To Örebro University

Human driver errors remain a central road-safety concern, but existing error and violation taxonomies were developed mainly for conventional traffic. In mixed traffic, human drivers must also interpret and respond to autonomous vehicles (AVs), whose behavior, intent, recognizability, and social role may differ from those of human-driven vehicles. This paper presents a case-level evidence synthesis and develops a taxonomy of human driver errors and unsafe adaptations in mixed traffic, focusing on drivers of manually driven vehicles who interact with AVs as surrounding road users. The proposed taxonomy links AV-specific interaction factors, including recognizability, driving style, predictability, trust, social evaluation, infrastructure, penetration rate, and repeated exposure, to observable unsafe behavior and to the information-processing stages at which driver errors are inferred to occur. The synthesis suggests that many mixed-traffic errors are concentrated in the projection and decision stages, with two recurring pathways: safety-margin reduction caused by poor anticipation of AV behavior or excessive reliance on AV capability, and exploitation caused by intentional priority-taking or social devaluation of AVs. The taxonomy provides a structured basis for mixed-traffic safety analysis and assessment. 

Human-induced errors-such as slips, lapses and mistakes-are natural and often unavoidable, posing significant safety risks in teleoperation, particularly in high-risk, dynamic environments like underwater operations. While previous work has emphasised system-level enhancements, the proactive mitigation of human-induced errors through empirical evaluation remains underexplored. This work presents the first empirical investigation of Intelligent Disobedience (ID) as a collaborative strategy for mitigating human errors in teleoperation. A semi-controlled, game-based Wizard-of-Oz study with 40 participants performing an underwater navigation task was conducted. Two ID strategies were evaluated: one with automatic robot mitigation and another with user-mediated robot mitigation following the disobedient intervention. Using a mixed-methods approach that triangulated quantitative performance metrics with qualitative insights from semi-structured interviews, the study examined the effects of these strategies on operator performance, acceptance, and error-mitigation outcomes across different error types. The results revealed no detectable performance degradation under ID and emphasised that human error types differ in their characteristics, necessitating the adaptation of ID intervention strategies. These findings underscore the importance of context-adaptive approaches when integrating ID into collaborative teleoperation systems.

We present our first stage results from deploying an LLM-augmented visualization software in a classroom setting to engage primary school children with Earth-related datasets. Motivated by the growing interest in conversational AI as a means to support inquiry-based learning, we investigate children's expectations, engagement, and evaluation of a spoken LLM interface with a shared, immersive visualization system in a formal educational context. Our system integrates a speech-capable large language model with an interactive spherical display. It enables children to ask natural language questions and receive coordinated verbal explanations and visual responses through the LLM-augmented visualization updating in real-time based on spoken queries. We report on a classroom study with Swedish children aged 9-10, combining structured observation and small-group discussions to capture expectations prior to interaction, interaction patterns during facilitated sessions, and children's reflections on their encounter afterward. Our results provide empirical insights into children's initial encounters with an LLM-enabled visualization platform within a classroom setting and their expectations, interactions, and evaluations of the system. These findings inform the technology's potential for educational use and highlight important directions for future research.

A central challenge in building continually improving agents is that training environments are typically static or manually constructed. This restricts continual learning and generalization beyond the training distribution. We address this with COvolve, a co-evolutionary framework that leverages large language models (LLMs) to generate both environments and agent policies, expressed as executable Python code. We model the interaction between environment and policy designers as a two-player zero-sum game, ensuring adversarial co-evolution in which environments expose policy weaknesses and policies adapt in response. This process induces an automated curriculum in which environments and policies co-evolve toward increasing complexity. To guarantee robustness and prevent forgetting as the curriculum progresses, we compute the mixed-strategy Nash equilibrium (MSNE) of the zero-sum game, thereby yielding a meta-policy. This MSNE meta-policy ensures that the agent does not forget to solve previously seen environments while learning to solve previously unseen ones. Experiments in urban driving, symbolic maze-solving, and geometric navigation showcase that COvolve produces progressively more complex environments. Our results demonstrate the potential of LLM-driven co-evolution to achieve open-ended learning without predefined task distributions or manual intervention.

Reinforcement learning algorithms are defined by their learning update rules, which are typically hand-designed and fixed. We present an evolutionary framework for discovering reinforcement learning algorithms by searching directly over executable update rules that implement complete training procedures. The approach builds upon REvolve, an evolutionary system that uses large language models as generative variation operators, and extends it from reward-function discovery to algorithm discovery. To promote the emergence of nonstandard learning rules, the search excludes canonical mechanisms such as actor-critic structures, temporal-difference losses, and value bootstrapping. Because reinforcement learning algorithms are highly sensitive to internal scalar parameters, we introduce a post-evolution refinement stage in which a large language model proposes feasible hyperparameter ranges for each evolved update rule. Evaluated end-to-end by full training runs on multiple Gymnasium benchmarks, the discovered algorithms achieve competitive performance relative to established baselines, including SAC, PPO, DQN, and A2C.

Single-image super-resolution aims to recover high-frequency detail from a single low-resolution image, but practical applications often require balancing distortion against perceptual quality. Existing methods typically produce a single fixed reconstruction and offer limited test-time control over this trade-off. This paper presents DR-SCAN, a dual-branch deep residual network for single-image super-resolution in which, during test-time inference, weights can be assigned to either of the branches to dynamically steer their respective contributions to the reconstructed output. An interactive interface enables users to re-weight the shallow and deep branches at inference or run a one-click LPIPS search, to navigate the distortion–perception trade-off without retraining the model. Ablation experiments confirm that both the second branch and the channel–spatial attention that is used within the residual blocks are essential for the network for better reconstruction, while the interactive tuning routine demonstrates the practical value of post-hoc branch fusion.

The purpose of this perspective is twofold: first, to highlight artificial intelligence (AI) implementation and patient safety, and second, a broader policy perspective on European AI sovereignty. European AI sovereignty and the access to trustworthy AI systems are highly relevant for orthopedic applications including clinical infrastructure, data governance, medical device regulation, and implementation in clinical workflows. AI has enabled disruptive technological advancements in medicine. AI-assisted navigation, robotics, and image-based diagnostics facilitate the next level of precision-based medicine. Medical documentation and coding based on AI have created hopes of reduced workload and improved working conditions in clinical production.

Fracture detection using AI has demonstrated diagnostic benefits among less experienced human image readers in high-volume emergency settings. AI-based computer tomography-radiostereometric analysis (CT-RSA) has demonstrated accuracy comparable to model-based RSA, apart from providing accurate and predictable preoperative implant planning. Large language models (LLMs) have been shown to generate orthopedic discharge documents faster than humans, maintaining comparable quality, and perform on par with or above junior residents in orthopedic exams, highlighting its potential as a learning and educational tool.

Estimating a target's six-degree-of-freedom (6-DoF) motion in underwater proximity operations is difficult because the chaser lacks target-side proprioception and the available relative observations are sparse, noisy, and often partial [e.g., ultra-short baseline (USBL) positions]. Without a motion prior, factor-graph maximum a posteriori estimation is underconstrained: consecutive target states are weakly linked and orientation can drift. We propose a generalized constant-twist motion prior defined in the tangent space of Lie groups that enforces temporally consistent trajectories across all DoFs; in SE(3), it couples translation and rotation in the body frame. We present a ternary factor and derive its closed-form Jacobians based on standard Lie group operations, enabling drop-in use for trajectories on arbitrary Lie groups. We evaluate two deployment modes: Mode A, an SE(3)-only representation that regularizes orientation even when only position is measured, and Mode B, a mode that uses boundary factors to switch the target representation between SE(3) and R-3 while applying the same generalized constant-twist prior across representation changes. Offline validation on a real-world dynamic docking scenario data set shows consistent ego-target trajectory estimation through USBL-only and optical relative measurement segments with improved relative tracking accuracy compared with the raw target measurements. Because the construction relies on standard Lie group primitives, it is portable across state manifolds and sensing modalities.

In a rapidly transforming technological landscape, higher education institutions struggle to meet the demands of industry in providing timely and appropriate upskilling and reskilling opportunities for the workforce. Under the umbrella of "advanced digitization", a term prominent in European discourse, covering a wide range of new digital technologies within the Industry 4.0 and Industry 5.0 paradigms in manufacturing, ICT and engineering, we use scoping review methods to explore the unique challenges and interventions in this area, highlighting potentials for further research. Our scoping review examines 81 peer reviewed Scopus-indexed papers, outlining the current global research on the challenges and interventions in upskilling the workforce for advanced digitization. Our results show that research in this area is most prevalent in Europe, North America, and Asia and is most frequently done within the manufacturing and ICT sectors. Due to the fast pace of advanced digitization, close collaboration between education providers and industry is required to develop relevant and appropriate upskilling opportunities. We underline the importance of developing strategic management and change management knowledge alongside technology skills. We also highlight the importance of inter- and trans-disciplinary knowledge development. While determining skill needs has attracted substantial research focus, businesses may struggle with developing strategic upskilling plans despite being aware of their skills needs. We conclude that interventions that foster situated engagement and hands-on experience are vital for the future of upskilling.

As district heating networks evolve to meet climate-neutral objectives, optimizing their control under heterogeneous building characteristics and dynamic environmental conditions remains a significant challenge. Traditional control strategies often lack the adaptability necessary to account for building-specific dynamics and to ensure real-time adherence to operational safety constraints. In this work, we present an integrated machine learning-based framework that combines an adaptive context-aware transformer model with deep reinforcement learning to address these limitations. The proposed approach introduces an adaptive context-aware transformer as a predictive environment within a Deep Q-Network (DQN) framework, enabling data-driven, building-specific control of district heating systems. Utilizing real-world data from 148 residential buildings across Sweden and Finland, the model incorporates contextual embeddings and temporal features to predict indoor temperature trajectories with high accuracy, achieving root mean square error values between 0.18-0.24 degrees C for Swedish buildings and 0.26-0.32 degrees C for Finnish buildings. The DQN agent leverages these predictions to optimize heating control while ensuring compliance with operational safety limits and preserving occupant comfort. Experimental results demonstrate significant energy savings, with mid-rise buildings achieving up to 14.85% reduction in energy consumption, and peak seasonal savings exceeding 20% during spring months. This integrated approach illustrates the potential for substantial energy optimization and reliable indoor climate management in future district heating networks.