Abstract
Most machine learning (ML) models are not intrinsically well calibrated, meaning that their confidence scores are not consistent with posterior probabilities. Moreover, the most commonly used metric to evaluate calibration, namely the Expected Calibration Error (ECE), has several limitations, including its dependence on hyperparameters such as the number of bins. The ECE estimates local accuracy and confidence by binning predictions in the confidence space. However, this yields a discontinuous estimate and introduces nonlocal effects, since imbalanced data can significantly affect the distribution within bins. To address these issues, we propose the Doubly Kernelized Expected Calibration Error (k(2) ECE), which introduces kernel based estimates of local accuracy and local confidence that are both centered at each observation and continuous. In addition, we propose a loss function based on this metric. Our results show that jointly optimizing for accuracy and calibration can be a viable approach, particularly when using bilevel optimization, although a trade off between accuracy and calibration is observed.

Abstract
The present research aims to analyse the perceptions of young university students regarding digital influencers and their role in brand communication strategies. A questionnaire was conducted using “Google Forms,” administered online between the months of February and April 2023, with a non-probabilistic convenience sample. A total of 196 questionnaires were collected and validated. The results show that young individuals recognize the role of digital influencers in brand communication to discover new products and brands. Insta- gram is the most widely used social network, and young university students follow digital influencers, primarily seeking content related to “fashion and beauty.” Respondents acknowledge when content is sponsored, but this does not diminish their trust and credibility in the influencer, especially when it aligns with their profile and lifestyle.

Abstract
At the practical level, the static stability constraint is one of the most important constraints in practical pallet loading problems, such as air cargo palletizing. Approaches to modeling static stability, which range from base support and mechanical equilibrium calculations to physical simulation, differ in workflow, focus, and assumptions, so choosing the right static stability approach has a substantial impact on the quality of the solution and, ultimately, on loading security. To date, little research has investigated the structural differences between approaches. The aim of this paper is to integrate knowledge and shed light on the applicability of the different approaches for the practical scenario of air cargo palletizing. We tackle this problem through (1) a reformulation and extension of static stability toward loading stability, (2) a conceptual analysis of current approaches, and (3) benchmarking that employs an independent multibody simulation on multiple heterogeneous datasets. Our results show that all approaches are prone to structure errors and vary significantly in their premises and information usage. Further, full base support is revealed to be the most restrictive approach by far, while physical simulation achieves the greatest accuracy. Given the trade-off between accuracy and runtime, the mechanical equilibrium approach is a good choice, while partial base support performs best for lower support values.

Abstract
This work presents a reinforcement learning (RL) approach for safe and efficient navigation of unmanned aerial vehicles (UAVs) using low-power onboard sensors. Our system is developed on the MANTIS aerial platform and employs the Twin Delayed Deep Deterministic Policy Gradient (TD3) algorithm to learn a velocity and heading control policy entirely within a ROS-Gazebo simulation framework, with future deployment on real hardware. The sensing stack comprises the simulation of two OV7251 grayscale cameras for stereo vision and an LD-19 2D LiDAR, forming a compact and energy-efficient alternative to high-end perception systems. These inputs are fused into a structured observation space that includes goal-relative position, UAV dynamics, and 18 proximity sectors capturing local obstacle information. The reward function is shaped to encourage goal progress, heading alignment, smooth maneuvers, and obstacle-aware behavior. Training is parallelized across up to five agents and evaluated in multiple 3D simulated environments of increasing complexity. We assess the learned policy based on success rate, trajectory efficiency, and time to goal. Our results demonstrate strong generalization to unseen scenarios and highlight the feasibility of deploying learned policies on lightweight UAVs powered by embedded compute platforms such as Qualcomm SoCs. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2027.

Abstract
This paper presents a graph-based SLAM approach using GTSAM for UAV localization in GPS-denied warehouse environments. By leveraging existing barcodes as passive landmarks and fusing dual visual-inertial odometry with incremental factor graph optimization, our system achieves robust indoor navigation without requiring pre-mapped infrastructure. The proposed approach integrates adaptive VIO fusion, barcode-as-unknown-landmark detection, and real-time iSAM2 optimization to enable autonomous warehouse operations. Experimental validation on the MANTIS UAV platform demonstrates sub-decimeter accuracy across laboratory and real warehouse scenarios, with RMSE values of 0.031 m under challenging conditions and 0.068 m in operational environments, confirming practical deployment viability for inventory management applications. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2027.

Abstract
Unmanned aerial vehicles (UAVs) have emerged as a technically viable and increasingly adopted solution to address key operational and logistical constraints, especially in offshore renewable energy platforms. A critical operational challenge lies in the execution of reliable and precise landing procedures. Independent of the control paradigm—be it fully autonomous or teleoperated—UAV landing operations demand accurate spatial coordination, robust control algorithms, and dynamic stability, particularly under adverse marine conditions characterised by restricted landing areas, platform motion, and high environmental variability. This paper introduces an adaptive deep reinforcement learning (DRL) strategy for UAV autonomous landing, suitable for moving and floating structures. Combining its pose estimation with the visual tags’ segmentation, the agent controls the vehicle through progressively constrained spatial zones, ensuring restricted precision during the final descent. The framework was developed on the Gazebo simulator, and the training and tests were divided into two stages: a simpler simulation environment with multi-vehicle training for static and moving targets, and a more complex offshore scenario for testing the model on a drone with identical features to the ones presented in the real vehicle. The results confirm the policy’s ability to generalise to realistic offshore dynamics, demonstrating promising performance in both the static and oscillatory conditions, achieving success rates of 100 % and 86 %, respectively. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2027.