Real-time simulations of the induced electric fields during transcranial magnetic stimulation play an important role in guiding and optimizing the coil positioning. In this paper, we present our ongoing work on a deep learning-based surrogate model designed to rapidly predict the induced electric field distribution across the entire cortex, offering a much faster alternative to traditional numerical solvers. Leveraging (conditioned) transformer architectures, our approach operates directly on mesh-based head geometries, achieving highly accurate simulations in just 0.08 seconds on consumer hardware. While we continue to improve the neural surrogate, its current accuracy and efficiency have already enabled integration into an augmented reality platform, demonstrating a promising foundation for live electric field-guided brain stimulation applications.

We introduce MotionDPS, a unified Bayesian framework for motion-compensated 3D MRI that jointly estimates the anatomical image, rigid-body motion, and coil sensitivity maps directly from motion-corrupted k-space data.
MotionDPS integrates pretrained 3D complex-valued score-based diffusion models as expressive anatomical priors within a physics-based forward model, tightly coupling learned generative modeling with data-consistent reconstruction.
Inference alternates diffusion posterior image updates with proximal optimization steps for motion and coil estimation, enabling fully unsupervised reconstruction without requiring paired motion-free training data or external motion tracking.
Experiments on simulated and in vivo brain MRI datasets demonstrate consistently improved image quality and motion robustness compared to state-of-the-art methods.

Machine learning applied to Irritable Bowel Syndrome (IBS) frequently reports high diagnostic performance despite small cohorts and heterogeneous preprocessing. We present a four-stage responsible AI workflow applied to the MARS-IBS-2020 dataset (n=73 at baseline), demonstrating that leakage ablation, harmonized preprocessing, and learning-curve diagnostics expose structural overfitting behind nominally high ROC-AUC values. A single evaluation yielded AUC 0.985 under temporal filtering, yet bootstrapped resampling (N=50) reveals a mean AUC of only 0.63 with 95% CI spanning [0.10, 1.00], and memorisation persists across all sample sizes. SHAP attribution consistently surfaces bile acid and tryptophan metabolites as dominant predictors, aligning with established IBS biology. The study offers a transparent methodological template for evaluating small-cohort biomedical ML systems where traditional performance metrics are unreliable.

Polypharmacology — the unintended binding of drug candidates to non-target proteins — remains a leading cause of late-stage clinical trial failure, yet it is largely invisible to conventional sequence-based screening methods. We present CavitOmiX, an AI-driven computational framework that addresses this problem through proteome-wide cavitome analysis: systematic mapping of all potential drug-binding sites across entire proteomes as high-dimensional physicochemical point cloud fingerprints. Leveraging GPU-accelerated protein structure prediction (AlphaFold2, ESMFold, OpenFold) and Innophore's Catalophore™ fingerprinting technology, CavitOmiX enables proteome-scale nearest-neighbour cavity searches. The human cavitome, comprising over 800,000 binding site fingerprints derived from 122,907 AI-predicted structural models (published open-source in Nature Scientific Data), serves as a reference for early off-target liability prediction. The same infrastructure supports drug repurposing by matching new targets against approved drug binding sites, and antiviral lead optimization via a genetic algorithm. We validate the framework across three use cases: resistance monitoring in SARS-CoV-2, structural proteome analysis of Monkeypox virus, and full-pipeline antiviral design for Chikungunya virus. All functionality is accessible via a public web application at copilot.cavitomix.bio.

Primary health care frequently relies on low-cost imaging devices, which are commonly used for screening purposes. To ensure accurate diagnosis, these systems depend on advanced reconstruction algorithms designed to approximate the performance of high–quality counterparts. Such algorithms typically employ iterative reconstruction methods that incorporate domain-specific prior knowledge. However, downstream task performance is generally assessed using only the final reconstructed image, thereby disregarding the informative intermediate representations generated throughout the reconstruction process. In this work, we propose IRTTA to exploit these intermediate representations at test-time by adapting the normalization-layer parameters of a frozen downstream network via a modulator network that conditions on the current reconstruction timescale. The modulator network is learned during test-time using an averaged entropy loss across all individual timesteps. Variation among the timestep-wise segmentations additionally provides uncertainty estimates at no extra cost.
This approach enhances segmentation performance and enables semantically meaningful uncertainty estimation, all without modifying either the reconstruction process or the downstream model.

Accurate human 3D avatars are essential for various medical applications, such as pain visualization and burn size estimation. Generating these avatars from simple anthropometric measurements offers a cheaper and more practical alternative to conventional 3D body scanners and image-based reconstruction methods, especially when patients have limited mobility. In this work, we investigate the reliability of predicting Skinned Multi-Person Linear model (SMPL) shape parameters from anthropometric measurements in the presence of real-world noise. We further introduce \beta-likelihood to quantify the anatomical plausibility of generated shapes against a learned distribution. Multiple regression models are evaluated on two external datasets, revealing a clear trade-off between metric accuracy and shape plausibility. The results indicate that regularized regression models are best suited to balance this trade-off when dealing with real-world measurement noise.

As Large Language Models (LLMs) transition from experimental tools to collaborative teammates in clinical settings, their behavioral alignment becomes a critical safety factor. This paper investigates "Clinical Sycophancy'', the tendency of an AI model to align its diagnostic output with a user's stated hypothesis even when that hypothesis contradicts clinical evidence. Drawing on Joint Cognitive Systems theory, we argue that sycophancy represents a catastrophic failure of the "check and balance'' function expected of a second-opinion system. We introduce an automated pipeline to stress-test general-purpose frontier models using free-form, counterfactual diagnostic prompts based on the MedQA dataset. Dose-response replication across three models reveals a model-dependent authority threshold: gpt-4o and o3-mini show a sharp institutional-authority step-change (~10-13% SR), while gemini-2.5-flash remains low across all authority levels. Multi-reason Devil's Advocate framing eliminates or near-eliminates epistemic sycophancy without model retraining, while cognitive forcing systematically backfires and all models exhibit progressive resilience decay under persistent multi-turn pushback.

Generating synthetic CT (sCT) from MRI or CBCT plays a crucial role in enabling MRI-only and CBCT-based adaptive radiotherapy, improving treatment precision while reducing patient radiation exposure. To address this task, we adopt a fully 3D Flow Matching (FM) framework, motivated by recent work demonstrating FM's efficiency in producing high-quality images.
In our approach, a Gaussian noise volume is transformed into an sCT image by integrating a learned FM velocity field, conditioned on features extracted from the input MRI or CBCT using a lightweight 3D encoder. We evaluated the method on the SynthRAD2025 Challenge benchmark, training separate models for MRI-sCT and CBCT-sCT across three anatomical regions: abdomen, head and neck, and thorax. Validation and testing were performed through the challenge submission system. The results indicate that the method accurately reconstructs global anatomical structures; however, preservation of fine details was limited, primarily due to the relatively low training resolution imposed by memory and runtime constraints. Future work will explore patch-based training and latent-space flow models to improve resolution and local structural fidelity.

Optical coherence tomography (OCT) layer boundary regression methods provide sub-pixel precision and topological guarantees but fundamentally assume that every layer exists across all A-scans. This mathematical constraint fails in severe pathologies such as Geographic Atrophy (GA), where specific retinal layers disappear. We extend the topologically constrained SD-RetinaNet framework to jointly perform boundary regression and explicitly detect missing layers using uncertainty quantification. We introduce a Gaussian Negative Log-Likelihood (NLL) formulation to calibrate aleatoric uncertainty, capturing spatial boundary errors. Concurrently, we employ an Evidential Deep Learning (EDL) module to model epistemic uncertainty directly from the network outputs, allowing the network to detect regions with zero structural evidence for a layer. Our framework addresses the largely overlooked challenge of anatomical absence in boundary regression, combining sub-pixel localization with direct atrophy segmentation.

Predicting disease progression from longitudinal data is an important open challenge in biomedical research. In this paper, we focus on Geographic Atrophy (GA), a late dry form of Age-Related Macular Degeneration (AMD), and aim to construct a personalizable spatio-temporal disease progression model.
Using a series of retinal Optical Coherence Tomography (OCT) scans, we infer the coefficients of a parametrized Partial Differential Equation (PDE) that best describes the true progression dynamics.
This recovered PDE acts as a soft constraint, guiding an Implicit Neural Representation (INR) to accurately extrapolate future GA segmentation maps.
To enable efficient training, we propose an iterative method, PINN-EM, specifically designed to recover the coefficients of non-linear PDEs from observational data.
At each iteration, the method decouples the objective into a PDE coefficient fitting step and a data fitting step, drawing inspiration from the Expectation-Maximization algorithm.
We evaluated PINN-EM on the complex task of GA progression modeling, where patients exhibit high variance in growth patterns and biological rates.
The proposed spatio-temporal model substantially outperformed existing baselines, demonstrating the strong extrapolation capabilities of PDE-constrained INR models.

Whole-heart multi-compartment CT segmentation is clinically important, but standard CNNs do not explicitly enforce anatomical plausibility. Based on statistics derived from the training data, we evaluate whether lightweight explicit shape priors, implemented as shape-aware losses and spatial label distribution heatmap-guided U-Net variants, improve 3D cardiac segmentation on MM-WHS CT and WHS++. Across all experiments, a standard 3D U-Net surprisingly remained a very strong baseline, with handcrafted priors yielding at best marginal and inconsistent changes and often degrading performance. These results suggest that the baseline already captures substantial implicit anatomical regularities and that future gains will likely require more expressive learned priors rather than simple handcrafted anatomical shape constraints.

Structure-based drug design aims to generate novel small molecules that bind favorably to a defined protein target. While recent generative models have demonstrated impressive results, they typically face challenges such as limited specificity, insufficient diversity, poor physical or chemical validity, weak binding affinity, and difficulties generalizing beyond the binders and targets seen during training. Here we present LigForge, a physics-informed pipeline that decouples key competences to overcome these challenges.

Longitudinal 3D medical imaging captures disease trajectories over time and is central to monitoring progression in many diseases. Existing self-supervised learning (SSL) methods for volumetric medical images largely operate in Euclidean latent spaces and do not explicitly encode temporal ordering. We propose a self-supervised framework that learns longitudinal representations in Lorentz hyperbolic space. For visit pairs from the same patient at different times, our model combines a hyperbolic contrastive objective with a temporal entailment loss and a time difference aware hyperbolic margin loss to enforce progression-consistent ordering in the time-like (radial) dimension. Experiments on MRI- and OCT-based degenerative disease cohorts show strong performance versus recent Euclidean and hyperbolic baselines, and ablations confirm that temporal constraints are important. To the best of our knowledge, this is the first hyperbolic representation learning approach for longitudinal medical imaging trajectories.

Time series forecasting of patient trajectories plays a critical role in the clinical environment by enabling the prediction of possibly treatment relevant patient events . Clinical data such as imaging studies, surgical records, laboratory measurements or tumor staging provide rich longitudinal information reflecting the progression of disease or treatment response. Modeling these data involves several challenges such as integrating multi-modal data, handling irregularly sampling over time, or managing missing values. Many existing forecasting approaches rely on regularly sampled data and perform poorly when facing irregularly sampled clinical data. Here, we evaluate three different deep learning models for predicting individual six month survival from irregularly sampled lung cancer patient trajectories. Results show that state-of-the-art models can integrate sparse clinical data and benefit from multi-modality, improving forecasting of clinical outcomes despite irregular sampling patterns.

Alzheimer’s disease (AD) is an inherently multimodal pathology driven by complex genetic and phenotypic interactions, making reliable early detection a critical challenge. Existing multimodal approaches often struggle to effectively align static baseline genetic risk with longitudinal physical changes. In this work, we introduce a novel two-stage contrastive learning framework integrating Single Nucleotide Polymorphisms (SNPs) and structural MRI volumes. To overcome the bottleneck of single-timepoint genetic measurements, we propose an age-conditioned augmentation strategy that generates time-aware genetic embeddings for longitudinal contrastive pairing. Utilizing a dynamic Gated Fusion mechanism for downstream classification, our approach effectively weights modality contributions. Evaluated on the ADNI database, our framework consistently outperforms strong classical baselines and state-of-the-art generative models, demonstrating particularly significant improvements in early-stage cognitive decline detection.

Intensive care units (ICUs) provide lifesaving treatments to patients with severe medical conditions, producing large amounts of clinical time-series data that reflect patient health trajectories. Forecasting future trajectory changes helps clinicians anticipate adverse events. While prior work addresses the challenges of missing values and irregularities in clinical time-series, designing effective forecasting architectures for such data remains an open research area. At the same time, limitations of Transformer-based models are prompting a renewed interest in recurrent architectures for processing time-series. Among them, the recently proposed xLSTM demonstrates strong forecasting capabilities across several domains, yet its potential for clinical use-cases remains largely unexplored. In this work, we address this gap by extending xLSTM to forecast irregular multivariate clinical time-series with missing values. To this end, we replace the temporal and cross-channel modeling components of an established forecasting architecture with xLSTM blocks. Our models achieve competitive predictive performance compared to several baselines on a subset of MIMIC-III, highlighting xLSTM's potential as a powerful backbone for clinical time-series forecasting.

Masked image modeling is a widely used self-supervised learning technique for pretraining and foundation models. In ophthalmology, most existing approaches are either unimodal or unpaired multimodal. In this study, paired color fundus photography (CFP) and optical coherence tomography (OCT) are utilized to investigate the impact of multimodal and unimodal pretraining on the downstream tasks. Experiment results show that multimodal pretraining generally improves downstream performance compared with unimodal pretraining, with notable gains observed on several CFP and OCT datasets.

Autonomous mobile robots must navigate dynamic environments safely, yet high end depth sensors are often expensive or impractical. Monocular cameras are widely available, but estimating metric depth and detecting obstacles in real time remain challenging. We address this by implementing a pipeline that combines monocular depth estimation with metric scale calibration, 3D back-projection, filtering, and clustering. Our marker-based calibration achieves a depth RMSE as low as 13mm, while the proposed pipeline successfully detects all 8 obstacles in our evaluation. With OpenVINO optimizations, the model achieves an inference rate of up to 17FPS, establishing a foundation for real-time processing. Overall, the pipeline demonstrates promising results for safe navigation using only monocular cameras on resource-constrained robots, evaluated in the context of the international robotic contest Eurobot.

Roadside vegetation management is vital for traffic safety, efficiency, and biodiversity. Conventional mowing relies on routine schedules and manual inspections, limiting route optimization and adaptation to changing vegetation growth. To address the challenges of roadside maintenance, we developed MeadowLevelSeg, a deep learning approach that employs Mask2Former to map meadow heights into precise 5 cm classes. Around 800 high-resolution roadside images were recorded and annotated. Performance is evaluated using a novel Distance-Aware Accuracy metric, which takes the ordinal nature of height classes into account. Initial results demonstrate that the model effectively identifies different meadow heights and high-growth zones, achieving a mean absolute error of less than 7 cm using monocular images. This provides a robust basis for automated maintenance scheduling.

Driver-related factors contribute to nearly 90% of traffic accidents. Estimating 3D driver poses can help track risky behaviors. However, the scarcity of annotated 3D pose data, together with the complexity and high cost of 3D annotation, limits the training of domain-specific estimators. We address this challenge by pre-training 2D-to-3D pose lifting models using synthetic 3D poses from a simulated dataset. In experiments on the Drive&Act dataset, we compare training from scratch with synthetic pre-training while gradually increasing the amount of real-world data. For example, when only 5% of training data is available, MPJPE is reduced from 90.0 mm to 70.9 mm for the GraFormer model. Our results demonstrate that synthetic pre-training consistently reduces estimation errors, particularly when real-world data are limited. Furthermore, synthetic pre-training improves the best fine-tuned results across different models from 48.1 mm to 46.0 mm in our tests.

As medical image databases expand, precise Content-Based Medical Image Retrieval (CBMIR) techniques are increasingly required to support case-based reasoning, clinical education, and data-driven decision-making. Recent deep learning–based CBMIR approaches typically rely on global embeddings to enhance retrieval performance. However, such image-level representations often dilute localized anatomical features and fail to capture clinically relevant organ-specific details. To address this limitation, we propose a region-based CBMIR framework that integrates organ-level information into both representation learning and retrieval. The ROI Embedding Selector extracts patch-level embeddings from user-specified regions of interest (ROIs). The Region-aware Organ Attention (ROA) module then learns structured organ representations through cross-attention between image patches and dedicated organ tokens. During inference, a visibility-weighted aggregation strategy guided by Organ Visibility Recognition incorporates query-relevant organs, enabling anatomically targeted and clinically meaningful retrieval. Experiments on the TotalSegmentator dataset demonstrate that the proposed framework consistently outperforms global embedding–based vision foundation models, particularly in region query settings.

Edge detection of complex, smooth, transparent, reflective and texture-less objects is an unsolved problem in computer vision. In this work, an existing approach using diffusion in the image space is adapted to enable fast and accurate edge detection. The method is applied to texture-less industrial objects from the T-LESS and XYZIBD datasets. The models are trained on datasets, generated synthetically using BlenderProc. Three training datasets are created using T-LESS objects to evaluate the impact of edge type and object texturing on prediction quality. Two more datasets are generated using XYZIBD objects to investigate the influence of the crease angle used in edge rendering. The diffusion models are evaluated using the NMSE, SSIM, DICE, and CRISP metrics, to assess accuracy, structural fidelity, and perceptual sharpness. Experiments show that our approach achieves competitive edge prediction quality and consistently outperforms existing diffusion based methods in computational efficiency at a lower resolution, while offering overall better prediction fidelity compared to the Canny edge detector. With a runtime of 95ms per image on an NVIDIA RTX3090, the approach demonstrates suitability for deployment in robotic vision systems. A quantitative edge prediction quality evaluation is conducted on real-world test sets which are extended with the edge ground-truth.

We discuss intelligent data augmentation strategies to help training object detection models from low-volume datasets. In particular, many industrial inspection tasks suffer from a lack of samples showing defects in the training data, and furthermore the failure cases are typically heterogeneous, leaving only a handful of samples for each of them. For our application scenario of printed circuit board (PCB) inspection, we propose and evaluate a strategy for synthesizing defects as well as a strategy to copy-paste difficult, challenging, or otherwise rare cases into the training images. Maintaining this library of challenging or rare cases offers an easy way to update the model and integrate feedback after deployment. We evaluate the benefits of the augmentation strategies in experiments and present a reliable and accurate PCB inspection model trained with only 25 images.

A non-destructive approach is proposed to assess the compressive strength of reclaimed bricks using short-wave infrared (SWIR) hyperspectral imaging (HSI) and a spectral–spatial 1D-Convolutional Neural Network (CNN). Hyperspectral images of 60 bricks, capturing both outer (weathered) and inner (pristine) surfaces, were analyzed. Regression reached R2 = 0.625, while a three class (low, medium, high) compressive strength classification achieved 83 % pixel level accuracy. At the brick level, aggregating predictions with a majority-vote scheme attained an accuracy of 91 % for outer and 98 % for inner surfaces. Score-CAM identified key wavelengths around 1200–1400 nm (moisture) and 2300–2500 nm (clay minerals) as driving the predictions. The results demonstrate that SWIR HSI can capture mineral- and moisture-related signatures relevant to compressive strength, offering a rapid, non-destructive screening tool for reclaimed bricks.

Recent diffusion-based 6-DoF grasp generation methods like GraspGen achieve state-of-the-art performance in simulation but face significant challenges when deployed on real robotic platforms. We present a unified adaptation pipeline for the Toyota Human Support Robot (HSR) that bridges these gaps without retraining the foundation model. Our approach combines symmetry-based point cloud completion to mitigate self-occlusion artifacts, three geometric feasibility filters that reduce motion planning failures from 66 % to 16 %, and a kinematic compensation for the HSR’s arc-shaped gripper trajectory. We show in our experiments, that our pipeline achieves an overall success rate of 85 % which is competitive with simulation of GraspGen while outperforming baselines M2T2 (56 %) and AnyGrasp (70 %) by up to 29 percentage points. Ablation studies confirm the necessity of each component: symmetry completion improves success by +13 percentage points, while geometric filtering enables 4× more grasp candidates to reach execution. These results demonstrate that post-hoc adaptations can unlock the real-world potential of simulation-trained grasping foundation models on diverse hardware platforms. The code and repository are available at:
https://github.com/Ziegenschmuggler/GraspGenforHSR

We consider the extraction of point-spread function (PSF) information for blind image deconvolution from blurred image in a way that preserves phase information, in contrast to using cues like autocorrelation that reveal only spectral information. Our approach is based on extracting suitable feature patches, depending on the type of images either spot highlights or edge segments. We discuss how edge patches in fact constitute a tomographic representation of the PSF. In integrating information from spot or edge patches into a PSF estimate, it is essential to compensate spatial misalignments. We achieve this by an iterative update rule that combines Fourier transformation with a nonlinear intensity transformation to achieve shift invariance.
Although designed with the goal of integration into alternating minimisation schemes, the two-step procedure of PSF estimation followed by non-blind deconvolution developed here performs surprisingly well as a fast blind deconvolution method in its own right on suitable image classes.

To enable the transition to cost-effective, real-time multi-
spectral sensors, this study introduces a novel Explainable
AI (XAI) framework for spectral band selection by adapt-
ing Score-CAM—typically used for 2D images—to 1D hy-
perspectral signals. This XAI-driven approach is rigor-
ously evaluated against established chemometric and ma-
chine learning baselines, including Weighted Regression Co-
efficients (WRC), Variable Importance in Projection (VIP),
and the Successive Projections Algorithm (SPA). This method
reduces data volume by over 90% while matching full-
spectrum baseline performance. Ultimately, this research
validates XAI as an interpretable, robust tool for design-
ing efficient, low-cost optical sorting systems for the circular
economy

Fragmented occlusion, as encountered in through-foliage observation, makes monocular motion estimation difficult because the target is visible only through sparse, discontinuous image fragments. We estimate motion by warping frames under a parametric model and maximizing the contrast of their integrated image. Although effective for 2DoF translation, this objective becomes ill-conditioned for 4DoF similarity motion. To analyze this, we derive a Fourier-domain reformulation that exposes the optimization structure and shows that static occlusion biases the objective toward zero motion. This motivates a decoupled 4DoF pipeline in which rotation and scale are estimated separately from translation. On synthetic videos with controlled fragmented occlusion, the Fourier formulation matches the spatial baseline at low-to-mid occlusion while converging faster, and the decoupled pipeline restores reliable translation recovery where joint 4DoF optimization fails.

This is work about using (polarization) filters in hand-vein biometric recognition. Experiments clearly demonstrate that the respective application of linear polarization, circular polarization, and band pass filters on the capturing lens improve hand-vein sample image quality across a considerable range of specific vascular image quality metrics. In case the illumination source is additionally equipped with a linear polarization filter (in relative perpendicular direction), further quality
improvement could not be demonstrated.

Motion capture systems are widely used to generate training data for machine learning and robotics applications. However, existing workflows often rely on fragmented toolchains, ad-hoc preprocessing scripts, and inconsistent data representations, which complicates reproducibility and rapid experimentation.

We present kinesis, an API-first schema-driven framework for capturing, validating, inspecting, and augmenting motion data based on structured 3D keypoints. The framework enforces consistent dataset structure through declarative skeletal schemas, enables interactive inspection through synchronized tabular and 3D visualization, and provides schema-aware augmentation operators for spatial and temporal transformations.

To demonstrate practical applicability, we implement a WebXR-based hand gesture capture pipeline using consumer-grade hardware (Meta Quest~3). While this setup serves as a concrete example, the framework is hardware-agnostic and extensible to other pose estimation sources. By unifying capture, validation, inspection, and augmentation in a single system, kinesis provides a reproducible foundation for motion-based machine learning workflows and rapid prototyping in XR and robotics scenarios.

The hospitality sector is facing a severe shortage of skilled personnel, which results in a significant need of automation and digitalization. In particular, automation of professional kitchen processes poses significant challenges due to the variability of commodities, the mixed presence of humans and machines, and harsh environmental conditions. The introduced concepts integrates a recipe-driven approach including warehouse intralogistics and automation for food processor tending. The ongoing work presented in this extended abstract reflects initial results of an in-depth conceptualization phase supported by simulation-based validation.

Simulation is increasingly used to accelerate the development of robotic manipulation systems, but integrating simulation environments with robotics frameworks remains challenging. This paper presents a framework that connects ROS 2 with NVIDIA Isaac Sim to support perception-driven dual-arm manipulation. A simulated workcell with two UR10 manipulators, parallel grippers, and an RGB-D camera is implemented to study system integration. The framework is evaluated on cooperative manipulation of rigid objects and textile manipulation using cloth simulation. While rigid object handling is achieved after careful parameter tuning, textile manipulation exposes limitations in current cloth simulation, particularly in reliable grasping. The paper discusses key integration challenges such as interface alignment, temporal and spatial synchronization, and coordinated dual-arm motion planning, providing practical insights for building reproducible simulation pipelines.

Tactile sensing is essential for dexterous robotic manipulation, enabling reliable contact detection, grasp assessment, and safe interaction with delicate objects. In this work, we present a finger-shaped tactile sensor system based on a 2D array of MEMS barometric pressure sensors, designed to mimic the compliance and geometry of the human fingertip. The system integrates real-time contact force measurements utilizing the pressure sensor array, in combination with acceleration data from an onboard Inertial Measurement Unit (IMU), allowing both precise point-of-contact estimation and dynamic impact detection. A dedicated microcontroller (μC) acts as a local processing and coordination node, responsible for closed-loop grasp and movement control, while a PC manages high-level communication between the μC and a robotic gripper. In addition, a hardware-level GPIO handshake between the control unit of a collaborative robot and the processing node enables deterministic synchronization between robotic arm positioning and grasp execution. Experimental validation of both the tactile sensor system and the robotic gripper control demonstrates robust operation across the conducted performance tests, with no malfunctions or object damage, as tactile feedback enables real-time grasping control throughout object manipulation. These results highlight the advantages of our tactile sensing solution as a cost-effective, versatile approach for enhancing robotic touch and advancing adaptive object-handling strategies.

When employing robots for tasks such as polishing or surface grinding, vibration and peak interaction forces exerted at the robot or the part must be limited. Therefore, such tasks are performed with the help of Active Contact Flanges (ACF). These devices are force controlled and enable fast processing speeds. High contact velocities of the tool result in significant impact forces during interaction with the environment. Consequently, a critical control aspect is the reliable estimation of these forces, which is the focus of this paper. A dynamical model is developed, resulting in a linear time invariant equation of motion, which is solved analytically. Hereby, a homogeneous and a particular solution is derived. The maximum contact force is then determined based on an optimization. Experimental results demonstrate high consistency between measured and calculated contact forces.

Automated garment handling in textile recycling remains challenging due to the deformability of textiles, their high shape variability, frequent self occlusion, and the presence of foreign objects in cluttered heaps. This paper presents a Multi- Modal robotic sorting system that combines semantic visual perception with tactile grasp monitoring. The proposed approach integrates Visual Language Model (VLM) based garment classification, Convolutional Neural Network (CNN) based grasp prediction using RGB-D images, and capacitive tactile fingertips mounted on a parallel gripper to detect grasp success, object loss, and approximate weight during manipulation. The estimated weight serves as a plausibility measure for the visually predicted garment class and as a coarse indicator of garment size. To support safe execution, a Digital Twin implemented in MoveIt2 is used for motion planning and collision avoidance in a synchronized real and virtual environment. A classification accuracy of up to 87.89 % across six classes was achieved in an experimental robotic sorting scenario including 219 items. Furthermore, the tactile finger sensor is evaluated under wet conditions and in contact with wet textiles to assess robustness, showing reliable sensing behavior even in these challenging scenarios. Overall, the results demonstrate the potential of combining semantic vision and robust tactile sensing for dependable textile sorting in recycling applications.

As Artificial Intelligence (AI) systems are increasingly deployed in safety-critical and societally consequential contexts, the question of how to evaluate their performance in a trustworthy and interpretable manner becomes increasingly important. Within the European Union, this issue is reflected in the AI Act, which requires training, validation, and testing datasets to be relevant and sufficiently representative with respect to the system's intended purpose. This raises a fundamental technical question: representative of what population of situations?

From a statistical perspective, performance metrics such as error rates or expected losses are always defined with respect to a probability distribution. We refer to this distribution as the Application Domain (AD). In practice, however, the AD of real-world AI systems is rarely known in explicit mathematical form and must instead be characterized operationally through the procedures by which valid samples are generated or selected.

To address this problem, we introduce the Stochastic Application Domain Definition (SADD), a textual specification of the sampling protocol that induces the distribution under which performance claims are intended to hold. The SADD links intended use, dataset construction, and statistical testing by making the underlying data-generation assumptions explicit. We formalize the notion of protocol-induced distributions, discuss how SADDs guide feasible sampling procedures, contrast the approach with qualitative domain descriptions such as Operational Design Domains, and examine implications for the certification of AI systems.

Passing as human in a room full of people requires more than fluent speech, it demands reading the room. While Large Language Models (LLMs) have transformed human-AI interaction in a one-to-one setting, they still fall short in multi-user conversational settings where social dynamics define the interaction. In such environments, a conversational agent that merely generates coherent text will struggle to maintain consistent socially plausible behavior. We propose a structured Theory of Mind (ToM) framework that equips conversational agents with the cognitive machinery to reason over participant beliefs, intentions, and evolving group dynamics in real time. Rather than relying on a single LLM prompt, our architecture decomposes social reasoning into explicit modules---a knowledge base, belief system, goal generator, and intention planner---coupled with a dual-process response architecture that balances immediacy with strategic depth. To evaluate this approach, we deployed the framework within the Turing Game and Reverse Turing Game environments, further enhancing the agent's plausibility with a simulated human-like response timing algorithm. Preliminary evaluations demonstrate that our ToM-equipped agent exhibits stronger conversational coherence, sustains longer exchanges, and is less frequently identified as a bot compared to its predecessor without structured social reasoning.

Maritime Autonomous Surface Ships (MASS) challenge safety frameworks originally developed for conventionally crewed vessels. Although autonomous navigation algorithms have advanced significantly, a critical gap remains in the hardware and control architectures required to deploy them safely in real maritime environments. This paper examines the legal and operational constraints affecting MASS under current international maritime frameworks and reviews the associated challenges of multi-sensor perception and remote human--machine interaction. To address these issues, the study applies System-Theoretic Process Analysis (STPA) to identify unsafe control actions and derive safety constraints at the organizational and supervisory control levels. Based on these results, the paper proposes a safety-driven hardware and control architecture for automated surface vessel systems. The architecture is intended to function as an assurance layer around AI-enabled autonomy by combining hardware redundancy, real-time diagnostic monitoring, independent safety controllers, and mechanisms for safe supervisory intervention. In doing so, it provides the structural basis for fault-tolerant operation, controlled degradation, and transition to a minimum-risk condition under abnormal or degraded conditions.

Anthropomorphic terminology with respect to artificial intelligence systems has become commonplace both in AI expert and non-expert user circles. While anthropomorphic terminology in general has deep roots and has been widespread in many areas of human life, it poses significant risks, ranging from misguided expectations to ill-considered legislation, when applied to artificial intelligence. This article aims to contribute to a better understanding of AI systems at a fundamental level by analyzing some of the most widely used anthropomorphic terms in AI: "reasoning", "autonomy", and "understanding". While admitting that avoiding the use of anthropomorphic terminology in AI seems impossible, the authors aim to equip non-technical, particularly legal, experts with knowledge and understanding that would assist them in their professional engagement with AI systems.

The increasing availability of machine learning algorithms has posed the challenge of selecting appropriate algorithms for specific data analysis tasks. In domains such as education and business, where many practitioners are not specialists in
artificial intelligence, algorithm selection is often performed through trial-and- error experimentation or guided by limited methodological knowledge. Meta- learning has emerged as a promising approach for addressing this challenge by recommending algorithms based on characteristics of previously analysed datasets.
However, many meta-learning approaches rely on complex models whose decision processes remain difficult to interpret, limiting their suitability in contexts where transparency and accountability are required. This paper investigates the use of
explainable meta-learning models for machine learning algorithm selection in social science domains. Using datasets originating from education and business contexts, we construct a meta-dataset based on dataset characteristics represented
as meta-features. These meta-features serve as inputs to interpretable meta-models designed to recommend suitable algorithms for new datasets. We analyse the contribution of individual meta-features to the meta-model decisions, thereby
identifying dataset characteristics that drive algorithm recommendations. The results demonstrate that a subset of meta-features plays a key role in determining the
predictive power of the meta-model and forms the basis for explainable algorithm selection. By making these relationships explicit, the proposed approach enables transparent and interpretable recommendations that can support non-expert users in selecting appropriate analytical methods. The study contributes to discussions on trustworthy and responsible AI, particularly relevant in the context of emerging AI governance frameworks and certification initiatives that emphasise explainability, accountability, and user trust in AI systems.

Pain indicators and preparturient nest-building are precursors of farrowing, yet continuous quantification remains challenging. We present a non-invasive computer-vision system that detects pain-associated behaviors (back-arching, tail-flicking, back leg forward, trembling) and nest-building behaviors (manipulation of pen components, pawing, exploration) from top-view videos. We analyzed 748 h of RGB footage (25 fps) from 11 sows on a single farm, spanning 64 h pre-farrowing to 4 h post birth of the first piglet. Using a defined ethogram, 46,010 events were annotated with inter-annotator agreement κ = 0.724. To assess generalization, we used a sow-level split, i.e., 8 in the training set and 3 in validation. Behaviors were detected with a modified DeepEthogram architecture, combining RGB data and optical flow. Both streams were processed by separate ResNet3D-34 encoders. Optical flow was estimated using a state-of-the-art DPFlow model. Training employed focal loss to address class imbalance, alongside geometric and photometric augmentations for robustness to camera placement and lighting. Clips of 11 frames at 8.33 fps (≈1.32 s) were used. On held-out sows, per-class F1 scores were 0.875 for manipulation of pen, 0.634 for pawing, 0.820 for exploration, 0.639 for back-arching, 0.778 for tail-flicking, 0.871 for back leg forward, and 0.443 for trembling. These results indicate that pen-installed vision can identify key behaviors in a non-invasive way, supporting scalable monitoring. Limitations include modest dataset size and limited diversity, i.e., a single farm and a single breed. Ongoing work will expand the dataset and leverage behavior dynamics for time-to-farrowing estimation.

Problem weeds pose a challenge for agriculture. Robust detection of these plants is crucial for their control and for assessing possible contamination of the crop. Current UAV inspections are usually carried out using a fixed flight route, regardless of the extent of weed infestation. Initial approaches of adaptive flight control attempt to save flight distance by adapting flight altitude and route. We would also like to investigate adaptive gimbal guidance, to find out whether this can have a positive impact on flight paths and flight times. To this end, we developed a pipeline between the server including the operating website and the drone. The initial results, presented in this extended abstract, examine the functional capability of the pipeline and different error sources for GPS accuracy of adaptive gimbal pitches.

Eye diseases in dogs are visually similar and difficult to distinguish without profes-
sional examination. Furthermore, assessing the severity of such conditions – and in
particular determining whether immediate veterinary attention is required – poses
a significant challenge for pet owners. To address this problem, we aim to assist
pet owners in performing an initial triage of canine eye conditions with minimal
technical expertise required. Given a photograph of the affected eye taken with a
smartphone, we provide an automated preliminary assessment indicating whether a
veterinary visit is advisable. To this end, we employ a computationally efficient
convolutional neural network (CNN) to classify the images, identifying potential
conditions and reporting the result to the user.

In veterinary medicine, urine specific gravity (USG) is among the most commonly
used indicators of renal function, typically assessed by means of manual refrac-
tometers. However, accurate readings require a trained eye and adequate lighting
conditions, limiting their use outside of clinical settings. This study investigates
whether digital refractometers can serve as a reliable alternative, and in particular
whether pet owners are able to use them independently for continuous at-home
monitoring of their animal’s USG. With a mean deviation of 0.0014 ± 0.0008
between devices, the results demonstrate that digital refractometers yield measure-
ments comparable to those of manual devices, paving the way for broader at-home
use by pet owners.

Comprehensive veterinary care is of central importance for animal health in agricul-
ture. However, veterinary services are often unavailable at night and on weekends.
To better represent availability and deploy existing resources more efficiently, we
collected publicly available data on veterinary practices in Austria and stored them
in a continuously updated database. To make this information accessible, we inte-
grated it into a web app based on OpenStreetMap, providing relevant information
for both animal owners and veterinarians. The app allows users to visualize the
structure of veterinary services across Austria and to find the nearest available prac-
tices. Route planning is based on real street distances and travel times, computed
using the Open Source Routing Machine (OSRM), ensuring that geographical
constraints such as mountain passes or river crossings are properly accounted for.
The system is designed to be extensible, with future versions planned to incorpo-
rate real-time availability updates, veterinary specializations, and seasonal road
conditions.

Hybrid approaches combining differential equations and machine learning, commonly referred to as physics-informed machine learning, have gained significant attention in recent years. Prominent examples include Physics-Informed Neural Networks (PINNs) and Physics-Informed Gaussian Processes (PIGPs), the latter naturally providing uncertainty quantification. PIGPs encode differential constraints directly in the covariance kernel, and existing approaches can be roughly grouped into two schools of thought.
Operator-based constructions apply differential operators to a base kernel, yielding systematic and algorithmic methods, but often have restrictions of which systems can be represented, or performance issues. In contrast, Mercer-type constructions build kernels from problem-specific solution components such as Green’s functions or fundamental solutions; while typically data efficient, they rely on analytical insight and substantial manual derivation. We propose Monge-GPs, a hybrid construction based on Monge parametrization that unifies operator-based kernels and Mercer kernels using fundamental solutions. By parametrizing the controllable dynamics algorithmically and restricting problem-specific design to a low-dimensional autonomous component, the approach substantially reduces the need for manual kernel design, and stays data efficient while lifting the restriction to controllable systems.

Numerical integration and ODE discovery are two sides of the same coin—converse problems of finding trajectories from known physics versus inferring physics from observed trajectories. Although these problems have been extensively studied in isolation, they can be unified through the minimization of a common quantity: the Euler–Lagrange residual. In this paper, we build on this insight and introduce the Integrated Squared Action Residual (ISAR), which enables both tasks to be performed simultaneously. We formulate numerical integration and model discovery as a joint Bayesian inference problem, allowing for the systematic incorporation of physical prior knowledge and domain constraints in settings with sparse and noisy observations, where traditional approaches often fail. While we demonstrate the performance on two mechanical toy problems, it can be readily extended towards multiphysics systems including dissipative dynamics.

It was recently shown that the loss function used for training physics-informed neural networks (PINNs) exhibits local minima at solutions corresponding to fixed points of dynamical systems.
In the forward setting, where the PINN is trained to solve initial value problems, these local minima can interfere with training and potentially lead to physically incorrect solutions.
Building on stability theory, this paper proposes a regularization scheme that penalizes solutions corresponding to unstable fixed points. Experimental results on four dynamical systems, including the Lotka-Volterra model and the van der Pol oscillator, show that our scheme helps avoiding physically incorrect solutions and substantially improves the training success rate of PINNs.

Accurate surrogate modeling in engineering is often constrained by the high computational cost of generating training data from large scale numerical simulations. In many industrial applications, only a limited number of simulations can be afforded, which severely restricts the achievable surrogate accuracy, particularly in high dimensional parametric spaces. A promising approach to mitigate this curse of dimensionality is the incorporation of derivative information into surrogate training, which can be obtained efficiently via graph based implementations or adjoint calculations. This additional information captures local function structure, offering the potential to significantly improve data efficiency. In this work, we quantify the potential gains of derivative-enhanced training both theoretically and numerically, using a representative linear elasticity problem and an analytical benchmark. The findings provide guidance on the efficiency improvements achievable and the order of derivatives that yields the greatest benefit.

This extended abstract presents a physics-informed reinforcement learning framework for optimal diffuser design to improve airflow homogeneity upstream of a heat exchanger. This approach addresses key challenges in simulation-based optimization, including high-dimensional design spaces, expensive CFD evaluations, and the lack of gradient information. Physics-based flow features related to early pressure loss occurrence and eddy formation were employed as low-cost proxies for the target homogeneity objective. The problem is formulated as a partially observable Markov decision process in which the agent sequentially selects the geometries to be evaluated. Using an expected improvement reward function, the method adaptively balances exploration and exploitation. The approach is demonstrated on a synthetic one-dimensional example, and a two-dimensional diffuser optimization problem is presented.

Accurate estimation of active stress parameters is essential for understanding cardiac function, but remains clinically challenging when only imaging-derived displacement and strain data are available.
We present a Physics-Informed Neural Networks (PINNs)-based framework for inferring active contractility in time-dependent cardiac biomechanical models directly from such data. The approach incorporates adaptive weighting, residual-based attention, Fourier features, and tailored regularisation strategies, enabling robust reconstruction of active stress fields under noisy conditions and at high spatial resolution. A Pareto front analysis is conducted to assess the influence of loss weight selection on parameter estimation. The algorithm is further validated on tissue inhomogeneity detection, with potential clinical impact in conditions such as cardiac fibrosis and myocardial infarction.

Bayesian optimization (BO) enables data-efficient optimization of expensive black-box functions, yet real-world applications often provide partial domain knowledge from physical models or expert intuition. It remains unclear how the placement of such knowledge within the BO pipeline affects performance in small-data regimes. We present a systematic empirical comparison of three knowledge-injection strategies: surrogate-level integration, probabilistic priors over the search space, and acquisition-level regularization. The approaches are evaluated across different synthetic and materials-inspired multi-objective tasks under limited evaluation budgets using cumulative normalized hypervolume regret. Results show that surrogate-level injection can accelerate early learning when knowledge is well aligned but is sensitive to misspecification, whereas acquisition-level regularization provides more robust guidance under imperfect knowledge. Prior-based strategies yield only modest improvements over standard BO.

Conceptual-model-guided physics-inspired feature engineering (CPFE) a lightweight representation-level strategy that formalizes the incorporation of domain knowledge into machine-learning pipelines. Instead of embedding physics in the loss or model architecture, the proposed approach structures physical and operational knowledge at the representation level through traceable domain concepts.The method is demonstrated in a 3D-printer case study, where the resulting physics representation achieves better performance than a compact baseline and automatically extracted features using tsfresh.

For many scientific and engineering problems, equations based on applicable laws of physics can be used to link observable physical quantities. Analytic expressions, however, provide only point estimates and therefore cannot express uncertainty. This limits trustworthiness of predictions, especially in setups with limited data, noisy observations or when extrapolating. Bayesian probabilistic models address this limitation by treating unknown model parameters as random variables initialized by prior distributions and yielding - through inference - posterior (predictive) distributions. Constructing Bayesian models and convergence of inference, however, still requires specialized knowledge in probabilistic programming and inference algorithms, hindering the broader adoption of Bayesian models and uncertainty quantification in many domains. To make uncertainty-aware equation modeling more accessible, we present Equayes (Equation to Bayesian Model), a scikit-learn-style estimator that converts a user-provided symbolic expression into a probabilistic model and performs posterior inference over its numerical constants. The core value of the method and tool to construction of hybrid models is that it implements a principled approach to hybrid model evaluation, linking laws of physics, random variables and inference in an accessible manner.

Robotic manipulation of unseen objects relies on zero-shot 6D pose estimation, which typically requires a 3D mesh as a reference. While constructing accurate meshes requires specialized scanning hardware and manual editing, recently proposed Novel View Synthesis (NVS) techniques, such as 2D Gaussian Splatting (2DGS) and Sparse Voxels Rasterization (SVRaster), produce accurate surface reconstructions as a byproduct, potentially eliminating the need for specialized equipment. This work presents an automated image-based mesh generation pipeline that integrates object segmentation, camera registration, point cloud generation, metric height estimation, and NVS mesh generation, eliminating the need for expensive hardware and human intervention. Leveraging 2DGS and SVRaster with MASt3R-SfM or Visual Geometry Grounded Transformer (VGGT), the pipeline produces accurate meshes in minutes, with the VGGT/SVRaster combination reducing reconstruction time to seconds. Grounding near-view object-centric images with far-view scanning scene images using MASt3R yields consistent object height estimates. On the BOP YCB-V benchmark, meshes generated with our pipeline achieve competitive performance with state-of-the-art zero-shot pose estimation methods. Real-life robotic grasping experiments further indicate robust performance even under moderate scale errors. The source code is available at https://github.com/St333fan/meshgen-zeroshop.

Detecting unseen anomalies in unstructured environments presents a critical challenge for industrial and agricultural applications such as material recycling and weeding.
Existing perception systems frequently fail to satisfy the strict operational requirements of these domains, specifically real-time processing, pixel-level segmentation precision, and robust accuracy, due to their reliance on exhaustively annotated datasets.
To address these limitations, we propose a weakly supervised pipeline for object segmentation and classification using weak image-level supervision called \textit{Patch Aggregation for Segmentation of Targets and Anomalies} (PASTA).
By comparing an observed scene with a nominal reference, PASTA identifies Target and Anomaly objects through distribution analysis in self-supervised Vision Transformer (ViT) feature spaces.
Our pipeline utilizes semantic text-prompts via the Segment Anything Model 3 to guide zero-shot object segmentation.

Evaluations on a custom steel scrap recycling dataset and a plant dataset demonstrate high efficiency compared to traditional approaches.
While being domain-agnostic, our method achieves target segmentation performance of up to 88.3% IoU and anomaly segmentation performance of up to 63.5% IoU across industrial and agricultural domains.

The rapid identification of minerals is critical for real-time geological analysis. This study investigates the efficacy of machine learning models in classifying mineral and rock samples using high-speed Raman sensors. We evaluate three distinct data representation strategies: (1) 1D spectral profiles, (2) 2D Raman spectral images, and (3) a fused multi-modal approach combining both spatial and spectral features. Using a diverse dataset of geological samples, we benchmark several model architectures to determine the trade-offs between computational efficiency and classification accuracy. Our results demonstrate how spatial context from imaging can enhance identification compared to traditional 1D methods, while also identifying the scenarios where signal-only processing remains optimal. This work provides a framework for selecting the most effective data representation for high-speed, automated mineralogical mapping.

Interactive robot learning methods typically treat the human teacher as an infallible oracle, limiting the agent's ability to surpass the expert or reject adversarial advice.
We introduce MAGIC (Modulated Asymmetric Games for Interactive Control), a framework that formulates interactive learning as an asymmetric leader--follower game between a Teacher and a Learner.
The Teacher is an inverse reward field---instantiated with energy-based and flow-matching heads---that scores trajectory segments in SE(3) via contrastive learning on expert demonstrations.
The Learner is a hierarchical flow-matching policy (Eye, Brain, Muscle) that maximizes a shaped reward mixing environment reward and Teacher signal.
A gradient-agreement coupling determines state-dependent trust: when the Teacher's directional signal agrees with the task critic's gradient, the Teacher is trusted; otherwise it is ignored.
We prove that the alternating update satisfies the regularity conditions of two-timescale stochastic approximation.
The core pipeline is implemented and unit-tested; we present the framework, its theoretical grounding, and the planned experimental evaluation on 9 ManiSkill3 manipulation tasks, LIBERO with noisy human demonstrations, and real-robot transfer on UR3e and SO-101 arms.

Gaussian Splatting (GS) has recently emerged as a promising representation for real-time 3D reconstruction. While several GS-based SLAM systems have been proposed, their robustness in practical RGB-D scenarios remains insufficiently understood. In this work, we present a compact evaluation of representative GS-SLAM pipelines and analyze the role of tracking components within a GS-based system. This study is motivated by the need to identify reliable geometric foundations for future semantic scene representations built on top of GS-based SLAM. Our results indicate that tracking accuracy remains the dominant limitation, while reconstruction quality can be partially decoupled from trajectory correctness. We further observe that preprocessing strategies, such as downsampling, significantly impact both runtime and accuracy. Based on these findings, we identify key failure modes and discuss implications for improving robustness in GS-based SLAM systems.

Spiking Neural Networks (SNNs) offer an energy efficient alternative to conventional Artificial Neural Networks (ANNs) but typically still require a large number of parameters. This work introduces Linearized Bregman Iterations (LBI) as an optimizer for training SNNs, enforcing sparsity through iterative minimization of the Bregman distance and proximal soft thresholding updates. To improve convergence and generalization, we employ the AdaBreg optimizer, a momentum and bias corrected Bregman variant of Adam. Experiments on three established neuromorphic benchmarks, i.e. the Spiking Heidelberg Digits (SHD), the Spiking Speech Commands (SSC), and the Permuted Sequential MNIST (PSMNIST) datasets, show that LBI based optimization reduces the number of active parameters by about 50% while maintaining accuracy comparable to models trained with the Adam optimizer, demonstrating the potential of convex sparsity inducing methods for efficient neuromorphic learning.

Spiking neural networks (SNNs) have emerged as a biologically plausible computational paradigm with strong links to real-world brain dynamics. Recently, interest has grown in parallelizable State Space Model (SSM)–inspired architectures, which offer improved scalability compared to recurrent networks. While effective at scale, these models represent a step away from biological realism. In particular, the impact of removing recurrent connections and membrane nonlinearities on the temporal processing capabilities of SNNs remains largely unexplored. In this work, we investigate the impact of these changes to the network dynamics on the temporal processing capabilities of SNNs with a focus on recurrent connectivity. To this end, a suite of sequential tasks was used to systematically compare parallelizable SSM-style networks with recurrent SNNs. The results demonstrate that while parallelizable models perform well on tasks with simple or weak temporal dependencies, they struggle to maintain persistent internal state when complex, state-dependent computation is required. In contrast, recurrent architectures exhibit superior memory retention and robustness under these conditions. These findings suggest fundamental limitations of parallelizable SSM-style approaches for sequence tasks that rely on long-term internal memory, highlighting the continued relevance of recurrence in spiking neural computation as suggested by biology.

Backpropagation through time (BPTT) remains the gold-standard for training recurrent spiking neural networks, but its need to store long temporal computation graphs makes it memory-intensive and incompatible with strict online updates. This has motivated a range of alternative online learning rules, such as e-prop, further trace-based methods, and forward-only approximations, which reduce sequence-length-dependent overhead but typically require custom implementations and often sacrifice task performance. In this work, we revisit the simplest possible alternative: truncated BPTT with truncation length k = 1 (tBPTT1). Although this setting is usually regarded as an overly limited-horizon baseline with poor temporal credit assignment, we show that it is a widely underestimated learning strategy. In a standard surrogate gradient learning setup, tBPTT1 achieves performance competitive with or better than more sophisticated online learning rules. Our experiments identify two key ingredients for this result: a substantially smaller learning rate than commonly used and an optimizer with slow temporal averaging through its momentum statistics. These findings suggest that, for many practical spiking network settings, elaborate online credit-assignment rules may not be necessary: plain one-step backprop, when paired with appropriate optimization, appears as an overlooked training strategy provides effective, memory-efficient, and implementation-friendly learning.

Training recurrent spiking neural networks (SNNs) with leaky integrate-and-fire (LIF) neurons is often slow, particularly during the early phase, when networks must first establish sufficient spike activity to form patterns.
Strategies such as low firing thresholds or high-magnitude weight initialization can increase early spiking, but typically introduce instabilities and impair learning.
Here we introduce a modification of classical LIF and parameterized LIF (PLIF) neurons, in which spikes are generated probabilistically, including a proper surrogate gradient formulation. The membrane potential parameterizes the instantaneous spike probability, and spikes are sampled as Bernoulli variables at each time step, whereas underlying LIF membrane dynamics remain unchanged.
This stochastic activation stabilizes early spike activity and substantially accelerates learning.
In two benchmark tasks, these probabilistic LIF networks surprisingly achieve substantially higher classification accuracy than its deterministic LIF baselines.
These findings suggest that probabilistic spike generation may provide a promising new perspective for building compact and effective spiking architectures.

Working memory --- the ability to store and recall precise temporal patterns of neural activity --- remains an open challenge for spiking neural networks (SNNs). We propose a recurrent SNN of $N$ neurons in which each synapse is equipped with $D = 41$ delays, modelled as a weight tensor $\mathbf{W} \in \mathbb{R}^{N \times N \times D}$ and trained end-to-end with surrogate-gradient backpropagation through time. The network stores $M$ arbitrary target spike patterns by representing each as a sequential chain of overlapping Spiking Motifs: contiguous windows of length $D$ that uniquely predict spikes at the next time step. On a synthetic benchmark of $M=8$ patterns ($N=512$ neurons, $T=1000$ steps), training achieves a mean F1 score of $0.966$, with recall emerging first near the clamped initialisation window and propagating forward in time. This result demonstrates that heterogeneous delays provide an efficient substrate for working memory in SNNs, enabling energy-efficient neuromorphic edge deployment.

Trajectory prediction is an essential component of the perception stack in autonomous mobile robots (AMRs).
AMRs operate in complex environments where their movements are influenced by various environmental elements, such as racks and storage locations.
Therefore, accurate and efficient trajectory prediction for intralogistics requires detailed environment modeling that goes beyond the lane-based context commonly used for road traffic.
We propose a new environment context encoder that can be seamlessly integrated into state-of-the-art motion forecasting models.
Our approach, tailored to the specific challenges of intralogistics, achieves highly accurate predictions using efficient baseline networks.

Learning dense, pose‑aware object descriptors is a key ingredient for generalizing robotic manipulation across novel instances and viewpoints. Intermediate features from self‑supervised models like DINO and Stable Diffusion can serve as powerful dense descriptors for semantic correspondence, yet these features degrade under large viewpoint changes.
To address this, we introduce D²DINO, a descriptor prediction model for pixel level object understanding. Our model attaches a lightweight convolutional head to a frozen DINOv3 encoder and trains it to produce low‑dimensional (16‑D), pixel‑wise descriptors at full input resolution. The head fuses multi‑scale ViT features and progressively upsamples them, yielding compact descriptors that can be used directly for dense matching. Supervision comes from Normalized Object Coordinate Space (NOCS) annotations exploiting consistent 2D–3D mappings across frames.
We optimize D²DINO with a contrastive objective and further distinguish between negatives on other objects or background and negatives on the same object, down‑weighting the latter to encourage intra‑object variation. We show that D²DINO yields higher point matching accuracy than raw DINOv3 features with upscaled inputs, while requiring only a single forward pass at the original image resolution and a much lower descriptor dimensionality.

Numerous LiDAR-based Simultaneous Localization and Mapping (SLAM) algorithms have been developed for structured indoor and urban environments over the last decades. The applicability to large-scale, GNSS-denied, and featureless environments remains insufficiently investigated. This paper evaluates whether open-source SLAM algorithms can effectively address these challenges. A high-fidelity simulation of the Zentrum am Berg tunnel facilities in Austria is leveraged to benchmark open-source LiDAR-based SLAM algorithms and conduct a robustness evaluation by varying robot motion dynamics, sensor configurations, and environmental feature density. Results show that with appropriate motion and sensor settings, open-source SLAM can achieve reliable performance in large-scale, GNSS-denied, featureless environments.

Event cameras offer high temporal resolution and high dynamic range, making them suitable for localization under fast motion and challenging lighting conditions. Meanwhile, LiDAR maps provide accurate geometric representations of the environment. However, aligning sparse event observations with dense LiDAR maps remains difficult due to the significant modality gap between the two sensing modalities. In this work, we address event-based localization in LiDAR maps through cross-modal correspondence learning. We formulate event–LiDAR alignment as a dense event–depth flow estimation problem, which enables recovering 2D–3D correspondences between event observations and projected LiDAR points for pose estimation. Building on this formulation, we further introduce an edge-aware learning framework that jointly estimates event–depth flow and structural edges. The two tasks are coupled through cross-task feature fusion and iterative refinement, allowing the model to exploit modality-invariant geometric structures for improved correspondence estimation. Experiments on public event-based datasets demonstrate improved cross-modal alignment and localization accuracy.

Autonomous navigation in densely vegetated off-road environments remains challenging because conventional geometric perception often treats traversable vegetation as non-traversable obstacles. In this work, we present a modular semantic–geometric perception pipeline for vegetation-aware navigation. The approach combines camera-based semantic data with LiDAR to generate a local grid map containing geometric and semantic information. A subsequent filtering stage uses this representation to correct vegetation-induced artifacts in standard elevation maps while preserving rigid obstacles for navigation. The system is designed to be portable across multiple robot platforms and sensor configurations. The pipeline was evaluated in challenging alpine off-road environments on three robot platforms, indicating improved distinction between traversable vegetation and solid obstacles and supporting more reliable navigation in dense natural environments.

Visual-based localization has made significant progress, yet its performance often drops in large-scale, outdoor, and long-term settings due to factors like lighting changes, dynamic scenes, and low-texture areas. These challenges degrade feature extraction and tracking, which are critical for accurate motion estimation.
While learning-based methods such as SuperPoint and SuperGlue show improved feature coverage and robustness, they still face generalization issues with out-of-distribution data. We address this by enhancing deep feature extraction and tracking through self-supervised learning with task-specific feedback. Our method promotes stable and informative features, improving generalization and reliability in challenging environments.