ARM processing system, and monocular perception, RTAB-Map mapping and A* planning run on the Jetson. An 8-bit RSSI-to-distance estimator is deployed through FINN and combined with an analytical path-loss estimate and per-beacon temporal filtering. The system was evaluated through a held-out neural-model test, a separate static four-beacon ranging experiment, post-synthesis accelerator analysis, monocular metric-depth tests, localization-trajectory comparisons, offline path-planning tests and representative end-to-end waypoint-navigation scenarios.
Download
Short Papers
Paper Nr: 12
Title:

A Novel Piecewise Autonomous Multi-Model Odometry Approach
Authors:

Andrea Righetti, Giuseppe Spampinato, Arcangelo Bruna and Giorgio Battistelli
Abstract: State estimation in multi-modal systems is a well-studied problem, particularly when system dynamics change, requiring different models for optimal performance. In such cases, running multiple filters in parallel and fusing their outputs can improve estimation accuracy. In this work, we propose a simple and innovative multi-filter approach, designed for a small differential-drive robot. The proposed method leverages low-cost IMU and wheel encoders to perform odometry, while compensating for slipping and disturbances during rotational maneuvers. The primary goal is to develop an algorithm that balances accuracy with computational efficiency, making it suitable for real-time deployment on low-cost microcontrollers.
Download
Paper Nr: 16
Title:

Liquids Detection Using SWIR Data and Deep Learning
Authors:

Antonio Buemi, Andrea Accettola, Arcangelo Ranieri Bruna, Cosimo Ieracitano, Nadia Mammone, Massimo Merenda and Valeria Tomaselli
Abstract: This paper introduces a novel approach for real-time liquid detection on floors using a Short-Wave InfraRed (SWIR) sensor equipped with a Band-Pass filters, specifically at 1380 nm wavelengths. This filter technique enhances the sensor’s ability to capture SWIR data, which is critical for identifying liquids. Coupled with a reduced YOLOv8n model, the proposed solution is designed to be applied in common use cases, e.g. it could be integrated into vacuum cleaners, enabling them to avoid liquids and improve household cleaning processes. We used a SWIR sensor prototype by STMicroelectronics to create a comprehensive SWIR image database, then we performed a fine-tuning of a reduced YOLOv8n model to suit the computational constraints of consumer-grade hardware. Our experimental results demonstrate the model’s effectiveness in detecting liquids on different floor types, underscoring its potential for practical application in robotics.
Download
Paper Nr: 23
Title:

Deep Global Point Cloud Registration for Long-Term Robot Localization
Authors:

Míriam Máximo, Enrique Heredia-Aguado, Esther González-Amorós, Francisco José Martínez-Peral, Luis Payá, David Valiente and Mónica Ballesta
Abstract: Accurate localization and orientation are fundamental prerequisites for the autonomous navigation of mobile robots and vehicles. To this end, this paper evaluates a method capable of estimating a robot's pose with six degrees of freedom (6DoF) by comparing a real time LiDAR scan with a previously acquired scan of the same environment. Our approach relies on establishing local feature correspondences between the respective point clouds. These local features are extracted via a deep learning architecture originally trained to generate a global descriptor for the entire scan. This study analyses the performance of the model's intermediate layers when repurposed for this task. The analysis is conducted using standard state of the art datasets, supplemented by a proprietary collection of captured scans.
Download
Paper Nr: 26
Title:

A Deep Learning Framework for Underwater Perception and Navigation Using Super-Resolution Enhanced YOLOv8 and Reinforcement Learning
Authors:

Khaled Benkhoud, Saliha Sebbane and Sarra Bouzayane
Abstract: This work presents an autonomous perception and navigation framework for underwater waste detection and classification. A dedicated underwater dataset was enhanced using a Super-Resolution (SSR) model to improve the quality of degraded underwater imagery. Several object detection architectures, including YOLOv5, YOLOv8, Faster R-CNN, and EfficientDet, were trained and evaluated. Experimental results show that the proposed SSR + YOLOv8 framework achieves the best balance between accuracy and real-time performance, reaching 90.0% precision, 90.9% mAP@50, and 80 FPS. In parallel, a Q-Learning–based navigation agent was developed and integrated into a Webots simulation environment to support autonomous underwater navigation. The results demonstrate the effectiveness of SSR for underwater perception and confirm the feasibility of combining deep learning–based detection with autonomous robotic navigation for marine debris management.
Download
Paper Nr: 28
Title:

Traversability-Aware Frontier Selection for Goal-Directed Exploration: An Ablation Study
Authors:

Carlos Tampier Cotoras, Frank Höller and Dirk Schulz
Abstract: Goal-directed exploration in unknown environments requires selecting frontiers that balance information gain with progress toward a distant target. Although existing methods combine distance metrics and local map information, they typically ignore terrain characteristics or rely on simple Euclidean approximations. This paper presents a comprehensive ablation study of frontier selection criteria for graph-based exploration planners, systematically evaluating the use of Euclidean versus geodesic distances, local map information, frontier size, and traversability-weighted path costs. Our experiments in urban simulation across three scenarios with different local minima characteristics reveal that geodesic distances with traversability weighting substantially improve performance in challenging environments, while frontier size provides no measurable benefit and can even degrade performance. Traversability-aware methods achieve a consistent 41–44% reduction in exploration time compared to geodesic approaches without traversability, which in turn outperform Euclidean baselines in complex scenarios by up to 45%.
Download
Paper Nr: 29
Title:

Workspace Computation via Constrained Forward Kinematics: Application to Robotic Manipulators
Authors:

Francisco José Martínez-Peral, Míriam Máximo, Mónica Ballesta, José Vicente Segura-Heras, Arturo Gil and Carlos Perez-Vidal
Abstract: The calculation of the workspace is essential in the design and deployment of robotic manipulation systems. Classical methods based on direct kinematics and the Denavit-Hartenberg formulation allow it to be estimated efficiently. However, these methods tend to overestimate it by including configurations that are kinematically valid in terms of position but do not meet the orientation requirements of the end-effector in practical applications. On the other hand, inverse kinematics algorithms offer more realistic estimations, although at a higher computational cost. In this work, we compare three methods for its calculation: (i) unconstrained forward kinematics, (ii) forward kinematics with end-effector orientation constraints, and (iii) reachability evaluation using TRAC-IK. The experimental results demonstrate that restricted forward kinematics approximates the TRAC-IK workspace with high accuracy, but at a fraction of the computational cost, positioning it as a practical alternative for fast and realistic estimates in robotic manipulation.
Download
Paper Nr: 33
Title:

Influence of FEA Modeling Parameters on the Accuracy of Robot-Link Compliance Identification
Authors:

David Eskoundos, Anatol Pashkevich, Kevin Subrin and Stéphane Caro
Abstract: The paper deals with robot link compliance matrix identification of size 6×6 from FEA virtual experiments. It presents an assessment of how FEA modeling parameters affect the accuracy of this matrix and evaluates this accuracy using a cantilever beam benchmark and its analytical Timoshenko compliance model as a reference. Three groups of parameters are studied: global discretization level, local discretization level for the virtual sensor region, and FEA element type with its interpolation order. The analysis is quantified by the Frobenius matrix norm error, the reconstructed displacement field error, and the corresponding computational cost. It was found that using moderate global meshes with a number of virtual sensor nodes between 10 and 100 gives the best compromise between the accuracy and the computational cost. If achieving this discretization level is impractical, local mesh refinement of the sensor region ensures the desired accuracy without increasing the computational time. However, using a fine local mesh for moderate global discretization gives almost no benefit. It was also found that using quadratic FEA elements such as HEX20 offers the most accurate results. However, TET10 provides a reliable and practical alternative if HEX20 is not available.
Download
Paper Nr: 35
Title:

A Certification-Based Predefined Time for Dynamic Average Consensus for Multi-Agent Systems
Authors:

Muzammil Ali, Muhammad Sabih, Khubaib Ahmad and Aamer Iqbal Bhatti
Abstract: This work addresses the predefined-time first-order exact dynamic consensus problem for multi-agent systems with time-varying local reference signals. Classical exact dynamic consensus (EDC) algorithms achieve zero steady-state tracking error and enable derivative estimation but do not provide guarantees on user-assigned convergence time. To overcome this limitation, a predefined-time EDC framework is proposed, ensuring that all agents recover both the average signal and its first derivative within a designer-specified time bound independent of initial conditions. The proposed approach preserves the single scalar communication structure of EDC and is formulated using a network-compatible predefined-time correction-pair concept. For implementation, a bi-power correction law inspired by fixed-time and sliding-mode design is introduced. Numerical simulations under large initial disagreement demonstrate that the proposed method achieves accurate predefined-time convergence and significantly outperforms classical linear dynamic consensus and standard EDC schemes.
Download
Paper Nr: 36
Title:

Automatic Labeling for Low-Light Pedestrian Detection
Authors:

Dimitrios Bouzoulas, Eerik Alamikkotervo and Risto Ojala
Abstract: Pedestrian detection in RGB images is a key task in pedestrian safety, as the most common sensor in autonomous vehicles and advanced driver assistance systems is the RGB camera. Low-light pedestrian detection lacks large public datasets and autolabeling pipelines. This research proposes a solution in the form of an automated infrared-RGB pipeline. The pipeline consists of 1) infrared detection, where a fine-tuned model for infrared pedestrian detection is used; 2) label transfer process from the infrared detections to their RGB counterparts; 3) training object detection models using the generated labels for low-light RGB pedestrian detection. The research was performed using the KAIST dataset. For evaluation, three object detection models, DETR, YOLO, and RCNN, were trained on generated and ground-truth labels. When compared on previously unseen images, the results showed that the models trained on generated labels outperformed the ones trained on ground-truth in 5 out of 6 cases for the mAP@50 and LAMP metrics, and outperformed ground-truth on mAP@50-95 in all cases. Acquired results indicate that the proposed autolabeling pipeline could be used for scalable annotation of low-light datasets for pedestrian detection. The source code for this research is available on GitHub (link).
Download
Paper Nr: 42
Title:

A Bounded Circle Covering Strategy for Mobile Robotics Inspired by Nesting Optimization
Authors:

Stépane Binon, Lucas Si Larbi and Eric Lucet
Abstract: Environment modeling in mobile robotics commonly relies on grid-based costmaps. This discretized representation of the environment is suitable for various motion planning approaches. Formal approaches based on optimization processes may require a geometric representation of obstacles in order to reduce the number of constraints while simplifying their representation. However, covering obstacles with simple geometric shapes often leads to overestimation, making navigation difficult in narrow passages. This paper proposes an overlay approach based on bounded circles to cover obstacles in a 2D grid-based costmap. Adapted from nesting optimization, this method defines the upper bound of obstacle overestimation. Thus, the number of covering circles and the degree of overestimation can be adjusted according to the situation. Furthermore, this approach is based on a conventional grid-based costmap, ensuring seamless compatibility with existing navigation pipelines. The approach based on bounded circles is first described, then tested through simulation scenarios and experiments.
Download
Paper Nr: 45
Title:

Redundant Safety Mechanisms of Artificial Intelligence Models for Autonomous Driving
Authors:

Florence Wagner, Heinrich Litagin and Rolf Jung
Abstract: Functional Safety is a standard well known and applied to common vehicles. Since artificial intelligence (AI) is introduced, with the intention to enable autonomous driving, ensuring road safety is becoming more difficult. It follows that the functional safety standard is insufficient for AI safety and new standards are introduced. This paper addresses the safety mechanisms of AI drafted in ISO/PAS 8800 in the context of an autonomous driving function. In order to advance the research and innovation in autonomous driving, applicability and benefits of safety standards must be established. In this case the reliance on safety mechanisms is examined. Redundancy of AI models is the safety mechanism of interest for this paper. Experiments are conducted where camera and/or LIDAR sensor data are used for object detection. In the experiment investigating the effectiveness of three different safety mechanisms, AI models are tasked with detecting objects in parallel, based on regular and manipulated or degraded data. This paper contributes to understanding the importance of selecting the appropriate safety mechanisms for the given use case. Diverse Redundant Models, Model Ensembles and N-Version Diverse Programming are mechanisms tested and compared in this paper regarding object detection for autonomous driving, leading to the conclusion that Diverse Redundant Models can have a significant impact on road safety.
Download
Paper Nr: 53
Title:

Multi-Arm Motion Planning Based on Conflict Elimination
Authors:

Matouš Kulhan and Pavel Surynek
Abstract: Motion planning for simultaneous movement of multiple robotic arms that interact spatially is a challenging problem in robotics. It is usually solved by planning in a composite configuration space with a high dimensionality. However, this approach can lead to a low success rate of motion planning with increasing number of arms. In this paper, we study the feasibility of using conflict elimination instead, where each motion planning step is done only in the configuration space of a single arm and uses a state-of-the-art sampling-based algorithm. To accomplish this, we adapt techniques from the related field of multi-agent path finding (MAPF), specifically we reformulated the CCBS algorithm from MAPF. We propose and implement a novel algorithm MP-CCBS that solves the continuous-time multi-arm motion planning problem. We evaluate the MP-CCBS algorithm on a diverse set of planning scenarios with 2-8 robotic arms. We compare the results to composite planning using state-of-the-art motion planning algorithm RRT-Connect. Except for the easiest scenario with 2 robotic arms, MP-CCBS outperforms composite planning by large margins in terms of success rate.
Download
Paper Nr: 54
Title:

Liquid Recurrent Reinforcement Learning for Constrained Multi-UAV 3D Coverage Path Planning
Authors:

Mohammad Reza Mohebbi, Mario Döller and Nauman Ahmed
Abstract: Coverage path planning (CPP) for Unmanned Aerial Vehicles (UAVs) in three-dimensional environments is challenging because agents must maximize target-area exploration while avoiding terrain, no-fly zones, collisions, and movement-budget violations. This problem is critical for inspection, monitoring, and search missions, where incomplete coverage or unsafe return can reduce mission reliability. This paper proposes a liquid recurrent deep reinforcement learning (DRL) framework for constrained UAV coverage. The task is formulated as a partially observable multi-agent reinforcement learning problem and optimized using Proximal Policy Optimization (PPO). The proposed policy combines Liquid Neural Network (LNN)-based temporal decision-making with spatial-temporal features, including local 3D voxel observations, terrain and target-area descriptors, uncovered-region cues, return-distance estimates, and budget information. To assess architectural design choices, we compare LNN and Long Short-Term Memory (LSTM) recurrent policies with and without attention mechanisms and engineered features. Experiments are conducted on procedurally generated 3D terrains, fixed random-seed evaluations, and real-world Passau terrain patches. Results show that engineered spatial-temporal features improve coverage and generalization, while attention provides limited benefit in the low-agent setting and may reduce stability. The best configurations achieve above 97% coverage while maintaining safe-return behavior. These findings establish liquid recurrent DRL as a promising direction for multi-UAV 3D CPP.
Download
Paper Nr: 74
Title:

Hamiltonian-Based Human Intention Estimation for Compliant Exoskeleton Robot
Authors:

Brahim Brahmi, Abderraouf Maoudj, Jawhar Ghommam and Mohammad H. Rahman
Abstract: Accurate estimation of a user’s motion intention is essential for achieving compliant, transparent, and effective assistance in upper-limb rehabilitation robots. However, many existing approaches are affected by measurement noise, non-causal behavior, and sensitivity to inter-user variability. To address these limitations, this paper presents a Hamiltonian-based estimator for reconstructing the Desired Intention of Motion (DIM). The proposed method combines an energy-consistent port-Hamiltonian framework with a Riccati-based adaptation law, ensuring passivity and robust convergence even under external disturbances. Experimental validation on a 7-DOF exoskeleton robot demonstrates accurate tracking of the user’s intended motion in both Cartesian and joint spaces, producing smooth trajectories with low estimation errors. The results indicate that the proposed estimator offers a reliable and dynamically consistent alternative to conventional methods, thereby enhancing transparency, user comfort, and operational safety in rehabilitation robotics.
Download
Paper Nr: 83
Title:

A Pose-Based AI Coach for Exercise Recognition and Quality Assessment
Authors:

Ahmed Aly, Al Amir Eldomiatty, Zeyad Elsayed and Walid Gomaa
Abstract: This paper presents an AI Virtual Coach system for automated exercise recognition and pose-based form assessment using consumer-grade RGB video. The proposed multi-stage pipeline integrates an exercise recognition module, a per-repetition assessment module, and a coaching agent. The recognition module employs systematic feature engineering, evolving from 19 base biomechanical features to 37 view-specific specialized features to resolve persistent confusion clusters, achieving macro F1 scores of 90.36% (side view) and 86.94% (front view) on a new dataset of 15 resistance exercises performed by 51 volunteers. The assessment module segments repetitions, encodes them with a compact temporal CNN, and predicts five aspect-level scores (0–10) via subject-level aggregation, yielding an overall MAE of approximately 3.5. A coaching agent synthesizes these scores with rule-based trend analysis and a large language model to generate personalized, actionable feedback. Evaluated under a subject-disjoint, multi-run protocol, the system demonstrates that lightweight pose-based models can provide practical, scalable exercise coaching from smartphone video, contributing a reproducible framework and dataset to the field of fitness AI.
Download
Paper Nr: 88
Title:

From Reactive to Map-Based AI: Tuned Local LLMs for Semantic Zone Inference in Object-Goal Navigation
Authors:

Yudai Noda, Kanji Tanaka, Sosi Tanaka, Changjun Wang and Tomoe Hiroki
Abstract: Object-Goal Navigation (ObjectNav) requires an agent to find and navigate to a target object category in unknown environments. While recent Large Language Model (LLM)-based agents exhibit zero-shot reasoning, they often rely on a “reactive" paradigm that lacks explicit spatial memory, leading to redundant exploration and myopic behaviors. To address these limitations, we propose a transition from reactive AI to “Map-Based AI" by integrating LLM-based semantic inference with a hybrid topological-grid mapping system. Our framework employs a fine-tuned Llama-2 model via Low-Rank Adaptation (LoRA) to infer semantic zone categories and target existence probabilities from verbalized object observations. In this study, a “zone" is defined as a functional area described by the set of observed objects, providing crucial semantic co-occurrence cues for finding the target. This semantic information is integrated into a topological graph, enabling the agent to prioritize high-probability areas and perform systematic exploration via Traveling Salesman Problem (TSP) optimization. Evaluations in the AI2-THOR simulator demonstrate that our approach significantly outper-forms traditional frontier exploration and reactive LLM baselines, achieving a superior Progress weighted by Path Length (PPL).
Download
Paper Nr: 93
Title:

Vision-Based Excavator Boom Pose Estimation with Sensor-Fused Uncertainty-Aware Tracking
Authors:

Robin Vetsch, Lucas Falch, Christoph Würsch and Sebastien Gros
Abstract: The boom pose estimation of an excavator, or the pose estimation of a manipulator in general, is critical for automation in construction and robotics. It underpins machine control, safety monitoring or productivity analysis. We present a real-time vision-based pipeline that estimates and tracks the boom joint angles of a large-scale mobile-walking excavator from a single onboard monocular RGB camera. A YOLO11 segmentation model isolates the boom. A fine-tuned EfficientNet then estimates the four joint angles together with a per-joint uncertainty value. An extended Kalman filter (EKF) tracks the joints when the boom leaves the field of view (FOV) of the camera or the image becomes unreliable due to suboptimal lighting conditions or occlusion. The EKF uses a learned long short-term memory (LSTM) dynamics model that is driven by 17 machine signals such as joystick commands, valve outputs and hydraulic pressure values, together with the previous joint velocities. It consumes the learned uncertainties as its measurement and process noise. In four different real-world recordings the EKF output of the proposed pipeline reaches a root mean square error (RMSE) of 0.63◦for the boom, 1.15◦for the stick, 2.70◦for the bucket and 55.6 mm for the telescope joint.
Download
Paper Nr: 104
Title:

Monitoring Endangered Animals Using Blockchain and Artificial Vision
Authors:

Ludwin Lope Cala and Roseli Aparecida Francelin Romero
Abstract: Global biodiversity is significant not only for its intrinsic value, but also because it is essential for human wellbeing, providing ecosystem services such as water and air purification, pollination, and climate regulation. The existence of each species is crucial for maintaining the balance of natural ecosystems. The objective of this work is to develop a computational system that uses blockchain and computer vision technologies to contribute to wildlife preservation, with a specific focus on monitoring and protecting endangered species. The system should be embedded in a single-board computer (SBC) with limited computational resources. To address this constraint, a deep, lighter (”tiny”) architecture, miniDETR, is being proposed, with an effective training strategy based on knowledge distillation. Specifically, in this architecture, the DETR and YOLOv8 models act as teacher models, guiding the learning process. Through this two-teacher approach, knowledge from both architectures is distilled, resulting in the miniDETR model, which is the compact student model. A comparison was made between the YOLOv8, DETR, and miniDETR models, and the miniDETR model demonstrated robust performance in both animal detection and classification.
Download
Paper Nr: 108
Title:

Robotic Manipulator Motion Planning and Obstacle Avoidance Using Physics-Informed Reinforcement Learning via Artificial Potential Fields
Authors:

Edmond Samia, Alexandre Parant, Laurent Arcese and Yuehua Ding
Abstract: Traditional planning algorithms often struggle with high-dimensional spaces or local minima traps, while standard Deep Reinforcement Learning (DRL) approaches suffer from sparse rewards and lack physical safety guarantees. To bridge this gap, this paper proposes a Physics-Informed Reinforcement Learning (PIRL) framework that integrates Artificial Potential Fields (APF) with Proximal Policy Optimization (PPO) for the continuous control of an n-Degree-of-Freedom (DoF) robotic manipulator. Rather than using APF as a standalone deterministic controller, the proposed approach leverages the APF gradient as a dense reward-shaping mechanism. This hybrid architecture guides the agent toward its target while strictly avoiding obstacles, effectively overcoming the sparse reward problem and mitigating classical local minima. The proposed method is validated by a comprehensive 17-task benchmark suite of varying complexities in a simulated environment. Experimental results with a 6-DoF manipulator show that the APF-PPO agent reaches the target under deterministic evaluation in all 17 tasks, against 4 for a classical APF controller operating on the same potential field, generating collision-free trajectories and escaping the configurations in which that controller is trapped by local minima. A demo video is provided along with this paper (Available at: https://youtu.be/XZNDLkh75Go).
Download
Paper Nr: 108
Title:

Robotic Manipulator Motion Planning and Obstacle Avoidance Using Physics-Informed Reinforcement Learning via Artificial Potential Fields
Authors:

Edmond Samia, Alexandre Parant, Laurent Arcese and Yuehua Ding
Abstract: Traditional planning algorithms often struggle with high-dimensional spaces or local minima traps, while standard Deep Reinforcement Learning (DRL) approaches suffer from sparse rewards and lack physical safety guarantees. To bridge this gap, this paper proposes a Physics-Informed Reinforcement Learning (PIRL) framework that integrates Artificial Potential Fields (APF) with Proximal Policy Optimization (PPO) for the continuous control of an n-Degree-of-Freedom (DoF) robotic manipulator. Rather than using APF as a standalone deterministic controller, the proposed approach leverages the APF gradient as a dense reward-shaping mechanism. This hybrid architecture guides the agent toward its target while strictly avoiding obstacles, effectively overcoming the sparse reward problem and mitigating classical local minima. The proposed method is validated by a comprehensive 17-task benchmark suite of varying complexities in a simulated environment. Experimental results with a 6-DoF manipulator show that the APF-PPO agent reaches the target under deterministic evaluation in all 17 tasks, against 4 for a classical APF controller operating on the same potential field, generating collision-free trajectories and escaping the configurations in which that controller is trapped by local minima. A demo video is provided along with this paper (Available at: https://youtu.be/XZNDLkh75Go).
Download
Paper Nr: 110
Title:

GARL: Grasp-Aware Reinforcement Learning for Robust and Dexterous in-Hand Rotation
Authors:

Guillaume Thomas, Jaonary Rabarisoa, David Filliat, Quoc-Cuong Pham and Mathieu Grossard
Abstract: Dexterous multi-fingered hands on robotic arms are often subject to varying orientations and accelerations, yet most reinforcement learning (RL) methods for in-hand manipulation assume a fixed hand pose, limiting robustness and generalization. Our approach incorporates both object-related and hand-related considerations within a RL pipeline (Proximal Policy Optimization in our case) to achieve desired behaviors for dexterous manipulation tasks. It targets key aspects of multifingered grasps through quantitative measures of guaranteed stability, controllability, and manipulability. This enables a comprehensive analysis of the learned policies, characterizing the safety margin against undesired grasp configurations. It favors finger-gaiting sequences characterized by perturbation-invariant in-hand rotation with up to 12× fewer training steps and improves generalization to unseen perturbation intensities by 25×. The approach also enhances task performance by 18.1% and reduces average joint torques by 26.7%. Object generalization experiments demonstrate that our method provides consistent and reliable training outcomes. These simulation-only results bring the required robustness for future real-world deployment on humanoid robots. Code and Videos are available on our website (https://whojo.github.io/GARL/index.html).
Download
Paper Nr: 121
Title:

Towards Single-Contact-Point Textile Unfolding with Hierarchical Point Selection and Trajectory Generation
Authors:

Santiago T. Puente and Carlos Mateo-Agulló
Abstract: This paper presents a hierarchical agent architecture for textile unfolding using a single robotic arm. The proposed approach is composed of two connected agents: a point agent that selects the grasping point and a delta agent that generates the trajectory to unfold the textile. The method takes the textile Point-Cloud (PC) as input and processes it through a frozen Dynamic Graph Convolutional Neural Network (DGCNN) backbone to extract local and global geometric features. The point agent uses these features to select a suitable grasping point, while the delta agent combines the selected point with the global representation to generate a sequence of displacements defining the unfolding trajectory. In addition, a reward function is designed to evaluate the unfolding state of the textile and to guide the learning process. The agents are trained with different parameter configurations, including the number of input points, the number of trajectory displacements, and the displacement scale. The experimental results show that the proposed method achieves more than 95% positive unfolding actions, indicating its potential as a starting point for single-arm textile unfolding.
Download
Paper Nr: 124
Title:

Scene Description Generation for Traffic Scenes Using Segmentation and Visual Language Models
Authors:

Boitor Diana, Sergiu Nedevschi and Raluca Didona Brehar
Abstract: There is an increase of live-stream monitoring places in which surveillance cameras capture the traffic flows and incidents in urban environments. These places provide a huge amount of visual data that forms the basis of the proposed scene description generation model proposed in this paper. The paper presents a two-stage pipeline in which three segmentation models - ERFNet, YOLOv8-seg or SegFormer-B5 - are paired with vision-language models that generate a fluent natural-language description of the scene. Moondream2 and Qwen2-VL-2B have been explored alongside with three visual-prompting strategies. The appropriate strategy for each VLM was selected depending on whether the backbone accepts one or multiple images per prompt. We report quantitative results and qualitative observations on how the choice of the segmentation model propagates into the generated description. We also outline the application of the pipeline to a full-day traffic live-stream recording as a step toward continuous monitoring.
Download
Paper Nr: 128
Title:

Fault-Tolerant Multi-Sensor Fusion via Factor Graph Optimisation for Mobile Robot Localisation
Authors:

Clovis Adeler, Maan El Badaoui El Najjar, Cindy Cappelle and Denis Pomorski
Abstract: This paper presents a fault-tolerant multi-sensor fusion architecture for the localisation of an Unmanned Ground Vehicle (UGV) operating in outdoor environments. The proposed approach combines wheel odome-try, Inertial Measurement Unit (IMU), and Global Navigation Satellite System (GNSS) measurements within a Factor Graph optimisation (FGO) framework solved using a Gauss-Newton optimizer. Since real sensors are prone to faults that simulation-based studies cannot fully capture, a diagnostic layer is integrated into the fusion architecture to detect and exclude faulty measurements before they corrupt the estimated trajectory. Fault detection relies on a Mahalanobis distance test combined with a chi-square threshold to isolate inconsistent GNSS measurements. To validate the proposed architecture, real-world outdoor data are first collected using a ground robot, after which GNSS faults are artificially injected into the dataset through post-processing. Compared against a baseline FGO without the diagnostic layer, the proposed diagnostic layer successfully detects all injected offset faults across three diagnostic configurations, achieving detection latencies ranging from 0.0 s in the reactive configuration to 4.7 s in the conservative configuration, while maintaining an accurate and reliable position estimate despite the injected faults.
Download
Paper Nr: 139
Title:

Mapless Rover Navigation Leveraging Obstacle Perception with k-NN Vector-Space Clustering
Authors:

Giovanni De Gasperis, Sante Dino Facchini and Luigi Iaconi
Abstract: This paper presents a mapless navigation paradigm for mobile robots based on vector-space obstacle perception, extending our previous work on retrofitting a Pepper robot with a Redis-based flow-network architecture. To overcome the time and scalability limitations of human-guided data collection, we introduce a greedy autonomous exploration algorithm that eliminates human intervention while shifting the exploration objective from global mapping to local obstacle characterization. The idea was preliminarily validated simulating a navigating iRobot iCreate 3 platform in CoppeliaSim, with Redis serving a dual role as a real-time publish/subscribe message broker and as a vector search database, enabling cosine-similarity-based K-Nearest-Neighbor retrieval for navigation decisions. Training data were collected across six progressively complex virtual environments, allowing empirical tuning of exploration parameters governing freedom, novelty, and hysteresis. Results show that the proposed architecture eliminates the oscillatory behavior typical of greedy exploration, while achieving real-time responsiveness suitable for reactive obstacle avoidance on lightweight hardware.
Download
Paper Nr: 147
Title:

Testing Video Anomaly Detection through State Space Models on Challenging Real-World Surveillance Datasets
Authors:

Esther González-Amorós, Alberto Chueco, Míriam Máximo, Carlos Viegas, Adrián Peidró and Luis Paya
Abstract: This work presents the first comprehensive evaluation of the VADMamba framework on the UCF-Crime dataset, a challenging benchmark for video anomaly detection (VAD) in real-world surveillance environments. While VADMamba has demonstrated strong performance on controlled-environment benchmarks, its applicability to semantically complex surveillance scenarios remains unexplored. Through systematic experimentation, we demonstrate that a lightweight, prediction-only configuration utilizing a depth profile of [3,3,3,3] achieves an AUC (Area Under Curve) of 72.23% and 1.8 times fewer parameters and 2.6 times fewer FLOPs than the full original dual-task pipeline, by eliminating the optical flow reconstruction bottleneck. This performance surpasses the previous state-of-the-art one-class classification methods on UCF-Crime, despite training on only 260 videos of the available 800 normal videos. Cross-dataset transfer experiments reveal substantial domain gaps when models trained on controlled benchmarks are evaluated on UCF-Crime, highlighting the limited representativeness of widely-used datasets for real-world deployment.
Download
Paper Nr: 149
Title:

Comparative Analysis of Eye-in-Hand and Eye-to-Hand Motion Estimation Techniques for Dynamic Target Tracking in Visual Servoing
Authors:

Mario Talarico, Ibrahim Ben-Othman, Andrea Monguzzi, Navvab Kashiri, Andrea Maria Zanchettin, Elena De Momi and Paolo Rocco
Abstract: Dynamic target tracking is a fundamental capability for robotic systems operating in environments where the target undergoes unknown translational and rotational motion. In such scenarios, target dynamics introduce additional disturbances into the visual-servoing loop, making accurate motion estimation essential for maintaining tracking accuracy and control stability. This paper presents a comparative analysis of Eye-in-Hand and Eye-to-Hand motion-estimation strategies for dynamic visual servoing. The investigated approaches employ Kalman filter-based estimators to reconstruct the target motion and generate feedforward compensation terms integrated within Position-Based Visual Servoing (PBVS), Image-Based Visual Servoing (IBVS), and Two-and-a-Half-Dimensional Visual Servoing (2.5DVS) frameworks. The comparison focuses on the characteristics of image-based and pose-based estimation methods, as well as on their integration within different visual-servoing architectures. Simulation and experimental results demonstrate the advantages and limitations of the investigated approaches in terms of estimation accuracy, robustness, and tracking performance, providing practical guidelines for selecting suitable motion-estimation strategies in dynamic visual-servoing applications.
Download
Paper Nr: 154
Title:

Tether-Length-Constrained Reference Governance for Cooperative UGV-UAV Agricultural Inspection
Authors:

Ammar Selim, Zhengjia Xu, Mohsen Zahmatkesh and James E. Pickering
Abstract: This paper presents a unified control framework for agricultural inspection using a tethered unmanned ground vehicle–unmanned aerial vehicle (UGV–UAV) system. The framework integrates the UAV flight controller, UGV path-following controller, and tether management unit (TMU). A supervisory reference governor (RG) monitors the Euclidean distance between the UAV and the UGV and modifies infeasible horizontal UAV commands through tether-aware position-reference projection. The associated horizontal velocity and acceleration feedforward commands are scaled consistently with the projection factor. Without the RG, the nominal mission requires approximately 28 m of tether and is infeasible with the available 20 m tether. The RG constrains the commanded UAV reference, and the closed-loop simulation shows that the realised UAV–UGV Euclidean separation remains below this limit while preserving complete inspection coverage and satisfactory tracking. The tether-length-compliant governed mission is completed in approximately 450 s. The nominal unconstrained trajectory would require approximately 250 s; however, it violates the 20 m tether-length limit and is therefore physically infeasible. Consequently, the 250 s duration is used as an unconstrained nominal benchmark rather than a feasible baseline.
Download
Paper Nr: 155
Title:

Energy-Aware Adaptive Communication for Efficient Mobile Robots
Authors:

Giorgio Vivoli and Navvab Kashiri
Abstract: Wireless communication underpins the operation of mobile robotic systems, enabling sensor data exchange, remote supervision, and multi-robot coordination. Sustaining reliable links, however, incurs significant energy overhead which is a critical constraint for battery-powered platforms where communication competes directly with actuation and computation for limited power budgets. This paper introduces a modular, ROS2-native adaptive communication framework that jointly optimises transmission power, data rate, and message publication frequency to minimise radio current consumption while preserving link reliability. At each control cycle, real-time RSSI (Received Signal Strength Indicator) measurements are processed by a logistic model that predicts the packet reception rate (PRR) and identifies the energy-optimal parameter configuration. A priority-aware scheduling mechanism selectively reduces the publication rate of lower-priority topics under bandwidth constraints, preserving the integrity of critical data streams. Simulation and real-world experiments on a mobile robot platform demonstrate up to 65% reduction in current consumption relative to a maximum-power fixed configuration, with PRR sustained above 86%, demonstrating a tunable, environment-independent trade-off between energy efficiency and communication reliability.
Download
Paper Nr: 159
Title:

AIVFusion+3D Features: An Approach for Features Extraction Aiming to Implement Sensors Data Fusion from Monocular Camera and LIDAR
Authors:

Fernando S. Osorio, Felipe Freitas Maia, Alexandre M. Ishi Jr, Marcos Vinicius da Rocha, Vitor Augusto Pereira and Leticia de Carvalho Pereira
Abstract: This work presents a new approach for 3D features extraction and sensors data fusion based on 3D depth data. The proposed approach aims to match different layers of information extracted from 3D LIDAR data (3D Point Clouds), and RGB camera data (Monocular Depth Estimation), in order to facilitate and improve sensors data fusion (registration/match). We estimate depth from monocular images (RGB), and also, measure depth using laser sensors (LIDAR). This work was inspired by the AIVFusion approach (Pereira and Osório, 2025) - Artificial Intelligence Vision Fusion, seeks to identify references in multiple information layers obtained from the perception sensors: Monocular Cameras, LIDARs and also from other sensors. In this work, we expanded the AIVFusion. adding 3D features detectors/extraction approach (corner, edges), and performing tests considering simulated data obtained using CARLA Simulator, and real data obtained from LIDAR and monocular cameras. The results demonstrate we are able to track 3D Depth Robust Points and then seek to match them both from the image (depth estimation), and from the LIDAR depth data (depth measure). This approach can be applied to Intelligent Vehicles (aerial and terrestrial), mainly in localization, mapping and navigation tasks.
Download
Paper Nr: 164
Title:

Kalman Filter Fusion of Forward Kinematics and Vision for Robot Pose Estimation in Construction Environments
Authors:

Juan Pablo Cahueque Perez, Christine Chevallereau, Corentin Monet, Loïc Buatois and Andrea Cherubini
Abstract: Automating insulating panel installation for building facade renovation presents significant challenges due to sensor noise and the need for precise compliance with industrial standards. This paper introduces a hybrid vision/force control framework that addresses these issues through robust pose estimation and adaptive task optimization. A Kalman filter fuses vision data from a monocular camera with fiducial markers and proprioceptive encoder information, achieving up to 97.1% noise reduction in rotational pose components. The installation process is decomposed into a sequence of Quadratic Programming problems, where sensor feedback directly informs both vision-based tracking and force-based compliance tasks. This unified optimization approach leverages the complementarity between modalities to simplify control complexity. Experimental validation on a KUKA LWR robotic arm with prototype panels confirms compliance with industrial fastening requirements and demonstrates the feasibility of automated facade panel installation.
Download
Paper Nr: 37
Title:

KGAT-RL: Kernelized Graph Attention Networks with Drift-Triggered PPO Residual Correction for Robust Multi-Horizon Traffic Forecasting
Authors:

Mohammad Reza Mohebbi, Mohammad Javad Hassanzada, Erfan Ranjbar, Mohammad Soltanirad and Mario Döller
Abstract: Short-term traffic forecasting on sensor graphs is essential for Intelligent Transportation Systems (ITS), yet performance can degrade under noisy measurements, missing values, and non-stationary traffic regimes. This paper proposes KGAT-RL, a hybrid framework that combines a Kernel Graph Attention Network (KGAT) with a drift-triggered Reinforcement Learning (RL) residual-correction module. The KGAT backbone replaces softmax attention with learnable kernel similarity to stabilize message passing and preserve locality on irregular freeway networks. To address systematic bias under distribution shift, KGAT-RL activates a safety-gated Proximal Policy Optimization (PPO) policy only when persistent drift is detected from rolling horizon-specific residual statistics. The policy applies bounded residual actions, reducing harmful over-correction under uncertain or low-quality data conditions. The framework therefore combines stable graph-based forecasting with selective test-time correction without continuously adapting the base model. We evaluate speed and occupancy forecasting on Caltrans PeMS District 7 using 30 loop-detector stations from 2021, 5-minute sampling, and direct 15/30/60-minute horizons. At the 60-minute horizon, KGAT-RL improves speed RMSE from 7.01 to 6.89 and occupancy RMSE from 0.0552 to 0.0541. Results indicate that KGAT-RL offers a robust and drift-aware pipeline for multi-horizon traffic prediction.
Download
Paper Nr: 72
Title:

Robust Fractional-Order Sliding Mode Control for Trajectory Tracking of a Mobile Manipulator
Authors:

Gabriela M. Andaluz, Paulo Leica, Luis Morales, Pablo Proaño and Oscar Camacho
Abstract: This paper presents a robust Fractional-Order Sliding Mode Controller (FO-SMC) for trajectory tracking of a mobile manipulator. The control objective is trajectory tracking of the end-effector in task space rather than joint-space regulation. The proposed controller integrates a fractional-order term into a PID-type sliding surface to improve transient response and mitigate chattering inherent to classical Sliding Mode Control (SMC). Performance is evaluated on a lemniscate trajectory tracking task under reference changes and parametric variations. Compared to classical SMC, the proposed FO-SMC achieves reductions of 29.33% in IAE, 1.22% in ISE, and 7.05% in TVU, demonstrating smoother control signals and superior tracking accuracy. A Lyapunov-based stability analysis guarantees convergence and stability of the closed-loop system.
Download
Paper Nr: 92
Title:

Self-Supervised Appearance-Based in-Plane Rotation Estimation
Authors:

Sreejani Chatterjee, Nichols Taylor and Daniel Nikovski
Abstract: Establishing correspondences between keypoints of an object in different poses in multiple images captured by a camera is a key step in estimating its pose for the purpose of manipulation, material handling, and assembly in robotics. Full-range in-plane rotations in the image plane are particularly challenging, because the majority of keypoint detection, description, and matching methods are neither rotationally invariant nor covariant. In addition, the appearance of the object tends to undergo significant changes due to varying illumination and specular effects. We investigate the performance of four keypoint matching and two keypoint tracking methods on a representative dataset, and propose a novel memory-based method to improve matching under large in-plane rotations, where matching is done against a set of reference images whose relative poses are estimated in a self-supervised manner.
Download
Paper Nr: 95
Title:

Optimization of a Higher Order Sliding Mode Controller for Tactical Missile Guidance Applications
Authors:

Mirza Hodžić, Naser Prljača and Jakub Osmić
Abstract: The traditional missile guidance system design consists of two loops that are designed separately: guidance and control. The guidance loop generates commanding accelerations, while the control loop ensures that the commanded accelerations are tracked. The disadvantage of such a design is that each of these subcomponents is designed separately, and each has its own time constant. Integrated guidance and control (IGC) offers a unified approach to controller design, where guidance and control are considered as one entity with a single controller and a single time constant. This paper proposes IGC implementation using higher order sliding modes (HOSM) and higher order sliding mode observer and its optimization using surrogate optimization. The proposed algorithm is validated through 6-DOF simulations with evasive target.
Download
Paper Nr: 101
Title:

A Retrofit State Monitoring Framework for Safe Robot-Instrument Interaction in Legacy Analytical Laboratories
Authors:

Mohamed Ali Tlili, Thomas Roddelkopf, Heidi Fleischer and Kerstin Thurow
Abstract: Analytical laboratories increasingly employ collaborative robots to automate sample handling and instrument interaction. However, the integration of robotic systems with existing analytical equipment remains challenging because many legacy instruments do not provide access to safety-relevant operational states required for reliable robot coordination. As a result, robot actions cannot always be executed under deterministic operating conditions. This paper presents a retrofit state monitoring framework that supports deterministic robot–instrument interaction without modifications to instrument hardware, firmware, or proprietary control software. The framework consists of three layers: State Acquisition, State Abstraction, and Motion Authorization. Safety-relevant operational states are acquired through externally installed sensors, transformed into abstract device states, and evaluated to determine whether robot interaction is permitted. The framework was implemented and validated in an analytical workstation comprising two commercial analytical instruments with different access mechanisms and a collaborative robot for automated vial handling. Experimental results demonstrate that the proposed approach enables deterministic robot motion authorization based on externally monitored device states and provides a practical method for integrating legacy analytical instruments into robotic laboratory workflows.
Download
Paper Nr: 112
Title:

Optimization-Based Inverse Kinematics for a 4-DOF Robotic Arm: A Comparative Analysis with Real-Time ROS Implementation
Authors:

Jofre Suárez, Viviana Moya, Juan Pablo Vásconez, Álvaro Prado-Romo, Andrea Pilco and William Chamorro
Abstract: This work presents a real-time ROS-based implementation of inverse kinematics solvers for the Hiwonder ArmPi Mini, a 4-DOF servo-driven robotic arm controlled from a Raspberry Pi. The evaluated methods were fully self-implemented without external optimization libraries and include an unconstrained Newton–Raphson approach, a constrained SQP formulation using quasi-Newton BFGS and SR1 Hessian approximations, and a heuristic Particle Swarm Optimization method. All solvers use the same regularized objective function, which minimizes the Cartesian position error of the end effector while penalizing large joint values. The methods were experimentally assessed on a pick-and-place trajectory using the same kinematic model, ROS architecture, and actuation interface. The results show millimetric tracking accuracy, with RMSE values between 0.00138 m and 0.00225 m, and convergence rates close to or equal to 100%. The quasi-Newton methods provided the best accuracy–time compromise, with average computation times of approximately 7.5 ms, while PSO achieved the lowest average computation time of 2.22 ms. These results demonstrate that self-implemented optimization-based inverse kinematics can be executed at competitive real-time rates on a low-cost embedded robotic platform without joint feedback.
Download
Paper Nr: 129
Title:

TASC: Fast Skill Chaining for Complex Sequential Robotic Tasks via Pure Imitation Learning
Authors:

Ryan James Hayward and Kasim Terzić
Abstract: We present TASC: Termination Action Skill Chaining, a method for pure imitation learning of complex sequential long-horizon tasks, requiring only simple sub-task annotations and no environmental reward. We define a new termination action dimension to allow each sub-task’s termination function to be learned jointly with its policy, giving it the ability to end its sub-task. These policies are trained in parallel on segments of the demonstration, and then chained together at inference time, providing a massive speed-up over sequential training. We evaluate TASC on three complex robotic manipulation tasks and show improved success rates compared to state of the art methods.
Download
Paper Nr: 135
Title:

Electro-Mechanical Modelling of Generatively Designed Robotic Arm Segments with Integrated Wire Routing
Authors:

Johannes Lamm and Thomas M. Wendt
Abstract: Computational design and simulation tools shorten the path from an additive-manufacturing concept to industrial use. The workflow is demonstrated on a generatively designed, fully 3D-printed robotic arm segment by embedding a discrete copper wire in a load-bearing branch so that the branch simultaneously carries mechanical load and electrical power and signals. The feasibility is decided by two mechanical questions: whether the conductor with more than 34 times higher stiffness measurably alters the structural deflection, and whether the strain mismatch between the stiff wire and the compliant polymer matrix endangers the wire-matrix bond. Cross-sections are extracted from the CAD geometry, which yields the branch second moments of area and the wire eccentricity due to the printing process. A reduced parallel-branch model bounds the deflection, and a transformed-section analysis quantifies the stiffness contribution. The modulus mismatch instead concentrates shear in the adhesion layer. A finite-element model identifies the governing mechanism as transverse shear, largest on the neutral axis, together with a rising stress concentration where the wire crosses the outer surface. The interfacial stress, not the stiffness change, governs the mechanical design.
Download
Paper Nr: 153
Title:

Point Cloud Streaming from Mobile Robots under Bandwidth and Scene Complexity Constraints
Authors:

Mohamad El Dika, Jalal Possik, Petra Bilane, Jad Nassar, Roger Achkar and Charles Yaacoub
Abstract: Streaming high-resolution LiDAR point clouds from mobile robots is essential for remote perception and monitoring but remains challenging due to the large data volumes involved. This paper presents a comparative evaluation of point cloud compression techniques with a focus on rate-distortion performance and the influence of scene complexity. A ROS2-based simulation framework is developed to benchmark geometry-based compression (G-PCC) against projection-based approaches using image and video codecs (PNG, JPG, H.264, and H.265). The evaluation is conducted across environments with varying geometric characteristics, including planar, sharp, smooth, and complex scenes. Results reveal a clear trade-off between compression efficiency and reconstruction quality: G-PCC consistently achieves superior fidelity and robustness to scene variations, while projection-based methods offer higher compression at the expense of geometric accuracy, with video codecs providing the most favorable compromise. The study also demonstrates that scene complexity significantly impacts compression performance, particularly for projection-based pipelines. These findings provide practical insights into the behavior of compression strategies for real-time robotic applications and motivate the design of adaptive approaches tailored to scene characteristics.
Download
Paper Nr: 157
Title:

Proposal of an Intuitive Robot-Summoning Method for Agricultural Mobile Robots Using a Retroreflective Handheld Device
Authors:

Atsuhiro Ohno, Sho Yokota and Akihiro Matsumoto
Abstract: This study proposes a method that enables agricultural workers to summon a specific transport robot from among multiple robots with a single-button action. In agricultural fields where multiple workers are operating simultaneously, the summoned robot must also identify the worker who initiated the call. To achieve this, a worker points a handheld device equipped with a retroreflector toward the desired robot and presses a button. The button press broadcasts a simple wireless trigger signal. Upon receiving the signal, each robot examines the point cloud obtained from its 3D LiDAR and determines whether a highly reflective point cluster is present. Only the robot toward which the retroreflector is directed detects the high-intensity reflection and recognizes that it has been summoned. Furthermore, the position of the highly reflective point cluster allows the robot to estimate the caller's location and identify the worker who initiated the request. The main advantage of the proposed method is that robot calling and caller identification can be realized without using GNSS-based localization or inter-robot communication, resulting in a simple and lightweight system. Preliminary and verification experiments demonstrated the feasibility of the proposed approach and confirmed that both robot summoning and caller identification can be achieved using the proposed method.
Download
Paper Nr: 162
Title:

A Real-Time Framework for Multi-Person Reciprocal Line-of-Sight Modeling and Recovery in Human-Robot Interaction
Authors:

Jesus Eduardo Hermosilla-Diaz, Jeeangh Jennessi Reyes-Montiel, Yaser Aguas-Garcia and Antonio Marin-Hernandez
Abstract: In this paper, a robot reciprocal multi-person line-of-sight framework is presented. The proposed approach seeks to evaluate multi-users visual alignment in real time and execute autonomous attention recovery behaviors during multi-person Human-Robot Interaction (HRI). The framework is composed of a hierarchical process that performs: a facial landmark detection, 3D head pose estimation, and finally, a re-identification to perform identity association of multiple persons inside robot’s field of view. In this work, it is proposed a geometric model to represent the line-of-sight of a human’s field of view. Furthermore, a reciprocal line-of-sight score to quantify the human-robot degree of alignment, has been introduced. In this context, when a user’s focus decreases, the system identifies them as a target and triggers a recovery behavior. The robot response consists of predefined speech phrases followed by a Pan-Tilt Unit motion at the target’s face. The framework was tested in different configurations with multiple persons in order to analyze performance under different arrangements. The scoring process successfully detected predetermined shifts of line-of-sight loss and resolved overlapping target conditions by prioritizing the lowest reciprocal score. Consequently, visual alignment was restored with the behavior recovery, showcasing the potential of the framework to manage and recover visual focus in dynamic, multi-person scenarios.
Download
Paper Nr: 163
Title:

Implementation of an Autonomous Liquid Pouring System Based on Source Container Modeling for Different Target Containers
Authors:

Jeeangh Jennessi Reyes-Montiel, Jesus Eduardo Hermosilla-Diaz, Yaser Aguas-Garcia, Yahir Hernandez-Mella and Antonio Marin-Hernandez
Abstract: The problem of accurate partial liquid pouring still remains a challenging task for autonomous robotics systems, both in industrial and domestic environments. Particularly, when multiple pours are required, it is desirable to maximize the use of the liquid in the source container. The use of precise injectors apart from expensive are large and cumbersome, then not suitable for the most part of the systems. In this paper, to tackle the problem of multiple pours from same source container, it is presented the implementation of a system for regulated partial liquid pouring. Unlike most previous work, the proposed implementation enables multiple partial pours of varying liquid quantities, as our approach tracks and updates the estimated remaining liquid volume in the source container after each pouring operation. This allows for pouring different amounts into various destination containers in succession without needing to refill the source container. Proposed solution considers the geometry of the source container, rather to focalize on the destination container. To test and validate the approach, a simple vision module is used to estimate the initial volume of liquid in the target container as well as to validate the final volume after filling. Two sets of experiments were done, the first one consist of a simulated industrial filling quality control, where the system require to complete the missing liquid, and the second where the system pour the same quantity of liquid on diverse containers. The results show that the proposed implementation successfully performs partial liquid pouring with an average relative error of 4.44% ± 2.66\%, showing that the proposed approach provides an effective solution for autonomous partial liquid pouring.
Download

Area 4 - Signal Processing, Sensors, Systems Modelling and Control

Full Papers
Paper Nr: 18
Title:

Extended Neural Ordinary Differential Equations
Authors:

Leonard Johard and Vittorio Lippi
Abstract: Stiff systems of ordinary differential equations (ODEs) present significant challenges for standard Neural Ordinary Differential Equation (NODE) frameworks, often leading to training instabilities and prohibitive computational costs when using dynamic step solvers. In this work, we analyze the root causes of these instabilities and propose eNODE. Rather than learning a continuous vector field as in standard Neural ODEs, we learn a timestep-conditioned discrete transition operator that explicitly models the numerical integration step. By conditioning the neural network on the integration step size (∆t), eNODE explicitly learns the discrete transition operator, allowing it to internally compensate for discretization errors. We demonstrate that our approach enables the use of computationally efficient explicit solvers, such as forward Euler, to solve the highly stiff Robertson and Schäfer (HIRES) equations, tasks typically requiring expensive implicit schemes. Our results indicate improved stability and accuracy across a wide range of time steps, maintaining robustness even when extrapolating to step sizes significantly larger than those encountered during training. This method provides a surrogate modeling approach that enables stable explicit integration of stiff dynamical systems with negligible computational overhead.
Download
Paper Nr: 32
Title:

Robotic Sorting of Dirty Laundry Using a Multispectral Vision-Based Grasping Strategy
Authors:

Ferréol Gagey, Gabriel Prudhomme and Mathieu Grossard
Abstract: This paper addresses autonomous robotic grasping of clothing items arranged in random piles, with the goal of separating them into white, light, bright color, and dark categories. The proposed hardware system integrates a multispectral vision module used for computing optical metrics such as saturation and lightness scores, a bidigital gripper with specifically redesigned jaws, and a collaborative robotic arm. A dedicated grasping strategy is developed to select the most suitable blob for pickup, while a Support Vector Machine (SVM) classifier is used to assign the grasped garment to the appropriate category based on its optical fitness scores. Experimental results demonstrate a high classification accuracy of 95%, highlighting the effectiveness of the proposed vision-based perception and classification pipeline. Overall, the method shows strong potential for improving the performance of robotic systems for real-world laundry sorting applications.
Download
Paper Nr: 43
Title:

Sliding Mode Control of Single-Mode Photon by Nonlinear Quadrature Actuation
Authors:

Sergey V. Drakunov and Bereket Berhane
Abstract: This paper develops a nonlinear switched feedback for a single quantized electromagnetic-field mode in which the controlled variables are the first and second moments of the canonical quadratures. The objective is formulated as photon-state moment control: the quadrature expectations ⟨Q⟩ and ⟨P⟩ are driven to prescribed reference values, while the total quadrature variance VQ +VP is reduced to the smallest value compatible with the Schrodinger–Robertson uncertainty invariant. A linearly driven cavity mode is first reviewed. Linear ¨ displacement controls can steer the quadrature means, but they do not affect the covariance matrix. A bilinear quadrature Hamiltonian containing the symmetrized product QP+PQ is then introduced. The resulting closed moment equations form a finite-dimensional bilinear control system with an invariant hyperboloid determined by the uncertainty relation. A feedback law for the bilinear input is proposed to minimize the total variance, and a sliding-mode design is developed for robust tracking of the quadrature means. The formulation provides a bridge between continuous-variable quantum optics and nonlinear robust control, with possible applications to programmable photonic oscillators, coherent-state preparation, and quantum optical signal shaping.
Download
Paper Nr: 47
Title:

A Biquaternion-Based State Observer and Its Application to a Two-Level Quantum System
Authors:

Andrei A. Markov and Sergey V. Drakunov
Abstract: This work develops a biquaternion-based multiplicative-error observer for a two-level quantum system. The construction is motivated by a quaternion kinematic observer previously developed for rigid-body attitude estimation, in which convergence is achieved through a multiplicative quaternion error rather than through an additive error. For a two-level quantum system, the natural algebraic setting is the algebra of complexified quaternions, or biquaternions. We formulate a stochastic biquaternion propagator model for a continuously measured two-level system, introduce the corresponding multiplicative-error observer, and derive error equations in the biquaternion algebra. The proposed approach is intended as a complement to standard quantum filtering methods. In particular, it highlights the analogy between a quaternion attitude observer and an estimator for quantum spin dynamics.
Download
Paper Nr: 49
Title:

Beyond Human Captions: Semantic Supervision and PCA-Regularized Decoding for Stable-Diffusion-Based fMRI Reconstruction
Authors:

Abdallah Mohammed AboHassan, Ehab Elshazly, Tetsuji Ogawa and Ahmed Fares
Abstract: Reconstructing perceived images from functional magnetic resonance imaging (fMRI) is a high-dimensional inverse problem that has recently become more tractable through the use of latent diffusion models. Recent reconstruction pipelines regress fMRI responses onto the image-latent and CLIP text-embedding spaces of Stable Diffusion, which is then used as a frozen decoder. However, two design choices in this framework can substantially affect reconstruction performance but have not been systematically investigated: the type of semantic supervision used to define the regression target and the regularization strategy applied to highdimensional regression. In this study, we address both issues. First, we introduce a target-space principal component analysis reformulation, in which the 59136-dimensional CLIP text-embedding target used for Stable Diffusion conditioning is projected onto its top (K=1024) principal components before regression and reconstructed by inverting the projection at inference time. This formulation yields a target-space PCA ridge decoder whose effective coefficient matrix has rank at most K. Second, we compare three caption-supervision regimes: human-annotated COCO captions and machine-generated captions from BLIP-2 and Florence-2 models. Experiments on four subjects from the Natural Scenes Dataset (NSD) show that the combined effect of target-side PCA compression and the dense captions generated by Florence-2 yields a mean improvement of 10.3 percentage points over the human-caption baseline when evaluated across seven feature-space similarity metrics, with the sparser BLIP-2 captions reaching an improvement of 8.7 points under the same compression.
Download
Paper Nr: 56
Title:

Negative Binomial Mixtures with Data-Driven Initialization for Streaming Overdispersed Count Data
Authors:

Hana Sedláčková and Evženie Uglickich
Abstract: This paper addresses the modeling of overdispersed and multimodal count data arising in complex urban intelligent transportation systems. We propose a statistical framework based on mixtures of negative binomial components, combined with a recursive Bayesian estimation algorithm that enables efficient parameter updates without relying on the iterative expectation-maximization paradigm. For validation, the proposed methodology is applied to bivariate count data representing: (i) electric vehicle (EV) charging demand, measured as the number of plug-in events per hour in a parking facility equipped with both private and public chargers, and (ii) traffic counts collected from neighboring urban streets. The main contributions of this work are: (i) an extension of recursive Bayesian estimation theory for mixtures of negative binomial distributions, tailored to overdispersed and multimodal count data; and (ii) an online adaptation mechanism that tracks regime changes by updating sufficient statistics with each incoming data point in a streaming setting. Experimental results demonstrate accurate online estimation and improved probabilistic modeling of overdispersed EV charging and traffic count data.
Download
Paper Nr: 68
Title:

EJUST-GYM-3: A Multi-View Exercise Dataset for Human Activity Recognition and Motion Analysis
Authors:

Mohamed Abouelella, Ahmed Ashraf and Walid Gomaa
Abstract: Human Activity Recognition (HAR) for exercise analysis is fundamental to virtual coaching and personalized fitness systems, yet existing datasets remain limited in scale, viewpoint diversity, and consistency of motion representation. This paper introduces EJUST-GYM-3, a large-scale multi-view indoor exercise dataset comprising synchronized frontal and side-view RGB recordings from 123 participants performing four common full-body exercises: high knees, jumping jacks, sit-ups, and steam engine. Each repetition is manually segmented, temporally normalized, and processed using a unified pose-based feature engineering pipeline. We extract 3D skeletal landmarks using MediaPipe and construct a compact distance-based joint representation, followed by data-driven feature selection to retain the most informative motion patterns across exercises. The dataset is evaluated under subject-disjoint protocols using two tasks: multi-class exercise classification and subject verification based on execution style. Experimental results with hybrid CNN–LSTM models achieve 99.82% classification accuracy on unseen subjects and strong verification performance, with additional validation on public YouTube videos. These results demonstrate that EJUST-GYM-3 provides a reliable benchmark for exercise recognition and motion analysis.
Download
Paper Nr: 80
Title:

Effects of Seasonality in Dengue Spread Equilibria and Transient Behaviors
Authors:

Flavia Altamura, Paolo Di Giamberardino and Daniela Iacoviello
Abstract: The present work proposes a simplified model for season-based vector life cycle in arbovirus epidemic spread, with particular reference, for the numeric choice in the numerical simulations, to dengue. The difficulties in describing insect life under climatic and seasonal effects is here overcome moving to the infective contact rates the annual seasonal variation, approximated with a pure sinusoid. With this approach, it is easier to catch the effects of the change of epidemic spread during different season conditions. Unexpected evolution characteristics under perturbations are easily explained through graphical representations of numerical simulation results.
Download
Paper Nr: 126
Title:

Implicit Predecessor-Based Region of Attraction Estimation and Robust Invariance Analysis for a Two-Wheeled Inverted Pendulum
Authors:

Lorenzo Fici and Guillaume Ducard
Abstract: Estimating the region of attraction (RoA) of nonlinear systems is fundamental for assessing closed-loop stability and ensuring safe operation. While Lyapunov-based approaches provide certified stability guarantees, they often yield conservative inner approximations of the RoA. This paper combines a certified Lyapunov-based positively invariant set with a predecessor-based implicit representation to compute a significantly less conservative inner approximation of the RoA while preserving formal stability guarantees. In addition, the robust positive invariance of the initial certified Lyapunov-based invariant set is analyzed under bounded additive input disturbances, providing formal robustness guarantees. The proposed methodology is demonstrated on a nonlinear two-wheeled inverted pendulum stabilized by a saturated linear quadratic regulator. The resulting RoA approximation is compared with the initial Lyapunov-certified invariant set and validated through Monte Carlo simulations and hardware experiments, showing a substantially enlarged certified operating region that matches the empirical closed-loop behavior. These results demonstrate the practical applicability of combining certified Lyapunov analysis with predecessor-based set propagation for RoA approximation and robustness assessment of nonlinear systems.
Download
Paper Nr: 160
Title:

Data-Driven Modeling of Complex MIMO LPV Dynamics: A Scalable Iterative Approach Using Khatri–Rao Matrix Augmentation and Kalman-Enhanced State Estimation
Authors:

Jorge Puerto, Edwar Jacinto and Fredy Martínez
Abstract: Stochastic identification of Multi-Input Multi-Output (MIMO) Linear Parameter-Varying (LPV) systems is challenging because affine scheduling introduces parameter-state and parameter-input products. At the same time, the latent state and innovation sequences remain unknown. This paper presents Conditional Canonical Correlation Iterative Identification of LPV Systems (3C2ILPV), a subspace-identification method that extends the ICCALPV formulation; the novelty of 3C2ILPV lies in how it reformulates and organizes the stochastic LPV identification problem computationally. Relative to ICCALPV, which already employs successive conditional-CCA updates to handle the unknown scheduling-state products, 3C2ILPV replaces the sequential construction of scheduling-dependent augmented regressors by a block-Hankel/Khatri-Rao formulation, introduces a Kalman-assisted correction of the CCA-derived state estimate within the iterative refinement, and recovers the complete affine innovation model, including the scheduling-dependent gains K. Thus, the contribution is not the individual use of CCA, Kalman filtering, Khatri-Rao products, or least squares, but their specific integration into a block-structured iterative procedure for stochastic affine MIMO LPV innovation-model identification.
Download
Short Papers
Paper Nr: 62
Title:

Interpretable Human Activity Recognition via Phase-Space Trajectories
Authors:

Walid Gomaa and Ahmed Gomaa
Abstract: Human Activity Recognition (HAR) from wearable inertial sensors is often addressed using handcrafted features or deep learning models, many of which remain largely empirical and lack interpretability or computational efficiency. In this paper, we propose a dynamical systems–based HAR framework that models human movements as low-dimensional processes and reconstructs their phase space using Time-Delay Embedding (TDE), grounded in Takens’ embedding theorem. From the resulting phase-space trajectories, we extract compact statistical and geometric features that capture invariant motion dynamics and enable the use of lightweight, interpretable classifiers. We evaluate the proposed approach on the PAMAP2 and EJUST-ADL-1 datasets across a wide range of sensor configurations and fusion strategies. On PAMAP2, the method achieves up to 93.15% accuracy, with results showing that spatial fusion of accelerometer data alone consistently outperforms more complex multi-modal fusion schemes. In contrast, experiments on EJUST-ADL-1 demonstrate that gyroscope-based features are more discriminative than accelerometer-based ones, highlighting the modality-dependent nature of phase-space representations. Overall, the results demonstrate that phase-space analysis provides an efficient, interpretable, and robust alternative to conventional HAR pipelines.
Download
Paper Nr: 62
Title:

Interpretable Human Activity Recognition via Phase-Space Trajectories
Authors:

Walid Gomaa and Ahmed Gomaa
Abstract: Human Activity Recognition (HAR) from wearable inertial sensors is often addressed using handcrafted features or deep learning models, many of which remain largely empirical and lack interpretability or computational efficiency. In this paper, we propose a dynamical systems–based HAR framework that models human movements as low-dimensional processes and reconstructs their phase space using Time-Delay Embedding (TDE), grounded in Takens’ embedding theorem. From the resulting phase-space trajectories, we extract compact statistical and geometric features that capture invariant motion dynamics and enable the use of lightweight, interpretable classifiers. We evaluate the proposed approach on the PAMAP2 and EJUST-ADL-1 datasets across a wide range of sensor configurations and fusion strategies. On PAMAP2, the method achieves up to 93.15% accuracy, with results showing that spatial fusion of accelerometer data alone consistently outperforms more complex multi-modal fusion schemes. In contrast, experiments on EJUST-ADL-1 demonstrate that gyroscope-based features are more discriminative than accelerometer-based ones, highlighting the modality-dependent nature of phase-space representations. Overall, the results demonstrate that phase-space analysis provides an efficient, interpretable, and robust alternative to conventional HAR pipelines.
Download
Paper Nr: 64
Title:

Stochastic State-Multiplicative Retarded Systems: Robust Measurement-Feedback Control
Authors:

Eli Gershon
Abstract: This paper addresses the measurement-feedback control problem for continuous-time, state-multiplicative stochastic systems with time delays, considering both nominal and robust scenarios. In the robust formulation, the system matrices are assumed to be uncertain but confined within a prescribed polytope. For this case, a set of Linear Matrix Inequality (LMI) conditions parameterized by a single tuning variable is derived, enabling the computation of the state-feedback gain matrix. The proposed approach relies on a modified version of Finsler’s lemma, in which a distinct Lyapunov function is assigned to each vertex of the uncertainty polytope. Building upon the robust solution, a single feedback gain is obtained across the entire uncertainty domain. A numerical example is provided to compare the different solution methods presented in the paper.
Paper Nr: 84
Title:

Distributed Fusion-Based Cooperative Object Perception Using Onboard and Roadside LiDARs
Authors:

Hiroshi Murachi, Masafumi Hashimoto and Kazuhiko Takahashi
Abstract: This paper proposes a cooperative object perception method for objects, such as cars, motorcycles, and pedestrians, using networked onboard and roadside light detection and ranging units (LiDARs) in road environments. Each LiDAR unit detects objects from its own LiDAR image using the VoxelNeXt deep-learning model, exchanges object information (e.g., the position and size) with other networked LiDAR units, and fuses the received information using the non-maximum weighted method. Subsequently, each LiDAR unit estimates the object poses and refines them by exchanging the pose information among the units. A distributed interacting multimodel estimator is employed to estimate object poses under various motion patterns in a distributed manner. The performance of the proposed method is validated on the V2X-Real open dataset.
Download
Paper Nr: 86
Title:

LiDAR-Based Cooperative Object Perception Using Multiple UAVs for Road Traffic Monitoring
Authors:

Yuto Nakayama, Masafumi Hashimoto and Kazuhiko Takahashi
Abstract: This paper proposes a cooperative object perception method that employs light detection and ranging (LiDAR) units mounted on multiple unmanned aerial vehicles (UAVs) operating in road environments. The leader UAV acquires the LiDAR point cloud data of a road and performs density-aware-based early fusion. Specifically, the leader UAV maps the point cloud data onto a single-layer three-dimensional voxel map and calculates the point density for each voxel. Subsequently, the leader UAV identifies low-density voxels and requests the follower UAV to transmit LiDAR point cloud data only within these low-density voxels for point cloud fusion. This selective data fusion strategy achieves high detection accuracy and reduces communication overhead. The fused point cloud is then processed using CenterPoint-based object detection, followed by interacting multiple model-based object tracking. Simulation experiments on cooperative object perception using two LiDAR-equipped UAVs demonstrate the effectiveness of the proposed method.
Download
Paper Nr: 87
Title:

Helmet LiDAR-Based Detection and Tracking of Moving Objects for Micromobility Safety
Authors:

Yuya Osaka, Masafumi Hashimoto and Kazuhiko Takahashi
Abstract: This paper presents a detection and tracking of moving objects (DATMO) method using a solid-state light detection and ranging (LiDAR) mounted on the helmet of a micromobility rider. The proposed system is intended for active safety applications involving micromobility vehicles such as bicycles, e-bikes, and mopeds. Motion artifacts in the LiDAR data are corrected using a quaternion-based unscented Kalman filter that utilizes helmet pose information obtained via simultaneous localization and mapping (SLAM) and an inertial measurement unit. LiDAR data related to moving objects (moving LiDAR data) are extracted by subtracting a SLAM-based local environmental map from the current LiDAR data and subsequently applying an occupancy grid-based method. The extracted moving LiDAR data are then tracked using an extended object tracking framework based on a multiplicative error model-extended Kalman filter. Experiments conducted in a public-road environment demonstrate the effectiveness of the presented DATMO method.
Download
Paper Nr: 96
Title:

Comparative Analysis of Sparse Identification of Nonlinear Dynamics and Pose Estimation Approaches for Automated Resistance Exercise Classification
Authors:

Walid Gomaa, Abdelrahman Wael Ammar and Ismael Abbo
Abstract: The growing adoption of remote fitness coaching and physical tele-rehabilitation has created substantial demand for automated recognition systems capable of identifying and assessing complex resistance exercises. This study presents a rigorous comparative analysis of two fundamentally different sensing modalities in human activity recognition for exercise classification: wearable inertial measurement units and markerless computer vision. In the first modality, a novel feature engineering pipeline was developed utilizing the sparse identification of nonlinear dynamics to extract governing equations of human motion. For the vision-based approach, MediaPipe Pose estimation was employed to extract three-dimensional body landmarks from videos, from which biomechanical and statistical features were engineered. Both frameworks were evaluated on multi-subject datasets comprising four different resistance exercises. Our results demonstrated that although the video-based approach achieved a higher accuracy, the wearable-based approach remained highly competitive and offers practical alternative ranging from privacy-sensitive to environmental conditions independence.
Download
Paper Nr: 117
Title:

Model-Based Reinforcement Learning for a Linear-Quadratic Optimal Controller Driven by a Full-Order Observer
Authors:

Hafsa Nadia and Zoran Gajic
Abstract: Reinforcement learning generally assumes that complete state information is available to implement full-state feedback control and evaluate the reward associated with each action. For real physical systems, such a requirement is mostly not feasible. One way to avoid this problem is to use state observers. This paper considers model-based reinforcement learning (RL) for continuous-time linear-quadratic (LQ) control when only output measurements are available. A Policy iteration method is implemented through the Kleinman algorithm, which solves a sequence of Lyapunov equations, and a full-order Luenberger observer provides the state estimates used for feedback. An augmented Lyapunov formulation is used to evaluate the performance loss incurred by observer-driven control relative to ideal full-state feedback. The analysis characterizes the approximate and converged observer-driven costs and examines the effect of observer initialization. Numerical simulations of a sixth-order LCL-filter grid-connected inverter model, with only grid-side currents measured, illustrate the method. In this example, least-squares initialization of the observer produces a substantially smaller performance loss than zero initialization.
Download
Paper Nr: 143
Title:

General Prescribed-Time Synchronization Framework for Lipschitz Class of Nonlinear Systems with Application to Chaotic Systems
Authors:

Bharat Bhushan Sharma
Abstract: Stability and synchronization of nonlinear systems have attracted considerable research attention over the years due to its wide-ranging applications in robotics, biological systems,secure communication, power systems etc. Among various synchronization control strategies, prescribed-time synchronization enables the systems to achieve synchronization within a user-defined time which is independent of the initial conditions. Such schemes are vital and preferred choice for time-critical applications across the domains. This paper develops a general prescribed-time synchronization framework for a class of nonlinear systems while exploiting the structural attributes of these systems. The proposed approach gives simple time-varying state-feedback control function which drives the synchronization error to zero within a designer specified time horizon for systems in master-slave configuration. Sufficient conditions for synchronization are analytically derived using quadratic Lyapunov stability theory under a global Lipschitz assumption for the nonlinearities. The efficacy of the proposed strategy is validated through synchronization of the benchmark Lorenz chaotic system in master-slave configuration. Numerical simulations illustrate successful tracking of the master system states by the states of slave system within the specified convergence time.
Download
Paper Nr: 145
Title:

Infra-Less Pedestrian Indoor Positioning and Commercial Network-Based Information Provision System
Authors:

Seong Yun Cho
Abstract: To support the active functioning of crime victims, it is important to secure location information of indoor spaces and monitoring it in real time. Pedestrian dead reckoning (PDR) technology is being researched and developed to continuously track the location of pedestrians in indoor spaces where wireless location tracking is unavailable. NavIn Labs Co., Ltd. is a company established for the commercialization of this technology. Pedestrian inertial navigation for safety (PiNS), developed by this company, is a pedestrian location tracking system utilizing pedestrian inertial navigation (PIN) technology within PDR, which transmits location information to a real-time monitoring system based on commercial networks. The compact sensor module was developed for easy attachment to shoes, and the integrated PIN algorithm corrects errors through self-correction technology and recognizes movement on stairs to distinguish location by floor. This system is expected to effectively support safety activities for crime victims.
Download
Paper Nr: 148
Title:

Reinventing Tracheostomy Safety: Comparative Review of Emerging Monitoring Technologies
Authors:

Laya Salazar Lafon, Thomas Schlinquer, Xavier Dupuis, David Olivier, Dima Siblani and Maria Alice Bertolim
Abstract: Tracheostomy is one of the most common procedures in intensive care, with over 1,000,000 surgeries annually in the United States and about 30,000 in Japan. Despite its widespread use, tracheostomy care remains challenging, exposing patients to life-threatening complications such as tube obstruction, accidental decannulation, and infection. These events require vigilant surveillance, which remains difficult to achieve, particularly outside specialized or closely monitored care settings. International and French data highlight the vulnerability of tracheotomized patients due to underlying comorbidities and device-related risks, emphasizing the need for reliable monitoring strategies. Advances in artificial intelligence (AI), miniaturized sensors, and contactless technologies now offer promising opportunities to enhance tracheostomy surveillance. This review examines four major technological approaches - sound-based monitoring, integrated flow sensors, contactless sensing, piezoelectric material - and synthesizes recent findings on their performance, benefits, and limitations. This review identifies current technological gaps and discusses future directions toward safer and more intelligent tracheostomy monitoring systems.
Download
Paper Nr: 24
Title:

An IoT-Based Intelligent Irrigation System for Smart Gardening
Authors:

Delia Maria Cremene, Adriana Mihaela Coroiu and Ioan Daniel Pop
Abstract: The research aims to introduce the design and implementation of an Internet of Things (IoT)-based intelligent irrigation system, which is applicable to small-scale and domestic horticulture. The proposed method can overcome the limitations of the scheduled irrigation method, which is characterized by water inefficiency and uneven moisture content in the soil, by utilizing the sensor data to control the irrigation process. The proposed intelligent irrigation system consists of a capacitive sensor for measuring the moisture content, a DHT22 sensor for measuring temperature and humidity, and a NodeMCU ESP8266 microcontroller for controlling the irrigation process. The sensor data is sent to the cloud database, where the data is stored for future use. The proposed intelligent irrigation control is based on the moisture content of the soil, which ensures the plants are watered only when required, reducing the possibility of over- and under-watering. The proposed intelligent irrigation control is compared with the scheduled irrigation control, which proves the efficiency of the proposed method in maintaining the moisture content with less water consumption.
Download
Paper Nr: 57
Title:

Beyond Mean-Square Exponential Stability: Automatic Verification of Asymptotic Stability in Probability
Authors:

Sigurdur Hafstein
Abstract: Mean-square exponential stability is the concept usually used in application, when stability of equilibria for stochastic differential equations is studied. The reason is that this stability concept has been the most assessable one algorithmically. Recently, an algorithm to verify the global asymptotic stability in probability (GASiP) for the equilibrium of autonomous, linear stochastic differential equations with constant coefficients was presented. For nonlinear systems this algorithm can be used for the linearization of the systems to assert (local) asymptotic stability in probability (ASiP). We argue that ASiP/GASiP, which is a considerably weaker condition than mean-square exponential stability, is more appropriate in most applications and, since its verification is now easily assessable using open-source software, should be used instead of mean-square exponential stability.
Download
Paper Nr: 115
Title:

Modelling Kill Chain Disruption for Artillery Unit Protection against Small UAS in Constructive Simulation
Authors:

Tomáš Havlík, Jan Ivan, Ondřej Pekař, Miroslav Mušinka and Jurij Fejer
Abstract: Small unmanned aircraft systems have significantly reduced the time between target detection and strike execution, leaving artillery units with only a limited opportunity to complete fire missions, manoeuvre, and preserve combat capability. Consequently, artillery survivability depends on both reducing the probability of a successful strike and gaining sufficient time to accomplish the mission before the adversary completes its kill chain. Existing research primarily addresses UAS detection, tracking, neutralisation, and the evaluation of individual counter-UAS technologies, while little attention has been devoted to scenario-based assessment of artillery battery survivability. This paper proposes a framework for modelling adversary kill chain disruption at the artillery battery level within a constructive simulation environment. The framework decomposes the adversary kill chain into detection, decision-making, and strike phases, linking passive, deceptive, and active protection measures to their effects on phase duration, strike probability, and mission effectiveness. The framework is demonstrated in the MASA SWORD constructive simulation environment through three operational scenarios. Initial simulation results show how different combinations of protection measures influence kill chain disruption, battery survivability, critical asset losses, and mission effectiveness. The principal contribution is a terminologically consistent framework that provides a foundation for future model validation and comparative evaluation of alternative protection concepts.
Download