Loneliness research has mainly focused on an inside-out perspective, examining how lonely individuals process social information. Less is known about the outside-in perspective: how people judge whether someone else appears lonely and which visual cues guide that judgment. Because loneliness is not always outwardly apparent, understanding how judgments of perceived loneliness are formed may help explain when loneliness becomes socially visible, when it remains overlooked, and when it is misattributed. This study examines how elements of a visual scene influence perceived loneliness and attention toward the person being judged. Participants (N = 54) viewed 102 AI-generated photorealistic images (six everyday scenes × 17 versions) while their eye movements were recorded. In each scene, a baseline image showed the foreground person alone, with a neutral expression in a neutral setting. Manipulated versions of each baseline scene changed the person’s facial expression, the weather, the presence of objects, the background density, or the surrounding social situation. In the social scenes, further versions changed whether the foreground person remained neutral or reacted to the social situation. Each image was presented for 3 seconds, after which participants rated how lonely the foreground person appeared. Perceived loneliness was quantified as the rating shift from the matched baseline (absolute shift for magnitude comparisons); visual attention was quantified as dwell time percentage (DTP) on the foreground person’s face. Among scenes depicting a single person, the facial expression manipulation produced the largest mean absolute rating shift and the highest face DTP of the four manipulation types. Changing the weather also produced relatively large mean absolute rating shifts, though these were not consistent across the weather variations, and reduced face DTP relative to baseline. The presence of objects and cluttered backgrounds mainly reduced face DTP without corresponding changes in perceived loneliness. In scenes depicting multiple people, the foreground person’s reaction to the surrounding social situation was associated with a stronger influence on both loneliness judgments and face DTP than the type of social situation itself: reaction to the social situation produced larger mean absolute rating shifts and lower face DTP than remaining neutral. Overall, manipulations that produced the largest changes in loneliness judgments did not consistently produce the highest face DTP. This suggests that perceived loneliness was influenced by facial, environmental, and social information rather than attention on the face alone. However, how observers combined these cues was not directly examined here. ... Read more

Large Language Models (LLMs) promise intuitive robot control through natural language, yet the gap between vague human intent and safe physical execution remains significant. This thesis investigates how the distribution of planning responsibility relates to the reasoning architecture of the model. In a study involving 226 participants who were new users in LLM-based robot prompting, a non-reasoning model (Gemini 2.0 Flash-Lite) and a reasoning model (Gemini 2.5 Pro) were compared on a baseline navigation task, followed by an evaluation of the reasoning model across tasks of increasing logical complexity. Results indicate a clear divergence in safety profiles: non-reasoning models showed more collision-prone goal-seeking behavior, whereas reasoning models demonstrated stricter adherence to the safety constraints in the system prompt, preferring to refuse a request rather than generating an unsafe plan when the task exceeded the model’s capabilities. However, even reasoning models showed declining performance in high-complexity tasks, which hyperparameter tuning (temperature/tokens) did not resolve. Analysis of user interaction reveals that effective prompting is less about linguistic precision and more about "distributed cognition": while models can autonomously plan simple tasks, complex scenarios require the human to reclaim the task planning effort and provide low-level guidance to reduce the solution space. These findings suggest that safe language-driven robotics depends on a dynamic partnership where the distribution of task planning effort shifts based on task difficulty and the capabilities of the specific LLM. ... Read more

With a growing elderly population, shoulder injuries are becoming more common, and part of the recovery plan is to go to physiotherapy. However, to alleviate the demand for physiotherapists, robots could help with shoulder rehabilitation. To do this safely and enjoyably, the robot will need to prevent re-injury caused by fatigue while keeping the patient interested and motivated to continue with their therapy. In this study, a method for managing fatigue of the two most commonly injured shoulder muscles, the supraspinatus and infraspinatus, in a game is proposed and tested. To validate the developed method, a human factors experiment was conducted. The fatigue-adaptive game was compared to a baseline in which participants controlled fatigue themselves. The participants played three cases for each version of the game. To minimize the risk of over-fatiguing during physiotherapy and not crossing the line of being too fatigued. Therefore, we measured the overshoot of fatigue in both versions of the game. The mean of the overshoot is compared with a Welch's t-test with Bonferroni correction for each fatigue case. The results show a significant difference for some of the fatigue cases, where the controller is either significantly better or there is no significant difference in the overshoot. The fatigue-adaptive game shows consistency across the cases, whereas the baseline does not. Therefore, the fatigue-adaptive game can compete with a person in managing fatigue, while being easier to learn and automatically identifying and removing risky shoulder positions where fatigue changes rapidly. The fatigue-adaptive game can also be played with an industrial robot arm and still demonstrates the capability to manage the fatigue of the two most commonly injured muscles. ... Read more

This paper presents a method which is capable of creating an object centered world description based upon consecutive measurements about an unknown number of static objects using Bayesian inference. The objects are represented by a two dimensional position, with the aim of adding more attributes in future works. This objective is reformulated into a clustering problem which is then solved using a structure learning method. It is implemented using RxInfer which uses the message passing algorithm in combination with factor graphs to perform Bayesian inference. The results indicate a promising performance of the structure learning model, but also show signs that the object representation has been over simplified. The future works section provides guidance on how the model complexity can be increased by adding additional attributes in order to improve performance. ... Read more

Background Preterm birth is associated with an increased risk for neurodevelopmental impairments, requiring brain monitoring using amplitude-integrated electroencephalography (aEEG). While tools exist to detect severe brain dysfunction, methods for mild dysfunction—such as the Burdjalov scoring system or expert identification of sleep-wake cycles—are limited by subjectivity and require expert training. Existing automated sleep-staging models are typically trained on term neonates using polysomnography, a resource-intensive method not widely feasible in neonatal intensive care units (NICUs) for preterm neonates, where simplified aEEG with fewer electrodes is more commonly used.
Methods aEEG recordings from neurologically healthy neonates between 32 and 42 weeks postmenstrual age (PMA) were annotated for quiet sleep (QS) and non-quiet sleep (NQS) by a single expert clinician.
Results Five classifiers were trained to classify QS and NQS. A k-nearest neighbors model achieved a mean Cohen’s Kappa of 0.71± 0.12 in preterm infants, decreasing to 0.48 ± 0.21 in term infants. Features from QS segments were strongly correlated with PMA, enabling a PMA predictor model to achieve an average error of 0.88 weeks.
Conclusion Although performance on QS/NQS classification was strong for neonates between 33 and 37 weeks PMA, generalization across the full 32–42 week range remains challenging. Nevertheless, the low average error of the PMA predictor highlights its potential as a tool for detecting mild neuromaturation delays.
... Read more

Safe quadrupedal locomotion control with reinforcement learning (RL) has attracted increasing attention in recent years, where existing approaches can be broadly categorized into recovery RL, distributional RL, and constrained RL. However, recovery RL cannot provide predictive safety guarantees; distributional RL lacks passive safe performance; and constrained RL-while capable of both safety-often restricts exploration. To address these limitations, we propose \textbf{UPPS-RL}, a unified framework that integrates predictive and passive safety into quadrupedal locomotion control through three main components: a risk-aware task-level policy, a self-supervised risk network, and a risk-triggered recovery policy, forming a hierarchical control architecture that embeds unified safety without imposing explicit exploration constraints. Extensive simulations across composite scenarios, including steps, pit, slope, and rough plane terrains, demonstrate that UPPS-RL significantly suppresses catastrophic failures while maintaining a favorable trade-off between robustness and efficiency. ... Read more

Platooning has become a useful area for better transportation efficiency on highway driving. As Cooperative and Automated Vehicles continue to evolve and integrate , it is important to have insights into their implications, emphasizing the need for rigorous real-world assessments. In general, platoon formation is monitored by Cooperative Adaptive Cruise Control (CACC), which uses real-time vehicle-to-vehicle (V2V) communication to exchange vehicle status information, improving the control reaction as platoon members adjust to their surroundings. Automated systems can normally drive vehicles to perform planned behaviors based on the pre-setting by humans, but if the platoon encounters disturbances, the extent to which the automated system can still follow human intentions is still unknown. This research uses field operational test (FOT) data from the CACC platoon on an arterial corridor to assess the platoon's performance when disrupted during the test. This research applies the concept of meaningful human control (MHC) with focus on tracking condition. Additionally, this study will focus on human 'reasons', both distal and proximal. An evaluation framework for platoons is created by categorizing 'Tracking' into three main metrics: comfort, safety, and local stability. Furthermore, this study demonstrates that disturbance has variable degrees of detrimental impact on the platoon's tracking state, and that these effects may be recovered when the disturbance has concluded; however, different disturbance situations indicate different recoveries. The evaluation methodology of this paper provides insight into the tracking performance of CAVs, which can help road authorities build infrastructure for their wider deployment of CAVs. Last but not least, this study may provide guidance to automation technology organizations and automobile manufacturers on how to develop vehicles so that they follow human reasons more closely. ... Read more

Stroke survivors often struggle with gait asymmetry post-therapy. Researchers are exploring Virtual Reality (VR) to address this problem with the help of visual feedback and virtual avatars. VR also makes repetitive tasks more enjoyable, improving patient compliance and outcomes. Motor adaptation, essential in rehabilitation, involves adjusting movements to new conditions. Studies have used forms of motor adaptation with visual feedback to distort participants' gait symmetry, making people walk asymmetrically. This approach is called implicit Visual Feedback Distortion (VFD), where visual feedback is manipulated without the user's awareness. This thesis explores using implicit VFD with avatars in immersive VR to address gait asymmetry. An experiment with 11 healthy participants tested implicit VFD by gradually increasing the step length of the avatar's right leg. Contrary to previous screen-based VFD studies, results showed no significant effect on gait symmetry. Additionally, no correlations were found between step symmetry and psychological states (presence, embodiment, motivation). We hypothesize that the high distortion level on the right foot, suboptimal virtual environment design, and reported neck pain contributed to these findings. Future research should explore different VFD designs, or look at multiple groups to see what conditions lead to more gait asymmetry during adaptation. ... Read more

As autonomous vehicle (AV) technology progresses, the necessity for a comprehensive understanding of interactions between AVs and human-driven vehicles (HVs) becomes paramount, particularly in critical manoeuvres such as merging. Mastering merging interactions is essential for enhancing road safety. Existing research in this field focuses on how the AV performs the merging manoeuvre but often fails to assess how they influence these interactions. By drawing inspiration from Human-Robot Interaction and Human Aware Navigation, this study aims to bridge this gap by examining how these interactions influence driver workload, measured through fixations duration, perceived safety and drivers’ subjective perception during merging scenarios. We employed a Virtual Reality environment to simulate realistic driving conditions and measure driver responses. We conducted an experiment where participants engaged in merging manoeuvres with each other and, subsequently and without being informed, with the AV described in “Planning for cars that coordinate with people” [1]. This approach allowed for an unbiased assessment of natural driver reactions to AV behaviours. Our findings reveal significant increases in driver workload and decreases in perceived safety during HV-AV interactions, compared to HV-HV interactions. These results suggest that current AV algorithms may not fully account for the complexity of human-AV interactions, highlighting a need for interaction evaluation in the AV development. Participants’ subjective feedback indicates a recognition of and negative reaction to AV driving behaviours, emphasizing the importance of designing AVs that are both efficient and intuitive for human drivers. The study’s implications suggest improving AV controllers’ evaluations by including their interactions with human drivers. By integrating interaction evaluation, AV technologies can achieve smoother and more successful integration into existing road systems, enhancing predictability and driver acceptance. This study marks a step towards understanding the interactions between AVs and HVs, offering insights that could steer future research and development in autonomous driving technologies.
... Read more

Robots can aid in post-stroke motor function recovery and motor learning through the use of haptic feedback during collaborative training. A clear objective in robotic-assisted motor learning is to adapt the haptic feedback to individual users, but personal characteristics are not yet considered in this adaptation.

We investigated the suitability of a haptic guidance feedback strategy, based on participants’ locus of control character trait, compared to training without haptic guidance. For this purpose, a motor learning experiment was conducted on 42 healthy participants, where the internal dynamics of a pendulum had to be learned in order to hit upcoming targets. For two groups, training either with or without haptic guidance, we assessed motor learning and its generalization to similar tasks through target hitting performance, as well as behavior during training and perceived user experience.

Evidence was found of a relatively better performance improvement in both training and long-term (generalization of) motor learning for participants with a more internal compared to external locus of control. Lower observed interaction force during training and increasingly better performance throughout training
in these participants may have caused these motor learning differences. More positive user experience in these individuals through a higher perceived control over the pendulum and lower perceived frustration with haptic guidance may have also contributed.

Combined, this suggests an intrinsically better compatibility
with haptic guidance for people with a more internal rather than
external locus of control, for motor learning, during training and
in user experience ... Read more

Road traffic safety is a pressing global concern, with millions of yearly fatalities and injuries. This study aims to address the detection of abnormal driving behaviour. Traditional supervised approaches face limitations due to the need for labelled abnormal driving data. To overcome this challenge, semi-supervised machine learning models are explored and developed in this research.
Machine learning is utilized for abnormal driving behaviour detection because it offers a data-driven approach that adapts to different scenarios and captures subtle patterns. Furthermore, its scalability allows for efficient analysis of large datasets, leading to accurate identification of abnormal driving behaviour and valuable insights for enhancing road safety measures. Most existing machine learning (ML) based abnormal driving detectors rely on (fully) supervised ML methods, which require substantial labelled data. However, in the real world, labels are only sometimes available, and labelling large amounts of data is tedious. Thus, there is a need to employ unsupervised or semi-supervised methods to make the detection process more feasible and efficient. Luckily, it is possible with the advent of deep neural networks, especially autoencoder-based ones. This thesis develops and compares three ML methods: supervised (e.g. XGBoost and Random Forest), unsupervised ML (e.g. Isolation Forest and Robust Covariance), and semi-supervised ML (Hierarchical Extreme Learning Machines). Comparison results show that the semi-supervised deep learning model outperforms unsupervised methods exhibiting higher prediction accuracy and delivering acceptable results compared to the fully supervised models.
Moreover, previous ML-based approaches predominantly utilize basic car motion features (such as velocity and acceleration) to label and predict abnormal driving behaviours. In contrast, this thesis introduces Surrogate Measures of Safety (SMOS) as features for ML models to identify abnormal driving behaviour.
The results indicate that the supervised model performs best under the same conditions. However, relying on a large amount of labelled data in supervised models can pose challenges in real-life scenarios or when dealing with massive datasets. The study highlights the significance of Surrogate Measures of Safety (SMOS) and demonstrates the potential of HELM in effectively identifying abnormal driving behaviour. The introduction of SMOS significantly improves the performance of both unsupervised and semi-supervised models. The unsupervised model shows the most substantial improvement, increasing accuracy from less than 50% to over 90%.
While the Isolation Forest and Robust Covariance models fail to detect abnormal driving behaviour without including SMOS, the semi-supervised HELM model exhibits promising results even without SMOS. However, further research is necessary to address limitations and enhance the findings. While valuable, the current dataset used in this study may only encompass some types of abnormal driving behaviour. Future research should incorporate a more diverse dataset that covers a broader range of abnormal driving behaviours. The analysis should include multiple SMOS features, such as Post Encroachment Time (PET), to comprehensively understand abnormal driving behaviour and improve safety measures. ... Read more

Severe muscle weakness is a symptom appearing in certain neuromuscular diseases (NMDs), such as Duchenne Muscular Dystrophy (DMD), affecting people's daily lives by reducing functionality, decreasing independence, and reducing the ability to perform essential daily activities. This patient group might benefit from using active-assistive devices by having the potential to provide precise support torque counterbalancing the passive forces acting on the arm, the movement intention of the user, and external forces exerted by lifted objects. However, the determination of support to counteract the weight of lifted objects is an ongoing challenge. This research aims to improve the understanding of external forces by using data classification algorithms to distinguish between different lifted weights in a human experiment. Fourteen healthy individuals participated in this experiment, lifting weights ranging from 0 - 1000 grams while an active-assistive device compensated for the passive torques acting on the arm. Data was collected using various sensors: a force sensor, an Inertial Measurement Unit (IMU), a joint encoder, and surface Electromyography (sEMG) electrodes. Subsequently, this data was processed and fed into a K Nearest Neighbour (KNN) classifier and a Support Vector Machine (SVM) classifier to determine the lifted weights during human elbow flexion and extension. The classifier showing the highest performance achieved an accuracy of 39.70% on the test dataset, indicating several misclassifications. However, a recall percentage of 76.95% for the 1000-gram class within the multi-class classification demonstrates the capability to distinguish larger weights. While demonstrating potential in weight discrimination, especially for larger weights, improvements in the compensation strategy, arm support alignment, and experimental design are crucial. Future research on the impact of picking and placing objects, the influence of muscle weakness, and the application of alternative data classification algorithms are essential to further enhance understanding of the interaction with objects and result in more accurate predictions. ... Read more

Background: The procedure to fit a prosthetic socket to a patient, which can assure the patient’s comfort during activities of daily living, is labour intensive. Such a lengthy procedure could benefit from an automated and more efficient data-driven method capable of automatically tracking the relative movement between the patient’s tibia and the prosthetic socket. To investigate such a method, we acquired in-socket bone displacement data during the physical activities of the prosthetic user. Manually tracking the location of the tibia from, e.g., B-mode (imaging) ultrasound (US) sequences might be a solution, but this is time-consuming, and the interpretation of the sequences is highly operator dependent. Therefore, an automated and efficient method to assess socket fit in US sequences is needed.
Methods: We used an existing 3D U-Net with a long short-term memory module (LSTM) and compared its ability to track a landmark location point on the tibia in US recordings by comparing the displacement and similarity in shape with data obtained from a semi-automatic single-point tracker. To evaluate the performance of the developed automated workflow, we obtained experimental data from three participants who performed three repetitive stepping tasks with their prosthetic leg in a sideways, forward, and backward motion. Three deep learning models were trained with a varying hold-out method (66% training data, 34 % test data) to test the ability to track a landmark location on the tibia in unseen data from one participant. To find the similarity of the deep learning models compared to a semi-automated single point tracker, the normalised root mean squared error (NRMSE) was calculated. We also evaluated the normalised maximum cross-correlation (NMCC) to account for the maximum similarity in displacement trajectory when a delay occurred between the true trajectory and that from the automated model. We analysed the repeatability of each step task per participant with the standard deviation from the mean tibia’s landmark location trajectories.
Results: Due to the delay between the semi-automated single-point tracker and the DL model, the NRMSEs ranged between 27% and 90%. The similarity threshold (0,95) was reached for five trajectories of the tracked point on the tibia in the anterior-posterior direction, with a delay between 1,5% and 8,5% of the step duration. The similarity in the anterior-posterior direction of the tibia’s landmark location trajectory was higher than that in the lateral-medial direction. The SD for all participants was around 1 mm but varied proportionally to the amount of movement observed per participant. The SD of the DL models was similar to that of the semi-automated single-point tracker.
Conclusion: We conclude that a DL model from a 3D U-Net with an LSTM module has the potential to assist prosthetists and researchers in tracking in-socket tibial bone movement in the anterior-posterior direction. ... Read more

Neurological disorders in the nervous and neuromuscular systems affect approximately 260 million people annually and among these 255 million would benefit from rehabilitation [4]. Patients with neurological disorders usually require multi-dimensional rehabilitation, involving physical, cognitive, psychological, and medical help. Children with trunk control problems arising due to some of these neurological disorders also require such multi-dimensional rehabilitation. A major part of this is administered to the patient through the activities of a physiotherapist in the clinical context. But the limited number of physiotherapists result in exercises often being prescribed for patients as in-home rehabilitation. During in-home rehabilitation, the patient and the primary care-giver may not be able to comply with the prescription without feedback from a physiotherapist.
To address this challenge, this paper proposes an automated method for assessing movement quality of children during trunk control rehabilitation exercises. We adopted a Human-centered AI approach to the development of our system. We identified the needs of physiotherapists for assessing patient’s functional abilities through semi-structured interviews with six physiotherapists. As a result, we co-designed and developed an artificially Intelligent decision support system that automatically assesses the quality of motion. We created a trunk-control rehabilitation exercise movement dataset based on a protocol co-designed by the authors and the physiotherapists. The data was collected
from 15 typically developing children (mean age 7 years, range 4–10 years) using a ZED-mini stereo-camera and the quality scores as ground-truth were obtained from a physiotherapist. The exercises involved reaching targets kept on a table in front
while being seated away from the table on a stool. We investigated the performance of Random Convolutional Kernel transform and XCM, two state-of-the-art multivariate
time-series classification algorithms on this dataset and achieved a quality prediction f1-score of 65% on the test dataset and similar promising results on the detection of compensatory movements in the exercise motion. In addition, to increase the trust-worthiness of our AI solution, we have provided explanations on the predictions of the black-box algorithms, which can aid the users of the system to understand the causal relationships between the input and output to the AI algorithm. ... Read more

Current commercial Driver Steering Assistance Systems (DSAS) focus on path-tracking performance without taking into account driver intentions. Improved driver-automation interaction can be achieved by sharing vehicle lateral control through torques. Furthermore, integrating a driver steering-torque model allows to better match driver intentions. In this research, an existing driver model is adapted and parametrized for estimating driver steering-torque. Driver behaviour is modelled by learning the parameters of a Hidden Markov Model (HMM) and estimation is performed with Gaussian Mixture Regression (GMR). A parameter selection framework enables to select model hyper-parameters objectively. First, feature relevance is determined with an extensive feature
selection step. Thereafter, an iterative overfitting criteria is employed to select the number of hidden states. Final model behaviour is determined by adjusting the metric weights of a linear cost-function with the aim to trade-off estimation accuracy and smoothness. Naturalistic driver steering-torque data from seven participants was gathered in a fixed-base driving simulator at Toyota Motor
Europe for the training, evaluation, and testing of the proposed model. The results demonstrate that a 92% model accuracy can be achieved while the estimated steering-torque is 37% smoother and requires 90% less data compared to a baseline model. ... Read more