Using autonomous, decentralised control of unmanned aerial vehicles (UAVs) for critical scenarios such as bushfire firefighting provides clear safety and cost benefits. In this paper, we introduce a novel framework for autonomous, decentralised control based on multi-agent reinforcement learning (MARL) and a planning heuristic, designed to support learning a search-and-interact task in a multi-particle environment, which models a range of multi-UAV control applications. Our approach hybridises a planning heuristic with generic MARL methods under a discrete multi-action paradigm. The planning heuristic extracts only focused information from the environment, while also augmenting agent-centric observation data via exploiting observation-action symmetry to improve training. We employ this heuristic within a Centralised and Training Decentralised Execution (CTDE) paradigm to train autonomous control that is agnostic to the complexity of the environment, and effective under a range of action dynamics. The controller is encoded in a lean policy with reduced footprint, and can be deployed within resource-limited devices. Experimental results show that using even simple Q-learners within the framework provides highly effective agents for accurate firefighting in complex scenarios. Our approach is applicable to decision problems with similar search-and-interact characteristics.
The frequentist two-sample t-test for comparison of two means is one of the most widely used hypothesis tests in practice. Recently, several Bayesian alternatives to the t-test were derived and shown to have some advantages, such as incorporation of prior expert knowledge, over the standard methodology. This manuscript introduces a new Bayesian t-test based on the minimum message length (MML) principle of inductive inference. The aim of the manuscript is to introduce the general reader to MML and demonstrate how MML can be applied to a problem of central importance in statistics.
Physical literacy, encompassing movement skills, coordination, and confidence, is a critical yet underexplored domain in both sports and computer vision. In this paper, we introduce PLNet-12, a novel video dataset consisting of 12 fundamental movement skills designed to support the evaluation of physical literacy through multi-task vision-language models (VLMs). Unlike existing datasets, PLNet-12 includes fine-grained annotations for temporal action detection, repetition counting, displacement and distance measurement, duration estimation, and physical literacy scoring. We focus on zero-shot evaluation using large-scale VLMs without task-specific fine-tuning to investigate their generalization ability in real-world settings. To ensure annotation quality, we employ a semi-automated, rule-based pipeline verified by motion capture data and domain experts. We benchmark several state-of-the-art VLMs under different temporal and spatial settings and reveal challenges including fine-grained temporal understanding, action confusion, and temporal hallucination. Our findings highlight critical limitations of current VLMs in structured physical tasks and suggest directions for future multi-modal reasoning research. PLNet-12 offers a lightweight yet challenging platform for advancing temporal video understanding in human movement analysis.
Federated Learning (FL) is a decentralized paradigm that enables collaborative model training without sharing raw data. However, traditional FL methods often struggle with slow and unstable convergence, especially under non-Independent and non-Identically Distributed (non-IID) data, which degrades model performance and training stability. Most existing aggregation strategies rely on static, uniform, or heuristic-based weighting schemes that fail to adapt dynamic client behavior and overlook variations in local training quality. Although some recent approaches explore client selection or value-based reinforcement processes (e.g., Q-learning), they fall short of directly optimizing the aggregation process in an end-to-end fashion. To address these limitations, we propose Federated Reinforcement Weighted Aggregation (FedRWA), a policy gradient-based aggregation approach that dynamically assigns client-specific weights based on local training characteristics on the server side. Each client generates a feature vector consisting of training loss and accuracy, which is fed into a central policy network. This network outputs aggregation weights and is trained via a reinforcement process, using the global model’s validation performance as the reward. Extensive experiments on the UNSW-NB15 and NSL-KDD datasets, under both IID and non-IID conditions, show that FedRWA consistently outperforms state-of-the-art baselines in terms of classification metrics and communication efficiency. In addition, the learning curves demonstrate the superior convergence speed and training stability of FedRWA, validating its effectiveness as a robust solution for adaptive aggregation in heterogeneous FL environments.
Remote Sensing Object Detection (RSOD) presents unique challenges due to arbitrary object orientations, complex backgrounds, and data scarcity. Existing methods often rely on complex network architectures or auxiliary tasks for multi-task optimization, which increase model complexity and require additional data. In this work, we proposed an Adapter-based Few-shot Fine-Tuning framework for RSOD, termed AFFT. Specifically, to effectively capture global context and extract high-quality features, we adopted a pre-trained hierarchical Swin Transformer based on shifted windows attention mechanism as the backbone. This enables robust feature representation by efficiently modeling both local and global dependencies. Moreover, recognizing the limited availability of annotated remote sensing data, we designed a Group Equivariant Convolutions Multi-cognitive Visual Adapter (GCMONA) for few-shot fine-tuning. This approach significantly enhances model adaptability to new classes with minimal labeled samples, making it particularly well-suited for remote sensing scenarios. It is important that AFFT requires fine-tuning only 5
Real-world electronic health records (EHRs) are decentralized and typically exhibit incomplete information across both modalities and features. This heterogeneous missingness poses a major challenge for federated graph learning, which aims to collaboratively model distributed medical data while preserving privacy. Existing methods often assume uniform data completeness or apply imputation strategies, which struggle to maintain consistent predictive performance under non-uniform and severe data incompleteness. To address this issue, we propose MissPCL, a personalized federated graph contrastive learning framework tailored to heterogeneous client-level missingness. Each client estimates its missingness rate and constructs a bipartite patient-modality graph from observed features. Clients with low missingness participate in global training using contrastive learning on local graphs, while high-missingness clients are excluded from aggregation and instead perform personalized fine-tuning. Experiments on two real-world EHR datasets under diverse missingness scenarios demonstrate that MissPCL consistently outperforms state-of-the-art baselines in classification accuracy and robustness. Notably, our framework achieves stable performance across varying degrees of data incompleteness, showing its practical utility in realistic federated clinical settings. Code implementations and the supplementary materials are available at https://github.com/TutaResearch/MISSPCL .
This paper presents the use of Paired Hierarchical Variational Autoencoders (PHVAEs) for image-to-image translation and cross reconstruction. We explore a novel method for constructing and applying HVAEs for these types of tasks, and compare this model against existing GAN, transformer, and diffusion based methods, including pix2pix, StegoGAN, ResViT, and BBDM. The method utilizes deep, hierarchical VAEs with paired optimization to take advantage of the bidirectional latent space. The proposed model outperforms all alternative models in PSNR on all tested tasks, and attains consistently competitive SSIM, demonstrating its effectiveness and benefit for prediction-focused translation tasks such as cross reconstruction, while remaining parameter efficient.
Continuous video-based seizure detection remains clinically challenging owing to occlusions, environmental variations, and subtle seizure manifestations. We introduce a privacy-centric, non-invasive video-based seizure detection system that leverages dense surface normals to encode geometric features. This approach achieves superior generalization as the features remain invariant to patient appearance. This work explores the first application of surface normal analysis to seizure detection, demonstrating that geometry-based features not only preserve privacy but also outperform traditional pose-based methods. Through a rigorous evaluation on 821 clinical video clips from 7 patients, we systematically compare surface normals against pose estimation, semantic segmentation, and multi-modal fusion approaches under both patient-dependent (5-fold CV) and patient-independent (LOPO CV) validation protocols. Raw surface normals achieve 89.4% accuracy, significantly outperforming pose estimation ( 82.8% ) with a 26.8% relative improvement in F1-score. Critically, it maintains exceptional robustness in scenarios where semantic methods catastrophically fail, including multi-person interactions and severe occlusion, where its F1-score is robustly maintained at 0.833 while the pose-based F1-score drops to 0.0. Our proposed geometry-based, privacy-centric approach enables continuous monitoring in both clinical and home settings without compromising patient privacy.
With evolving technology, enabling a single operator to control multiple robots or a robotic swarm is becoming technically feasible. However, policy and regulatory frameworks do not yet address these scenarios, and we do not yet have a strong understanding of the impact of controlling large numbers of robots on human situation awareness. This paper addresses this gap through an empirical study that systematically degrades interface information to assess its impact on operator situation awareness and task performance in a human-swarm interaction scenario. Using a tablet-based interface, participants guide a robotic swarm through a simulated spreading fire task, and situation awareness and task performance are monitored. Results reveal that degrading information leads operators to focus on a single robot, often neglecting the rest of the swarm, a narrowing of attention that raises concerns for single-operator-multi-robot operations. Additionally, only certain information types were found to significantly affect situation awareness and performance, suggesting that some elements are critical while others are not. These findings underscore the need for regulatory support mechanisms such as secondary operators and call for flexible, task-specific policies that prioritise the most critical information elements in multi-robot operations.
Human-induced marine pollution has led to increasing underwater debris, threatening ecosystems. Although vision-based robots like AUVs and ROVs assist in seafloor monitoring, effective detection remains difficult due to poor visibility and degraded image quality in underwater environments. In this paper, we propose TrashTracer, a novel and lightweight framework for real-time underwater debris detection. To address visibility degradation, we introduce CG-Enhance, an online image enhancement method combining CLAHE and Gamma correction, which improves image quality without adding inference overhead. We further design an Efficient Reparameterized Backbone (ERB) utilizing the Diverse Branch Block (DBB) to enhance feature learning during training and simplify computation during inference. Additionally, we propose the Reparameterized Adaptive Spatial Fusion Head (RASF-Head) to adaptively integrate multi-scale features for robust detection. Extensive experiments on the TrashCan dataset demonstrate that our model achieves superior detection accuracy while maintaining low computational complexity and real-time performance, highlighting its strong potential for practical deployment in underwater debris monitoring applications.
Forward Search techniques such as Monte Carlo Tree Search (MCTS) take advantage of the ability to query a simulator and see the results of taking certain actions. While they are very effective in many board games they remain relatively ineffective in real-time environments. This is often due to the prohibitive expense of building a sufficiently deep search tree for effective play in these environments. We address this by introducing a new Evolutionary Algorithm which evolves macro-actions for a particular domain. A macro-action is a sequence of actions treated as a single action by MCTS; domain appropriate macro-actions increase search tree depth and avoid unproductive search. Our technique evolves effective macro-actions that significantly improve MCTS performance in multiple real-time domains.
In real-world applications, the demand for unmanned, autonomous platforms rises; yet these robot platforms are often deployed in highly dynamic environments and are operated by personnel with little to no expertise in formal planning, for whom plan validation must be simple and follow a human-in-the-loop (“four-eyes”) principle. To address this challenge, we propose a two-stage pipeline for automatic goal and constraint extraction for problem instance acquisition, where extracted goals complement the problem instance and constraints guide execution in open-world mobile robot mission planning. Our Large Language Model (LLM)-based goal and constraint extraction (GCE) module parses the operator-defined natural language mission instruction into a structured planner-independent goal and constraint representation while being robust to slightly misphrased mission instructions. In a second stage, a planner-specific projection layer maps these general goals and constraints to the syntax of any downstream planner. In a first validation step, on a fixed-domain benchmark of operator-authored mission instructions, our LLM-based extractor achieves higher exact-match goal identification and constraint coverage than a strong rule-based baseline, and we situate these results within current work. In a second step, we integrate the full pipeline into AUSPEX and showcase the planner-independent to planner-specific projection functionality based on different planners for a sample Search and Rescue (SAR) scenario with UAVs.
For ground-based optical telescopes, automated planning and control systems play a vital role in ensuring efficient acquisition of scientific data. These systems rely on All-Sky Cameras (ASC) to capture real-time sky conditions, including sky obscuration and lunar position. We address the task of assessing sky conditions under ambient nighttime illumination and in the presence of optical artifacts such as lens flares. Seeking to develop a solution based on a neural network model, we are faced with a lack of real nighttime cloud segmentation benchmark datasets. To address this, we propose an Augmentation and Transformation for Nighttime Cloud Segmentation in All-Sky Camera Images (AT-NCS) that converts existing daytime and dusk cloud segmentation datasets into a synthetic benchmark simulating realistic nighttime observing conditions. Our approach incorporates key challenges such as night sky glow, and optical artifacts (e.g., lens flare), thereby providing a training and evaluation framework for ASC-based observation systems. This work aims to enhance the reliability and value of autonomous observatory operations under diverse nighttime environments.
Multi-Agent Reinforcement Learning (MARL) has evolved from independent learning to sophisticated communication-based coordination systems like Adaptive Topology Communication (ATC). However, current communication-based approaches face limitations in domain knowledge integration, adaptation speed, and learning paradigm flexibility that hinder real-world deployment. This study proposes three synergistic MARL framework optimization techniques—Physics-Informed MARL, Meta-Learning MARL, and Hybrid Learning MARL—that extend beyond communication-based coordination to address these fundamental limitations. These methods integrate domain-specific knowledge, enable rapid adaptation to new scenarios, and provide flexible learning paradigms that dynamically balance centralized and decentralized approaches. Experimental evaluation in warehouse coordination tasks demonstrates statistically significant reward improvements of 2.4
Alignment methods for large language models frequently fail to preserve safety under distribution shifts owing to the coupling of task objectives and safety preferences. We introduce Guardrail Guided Policy Optimisation (GGPO), a technique that decouples task rewards from safety constraints by inferring them separately from positive and negative demonstrations. This approach builds upon inverse constrained reinforcement learning, where constraints are inferred from expert behaviour, but extends it with an adversarial mechanism inspired by adversarial inverse reinforcement learning to jointly model rewards and costs as discriminators for goal-oriented and unsafe actions. The functional forms for the reward and cost are selected based on empirical factors promoting training stability. Policy optimisation is then conducted through a Lagrangian relaxation that imposes the inferred constraints, with algorithmic stability derived from two-timescale stochastic approximation principles. In a series of text-based navigation environments featuring moving hazards, GGPO eliminates constraint violations entirely for a transformer-based agent while adapting to safety requirements after distribution shifts. These findings indicate that inferring explicit, decoupled cost functions within a theoretically grounded framework enhances safety in settings requiring generalisation and safety adaptation.
The scarcity of medical images poses major challenges for downstream tasks and AI deployment. While generative models like VAEs, GANs, and diffusion models offer potential solutions, they often fail to capture the fine-grained anatomical details essential in medical imaging, particularly for scoliosis. This highlights the need for a specialized model that preserves structural and lesion-specific features. We propose A-PriDiff, a diffusion-based generative model tailored for ultrasound scoliosis imaging. It adopts a two-phase framework: a pre-trained segmentation U-Net extracts anatomical features, which are refined by an Anatomical Prior Encoding Module (APEM) to guide a diffusion module in synthesizing high-fidelity images. Tested on a dataset of 1,170 raw and 109 segmented ultrasound images, A-PriDiff achieves superior performance (FID: 9.69, KID: 0.013), outperforming DDPM and GAN baselines. It also yields the lowest mean curve angle error (1.80 ± 0.90 ^∘ ). Ablation results confirm APEM’s critical role in enhancing anatomical consistency.
Person identification using behavioural biometrics has been an active area of research. Eye-blink patterns provide a unique set of characteristics that can identify individuals, as they are difficult to imitate. Current blink pattern-based biometric identification approaches use obtrusive head-mounted sensors to accurately capture eye blinks, which limits their daily use. In addition, they rely on one-dimensional temporal signals, such as eye-aspect ratio (EAR), or require fusion with additional biometric data. In contrast, we introduce novel two-dimensional spatio-temporal features, drawing from earlier successful gait recognition templates. In particular, our approach turns eye-blink sequences into discriminative spatiotemporal patterns. We construct and compare two types of “Time Images” to determine the most effective spatio-temporal template: (i) a Colour Outline Time Image (COTI), which uses a red-to-violet hue cycle to convey blink progression; and (ii) a Mean Intensity Time Image (MITI), which is an average intensity image of the periocular region. Our approach uses dynamic features extracted from the EAR and creates superimposed blink frame representations to capture the temporal evolution of eye contours. We used two different Convolutional Neural Network (CNN) architectures to classify these time images. Our method achieves strong results on two publicly available datasets that cover a range of covariates and demonstrates robust identification capability. This research serves as a baseline for future studies and demonstrates the validity of using eye-blink patterns as a standalone biometric feature for person identification.
Considerable progress has been made recently in the development of methodologies, software libraries, problem sets and landscape analysis towards improving the benchmarking of black box optimisation algorithms. The majority of this work focuses on algorithm evaluation at a large/broad scale, by evaluating algorithms on large sets of different problem instances, across multiple dimensionalities. Broad evaluations are useful in providing a more complete picture of the behaviour of an algorithm. However they necessarily involve the summarization of experimental results, which can mean insights at the individual problem level are lost. This paper presents a case study of exploratory, per-instance algorithm evaluation, using real-world-representative problem instances from training small neural networks. The results illustrate that insights can be gained from looking at individual problem instances that would not be seen in a broad summative analysis. The work also highlights some potential problems that can exist in summative comparisons if care is not taken at the instance level.
Modeling complex spatiotemporal systems remains a challenging task due to nonlinear interactions, chaotic behavior, and the frequent lack of complete mechanistic knowledge. Traditional approaches rely on approximate mathematical representations, while purely data-driven models often demand large volumes of high-quality data and struggle to generalize to unseen regimes. In this work, we explore hybrid deep learning approaches that integrate imperfect prior knowledge with observational data to model complex dynamics. We focus on two representative systems—the Lorenz system and the Kuramoto–Sivashinsky (KS) equation—to study how knowledge quality and data availability influence predictive performance. Two architectures are employed: Echo State Networks (ESNs) and Long Short-Term Memory (LSTM) networks, both adapted to incorporate knowledge-based predictions. Our empirical analysis shows that incorporating prior knowledge substantially improves the performance of ESNs, while LSTMs demonstrate strong baseline predictive capability with gains from knowledge integration and large data. The results highlight the importance of balancing knowledge and data in hybrid modeling and provide empirical evidence for designing robust deep learning frameworks for spatiotemporal chaotic systems.
In this paper we present generalizations of the k-means clustering algorithm, with guaranteed monotonic convergence, which do not require that data points be localized in a vector space. Instead, only a dissimilarity matrix between data points is needed. This generalization allows k-means-like clustering to be applied to data-sets that cannot be embedded exactly in an Euclidean space, such as road networks, and data-sets with categorical data. If the data points have localizations in an Euclidean space and the dissimilarity matrix is the matrix of mutual squared Euclidean distances between points, one of our algorithms perfectly mirrors the classical k-means. If the dissimilarity matrix does not correspond to an Euclidean scenario, our experimental results using both real-world and synthetic data sets show that our proposed generalizations of k-means generally lead to more compact clusters than k-medoids clustering.
