Centre for Automation and Robotics

Deep neural networks are widely used for image classification in different fields, although selecting an appropriate architecture often remains a trial-and-error process. The purpose of this work is to investigate a convolutional neural network architecture used to detect whale pulses in spectrograms in order to better understand the causes of its underperformance. By examining the behaviour of its internal layers, we show that the early convolutional blocks capture the most informative acoustic features, while deeper layers provide limited additional benefit and, under the considered training conditions, may even degrade classification accuracy. Based on these observations, we derive a simplified architecture consisting of only the first two convolutional layers followed by a lightweight classifier. This network achieves near-optimal performance, improving accuracy from 87% to 98%, and exhibits substantially lower variability between repetitions compared to the original model.

ABSTRACT The wear performance of additively manufactured (AM) AlSi10Mg components is critical for their deployment in tribological applications, yet a comparative analysis of the wear behavior under as‐built, T6, and stress‐relief (SR) conditions remains insufficiently explored. To broaden the industrial adoption of AM components, it is crucial to evaluate their wear behavior, as this underpins reliability and safety while promoting creativity in both design and material choices. This research examines the wear characteristics of the AlSi10Mg alloy created using Selective Laser Melting (SLM), evaluated in the as‐built condition and following T6 and SR heat treatments. The microstructural variations under these three states were analyzed using scanning electron microscopy (SEM). The0020as‐built specimens exhibited the highest hardness (137.3 HV), due to the presence of a refined α‐Al cellular framework embedded with Si particles generated by rapid solidification. Heat treatment altered this structure, leading to Si phase coarsening and a corresponding reduction in hardness to 103.35 HV in the T6 condition and further down to 73.75 HV in the SR condition. Wear experiments were carried out under applied loads ranging from 5 to 15 N (max load 15 N) for a duration of 300 s, along with assessments of the coefficient of friction (COF), the surface morphology following wear, and the loss of material. The findings indicated that the as‐built specimens consistently demonstrated lower wear volume loss across all load levels in comparison to the samples that underwent heat treatment. Additionally, the heat‐treated specimens developed compressive residual stresses, while the as‐built SLM parts primarily exhibited tensile stresses.

In this paper, we present a learning-based tracking controller based on Gaussian processes (GP) for collision avoidance of multi-agent systems where the agents evolve in the special Euclidean group in the space SE(3). In particular, we use GPs to estimate certain uncertainties that appear in the dynamics of the agents. The control algorithm is designed to learn and mitigate these uncertainties by using GPs as a learning-based model for the predictions. In particular, the presented approach guarantees that the tracking error remains bounded with high probability. We present some simulation results to show how the control algorithm is implemented.

Unmanned Aerial Vehicles (UAVs) equipped with multispectral sensors enable high-resolution, cost-effective crop monitoring by deriving vegetation indices such as NDVI for quantitative assessment of canopy vigor. However, illumination variability, atmospheric effects, and sensor limitations can adversely affect measurement accuracy. This study proposes a radiometric compensation method that emulates Earth surface reflectance using the reflectance of reference targets, based on physics-based modeling of solar irradiance, atmospheric effects, sensor response, and reference panel properties, while quantifying uncertainty through Monte Carlo sensitivity analysis. The method is first validated using multispectral imagery captured via UAV-mounted sensors from an in-field reference target, then applied to images from pistachio orchards to compensate for atmospheric effects prior to NDVI evaluation. Corrected NDVI values are evaluated against handheld NDVI sensor measurements and against Agisoft Metashape’s standard compensation approach. The proposed method significantly improves NDVI accuracy, closely matching reference data and outperforming the conventional approach, thereby enhancing the reliability of UAV-based remote sensing in precision agriculture.

Motion prediction plays a vital role in the safe deployment of autonomous vehicles, particularly in highly interactive and dynamic environments. However, challenges such as limited visibility due to obstructions, sensor range limitations, and misdetections can compromise situational awareness and significantly increase safety risks. This work extends the MultIAMP framework—which integrates a Dynamic Bayesian Network (DBN) and Markov chains to generate multimodal probabilistic predictions—by incorporating enhancements that enable it to handle occlusions using real perception inputs. The proposed approach systematically detects, processes, and anticipates the behavior of occluded areas, enabling safer navigation. The system is quantitatively evaluated using the CARLA simulator in combination with state-of-the-art perception and planning modules. Results, expressed through multiple quantitative metrics across diverse driving experiments, demonstrate that integrating occlusion awareness into the prediction module enhances both comfort and safety for autonomous vehicles.

Abstract This work presents a novel end-to-end approach to early defect detection in thermo-insulating panels, widely used in the construction sector. Unlike existing studies that focus on isolated aspects, our contribution combines: (i) a composable and interoperable software platform designed to support data augmentation, segmentation, and classification; (ii) a multi-technique synthetic image generation strategy that overcomes the limitations of scarce and unbalanced datasets; and (iii) a hybrid segmentation and classification pipeline that extends traditional U-Net architectures with post-segmentation polygonal analysis for robust defect identification. The proposed solution is validated on the GAMHE 5.0 Industry 5.0 pilot line, demonstrating not only improvements in detection accuracy and control time but also its feasibility in human-in-the-loop, near-production environments, a gap rarely addressed in the literature. These results confirm that the approach goes beyond a mere consolidation of existing methods, providing instead a scalable, human-centric, and sustainable framework for Zero-Defect Manufacturing.

This work presents the structural analysis and validation of a sub-250 g FPV drone chassis, emphasizing both theoretical rigor and practical applicability. The novelty of this contribution lies in four complementary aspects. First, the structural philosophy introduces a screwless frame with interchangeable arms, joined through interlocking mechanisms inspired by traditional Japanese joinery. This approach mitigates stress concentrations, reduces weight by eliminating fasteners, and enables rapid arm replacement in the field. Second, validation relies on nonlinear static and transient FEM simulations, explicitly including crash scenarios at 5 m/s, systematically cross-checked with bench tests and instrumented flight trials. Third, unlike most structural studies, the framework integrates firmware (Betaflight), GPS, telemetry, and real flight performance, linking structural reliability with operational robustness. Finally, a practical materials pathway was implemented through a dual-track strategy: PETG for rapid, low-cost prototyping, and carbon fiber composites as the benchmark for production-level performance. Nonlinear transient FEM analyses were carried out using Inventor Nastran under multiple load cases, including maximum motor acceleration, pitch maneuvers, and lateral impact at 40 km/h, and were validated against simplified analytical models. Experimental validation included bench and in-flight trials with integrated telemetry and autonomous features such as Return-to-Home, demonstrating functional robustness. The results show that the prototype flies correctly and that the chassis withstands the loads experienced during flight, including accelerations up to 4.2 G (41.19 m/s2), abrupt changes in direction, and high-speed maneuvers reaching approximately 116 km/h. Quantitatively, safety factors of approximately 5.3 under maximum thrust and 1.35 during impact confirm sufficient structural integrity for operational conditions. In comparison with prior works reviewed in this study, the key contribution of this work lies in unifying advanced, crash-resilient FEM simulations with firmware-linked flight validation and a scalable material strategy, establishing a distinctive and comprehensive workflow for the development of sub-250 g UAVs.

Abstract Recruiting passive candidates, i.e. individuals not actively seeking jobs but open to compelling opportunities, remains one of the hardest challenges in digital recruitment. Motivated by a real collaboration with an industry partner, we introduce the Independent Halting Cascade (IHC) model: a simple but rich agent-based framework that couples network diffusion with the possibility of halting through job applications. Agents can either recommend vacancies to peers or apply themselves, and incentives increase the likelihood of recommendation, mobilizing otherwise passive candidates. The IHC bridges research on social network diffusion, coordinated task completion, and labor economics by modeling heterogeneous skills, job specificities, and network structures, including homophily. We derive analytical boundaries that characterize diffusion and failure regimes, and we show through simulations that the IHC reproduces the empirical chain-length distributions of Travers and Milgram and Dodds with only coarse calibration. Across synthetic (ER, BA, homophilic) and real networks (SMS, e-mail, Twitter), the IHC achieves comparable or higher success rates than direct-recommendation baselines, while requiring fewer applicants. Our findings suggest that the IHC captures core mechanisms of coordinated task completion, offering both a theoretical contribution and a practical foundation for recruitment systems designed to reach and engage passive candidates.

Search and rescue (SAR) operations require optimal solutions to maximize efficiency and minimize risks for human responders. Quadruped robots have emerged as viable agents due to their locomotion and agility capabilities. This work presents a navigation framework designed to enhance stability and adaptability of legged robots in complex environments. A high-fidelity simulation environment was developed using NVIDIA IsaacSim, incorporating realistic disaster conditions such as unstructured terrain, fire, and smoke. Multiple stochastic routes were generated and analyzed in simulation in terms of energy consumption, stability, traversal time, and fall occurrences, to categorize them and determine the safest and most efficient path before real-world deployment, reducing the likelihood of failures. A fuzzy logic controller was proposed to regulate speed and improve locomotion adaptability. The proposed approach was validated in both simulation and real-world experiments. The results demonstrate the effectiveness of the proposed strategy in enabling safe and efficient navigation for quadruped robots in SAR missions. The code is publicly available at Robcib-GIT/IsaacSim_LeggedRobotSupplementary Video: https://youtu.be/cxEP5YVy8Qo

The origin of life initiated an evolutionary continuum yielding biologically optimized systems capable of operating under extreme environmental constraints. Biomimetics, defined as the systematic abstraction and transfer of biological principles into engineering domains, has become a strategic design paradigm for addressing the multifactorial challenges of space systems. This study introduces two core contributions to formally establish the discipline of Aerospace Bionic Robotics (ABR): First, it elucidates the relevance of biologically derived functionalities such as autonomy, adaptability, and multifunctionality to enhance the efficiency of space robotic platforms operating in microgravity environments. Second, it proposed the BEAM-D (Biomimetic Engineering and Aerospace Mechatronics Design), a standard for the development of Aerospace Bionic Robotics. By integrating biological abstraction levels (morphological, functional, and behavioral) with engineering protocols including ISO, VDI, and NASA’s TRL, BEAM-D enables a structured design pathway encompassing subsystem specification, cyber–physical integration, in situ testing, and full-scale mission deployment. It is implemented through a modular BEAM-DX framework and reinforced by iterative BIOX design steps. This study thus establishes formalized bio-inspired design tools for advanced orbital and planetary robotic systems capable of sustained autonomous operations in deep space exploration scenarios.