Bibsonomy APA

Motor learning is an important property for em-bodied agents and a wide variety of approaches has been developed for humanoid robots. One strategy is to study the human brain and replicate the structure and functionality of the brain in neuro-computational models. These models are built at different levels of abstraction to perform a variety of tasks. The presented work integrates a recent system-level neuro-computational model incorporating the motor cortex basal ganglia loops, the cerebellum and central pattern generators with the iCub robot. The iCub is a humanoid robot explicitly developed for cognitive research and well suited for manipulation and other motor tasks due to its high degrees of freedom. To test the performance of the model two experiments are conducted. In the first step a reaching task is carried out. Here, the task is to perform movements from a fixed initial position to different goal locations in the 3D-space. The results are compared to the prior performance of the model to verify the model and code adjustments and the learning abilities. In the second setup a simple motor adaptation task is realized, where an external shift perturbation is introduced to the arm motion, and the adaptation abilities of the model are investigated.

Gesture-based robot control offers intuitive interaction between humans and robots, with applications ranging from industrial automation to assistive robotics. However, existing solutions face challenges in achieving real-time requirements while ensuring accurate gesture recognition. This paper presents a new edge computing-based approach for real-time control of robots using Electrical Impedance Tomography (EIT) measurements to classify hand gesture numbers in American Sign Language (ASL). Existing solutions for gesture recognition struggle to achieve real-time performance while maintaining accuracy and energy efficiency. This challenge becomes higher in the case of EIT because of its relative complexity. We focus therefore on leveraging the capabilities of the edge device to implement effectively the Convolutional Neural Network (CNN) acceleration. The proposed solution combines hardware-aware optimization techniques to achieve fast and accurate gesture recognition by enabling rapid inference while minimizing energy consumption on a low-power resource-constrained device with Tiny Machine Learning (TinyML) capabilities. The lightweight CNN model required only 10.2 s to train using the Keras library of TensorFlow and achieved an accuracy of 89.37% for 10 sign language classes, with only 66 μs taken to run inference on the hardware-accelerated microcontroller-based device.

Gaze is an important and potent social cue to direct others’ attention towards specific locations. However, in many situations, directional symbols, like arrows, fulfill a similar purpose. Motivated by the overarching question how artificial systems can effectively communicate directional information, we conducted two cueing experiments. In both experiments, participants were asked to identify peripheral targets appearing on the screen and respond to them as quickly as possible by a button press. Prior to the appearance of the target, a cue was presented in the center of the screen. In Experiment 1, cues were either faces or arrows that gazed or pointed in one direction, but were non-predictive of the target location. Consistent with earlier studies, we found a reaction time benefit for the side the arrow or the gaze was directed to. Extending beyond earlier research, we found that this effect was indistinguishable between the vertical and the horizontal axis and between faces and arrows. In Experiment 2, we used 100% “counter-predictive” cues; that is, the target always occurred on the side opposite to the direction of gaze or arrow. With cues without inherent directional meaning (color), we controlled for general learning effects. Despite the close quantitative match between non-predictive gaze and non-predictive arrow cues observed in Experiment 1, the reaction-time benefit for counter-predictive arrows over neutral cues is more robust than the corresponding benefit for counter-predictive gaze. This suggests that–if matched for efficacy towards their inherent direction–gaze cues are harder to override or reinterpret than arrows. This difference can be of practical relevance, for example, when designing cues in the context of human-machine interaction.

It is often assumed that rendering an alert signal more salient yields faster responses to this alert. Yet, there might be a trade-off between attracting attention and distracting from task execution. Here we tested this in four behavioral experiments with eye-tracking using an abstract alert-signal paradigm. Participants performed a visual discrimination task (primary task) while occasional alert signals occurred in the visual periphery accompanied by a congruently lateralized tone. Participants had to respond to the alert before proceeding with the primary task. When visual salience (contrast) or auditory salience (tone intensity) of the alert were increased, participants directed their gaze to the alert more quickly. This confirms that more salient alerts attract attention more efficiently. Increasing auditory salience yielded quicker responses for the alert and primary tasks, apparently confirming faster responses altogether. However, increasing visual salience did not yield similar benefits: instead, it increased the time between fixating the alert and responding, as high-salience alerts interfered with alert-task execution. Such task interference by high-salience alert-signals counteracts their more efficient attentional guidance. The design of alert signals must be adapted to a ``sweet spot'' that optimizes this stimulus-dependent trade-off between maximally rapid attentional orienting and minimal task interference.

Several studies have reported methods for capturing ground reaction forces in the field. However, these forces are usually measured indirectly with non-interpretable ``black box'' models. Here we report an interpretable model with a function approximation algorithm. The model uses time series of plantar pressure from instrumented insoles to estimate vertical ground reaction forces. The study included data from 16 persons moving at different speeds on a two-belt treadmill equipped with force plates. The introduced regression model, based on a high-dimensional approximation, using only low-dimensional variable interactions, demonstrated its ability to estimate vertical ground reaction forces. The normalized mean square error for velocities between 3 and \\($\$\)9backslash,backslashhbox kmbackslash,backslashhbox \h\^\-1\\\($\$\)ranged from 10.6 to \\($\$\)24.4\backslash,\backslash\%\\($\$\). The accuracy of the presented approach, which can be used to analyze and interpret the learned model, was comparable to that reported in the literature. Furthermore, the evaluation of the learned model is particularly suitable for embedded and portable systems and, after a one-time calibration measurement, allows permanent and laboratory-independent measurement of vertical ground reaction forces.

Electrical impedance tomography (EIT) can monitor the influence of the muscle activity on the conductivity distribution across the forearm and is therefore suitable for Hand Sign Recognition (HSR). However, the method is relatively complex, so realizing real-time classification based on EIT measurements is still considered a major challenge. In addition, the accuracy of EIT for HSR is highly dependent on the injection configuration parameters applied. In this study, we investigate the influence of the injection configuration on the achieved accuracy of the hand robot control and the feasibility of a real-time classification based on EIT measurements. To assess real-time feasibility, experimental data have been collected for a sign set of 36 American Sign Language (ASL) performed by one healthy subject. Moreover, a comparative study of the adjacent, opposite, and cross injection configurations has been conducted on 10 subjects performing the ASL set. An ambiguity study based on the measured signals was performed to assess the discrimination potential between the hand signs. The adjacent EIT configuration exhibited higher sensitivity in discriminating hand signs. For a real-time classification, we propose the use of a Support Vector Machine (SVM) model realizing an overall accuracy of 71.38%. To validate the system's real-time behavior, the classification result is visualized by a robotic hand reproducing 6 different hand signs from the 36 ASL set reaching an accuracy of 80%. The results demonstrate the potential of EIT to facilitate real-time Hand Sign Recognition for robot control applications.

Several studies have shown that electrical impedance tomography (EIT) on the human forearm can be used to classify hand signs without the need for cameras or gloves. However, the amount of measured data is in relatively high, leading to long execution times and high processing complexity. In this study, we investigate the influence of reducing the number of EIT measurements on the classification accuracy. Four different algorithms were used to halve the number of measurements, namely Principal Component Analysis (PCA), Chi-square tests, Minimum Redundancy Maximum Relevance (MRMR) and Laplacian scores. These algorithms were applied to a dataset of forearm EIT measurements corresponding to the practice of American Sign Language (ASL) in a healthy subject. The results show a slight drop in accuracy when using the quadratic support vector machine (SVM) classifier, with accuracy dropping from 95.3% to 93.4% in the best cases. A method based on a genetic algorithm (GA) was also used. This approach exploits the strengths of genetic algorithms in exploring a large solution space and converging on the near-optimal combination. This approach resulted in an increase in accuracy to 96.7%. The study demonstrated that the number of EIT measurements can be reduced without compromising overall accuracy.

Handover actions are part of our daily lives. Whether it is the milk carton at the breakfast table or tickets at the box office, we usually perform these joint actions without much conscious attention. The individual actions involved in handovers, that have already been studied intensively at the level of individual actions, are grasping, lifting, and transporting objects. Depending on the object's properties, actors must plan their execution in order to ensure smooth and efficient object transfer. Therefore, anticipatory grip force scaling is crucial. Grip forces are planned in anticipation using weight estimates based on experience or visual cues. This study aimed to investigate whether receivers are able to correctly estimate object weight by observing the giver's kinematics. For this purpose, handover actions were performed with 20 dyads, manipulating the participant role (giver/receiver) and varying the size and weight of the object. Due to the random presentation of the object weight and the absence of visual cues, the participants were unaware of the object weight from trial to trial. Kinematics were recorded with a motion tracking system and grip forces were recorded with customized test objects. Peak grip force rates were used as a measure of anticipated object weight. Results showed that receiver kinematics are significantly affected by object weight. The peak grip force rates showed that receivers anticipate object weight, but givers not. This supports the hypothesis that receivers obtain information about the object weight by observing giver's kinematics and integrating this information into their own action execution.

Effective feature extraction and selection are crucial for the accurate classification and prediction of hand gestures based on electromyographic signals. In this paper, we systematically compare six filter and wrapper feature evaluation methods and investigate their respective impacts on the accuracy of gesture recognition. The investigation is based on several benchmark datasets and one real hand gesture dataset, including 15 hand force exercises collected from 14 healthy subjects using eight commercial sEMG sensors. A total of 37 time- and frequency-domain features were extracted from each sEMG channel. The benchmark dataset revealed that the minimum Redundancy Maximum Relevance (mRMR) feature evaluation method had the poorest performance, resulting in a decrease in classification accuracy. However, the RFE method demonstrated the potential to enhance classification accuracy across most of the datasets. It selected a feature subset comprising 65 features, which led to an accuracy of 97.14%. The Mutual Information (MI) method selected 200 features to reach an accuracy of 97.38%. The Feature Importance (FI) method reached a higher accuracy of 97.62% but selected 140 features. Further investigations have shown that selecting 65 and 75 features with the RFE methods led to an identical accuracy of 97.14%. A thorough examination of the selected features revealed the potential for three additional features from three specific sensors to enhance the classification accuracy to 97.38%. These results highlight the significance of employing an appropriate feature selection method to significantly reduce the number of necessary features while maintaining classification accuracy. They also underscore the necessity for further analysis and refinement to achieve optimal solutions.

We use hyperbolic wavelet regression for the fast reconstruction of high-dimensional functions having only low-dimensional variable interactions. Compactly supported periodic Chui-Wang wavelets are used for the tensorized hyperbolic wavelet basis on the torus. With a variable transformation, we are able to transform the approximation rates and fast algorithms from the torus to other domains. We perform and analyze scattered data approximation for smooth but arbitrary density functions by using a least squares method. The corresponding system matrix is sparse due to the compact support of the wavelets, which leads to a significant acceleration of the matrix vector multiplication. For non-periodic functions, we propose a new extension method. A proper choice of the extension parameter together with the piecewise polynomial Chui-Wang wavelets extends the functions appropriately. In every case, we are able to bound the approximation error with high probability. Additionally, if the function has a low effective dimension (i.e., only interactions of a few variables), we qualitatively determine the variable interactions and omit ANOVA terms with low variance in a second step in order to decrease the approximation error. This allows us to suggest an adapted model for the approximation. Numerical results show the efficiency of the proposed method.

Embodied Digital Technologies (EDTs) are increasingly populating private and public spaces. How EDTs are perceived in Hybrid Societies requires prior consideration. However, findings on social perception of EDTs remain inconclusive. We investigated social perception and trustworthiness of robots and telepresence systems (TPS) and aimed at identifying how observers' personality traits were associated with social perception of EDTs. To this end, we conducted two studies (N1\thinspace=\thinspace293, N2\thinspace=\thinspace305). Participants rated five different EDTs in a short video sequence of a space sharing conflict with a human in terms of anthropomorphism, sociability/morality, activity/cooperation, competence, and trustworthiness. The TPS were equipped with a tablet on which a person was visible. We found that the rudimentarily human-like TPS was perceived as more anthropomorphic than the automated guided vehicle, but no differences emerged in terms of other social dimensions. For robots, we found mixed results but overall higher ratings in terms of social dimensions for a human-like robot as opposed to a mechanical one. Trustworthiness was attributed differently to the EDTs only in Study 2, with a preference toward TPS and more human-like robots. In Study 1, we did not find any such differences. Personality traits were associated with attributions of social dimensions in Study 1, however results were not replicable and thus, associations remained ambiguous. With the present studies, we added insights on social perception of robots and provided evidence that social perception of TPS should be taken into consideration before their deployment.

This paper introduces a novel approach to addressing the challenge of accurately timing short distance runs, a critical aspect in the assessment of athletic performance. Electronic photoelectric barriers, although recognized for their dependability and accuracy, have remained largely inaccessible to non-professional athletes and smaller sport clubs due to their high costs. A comprehensive review of existing timing systems reveals that claimed accuracies beyond 30 ms lack experimental validation across most available systems. To bridge this gap, a mobile, camera-based timing system is proposed, capitalizing on consumer-grade electronics and smartphones to provide an affordable and easily accessible alternative. By leveraging readily available hardware components, the construction of the proposed system is detailed, ensuring its cost-effectiveness and simplicity. Experiments involving track and field athletes demonstrate the proficiency of the proposed system in accurately timing short distance sprints. Comparative assessments against a professional photoelectric cells timing system reveal a remarkable accuracy of 62 ms, firmly establishing the reliability and effectiveness of the proposed system. This finding places the camera-based approach on par with existing commercial systems, thereby offering non-professional athletes and smaller sport clubs an affordable means to achieve accurate timing. In an effort to foster further research and development, open access to the device’s schematics and software is provided. This accessibility encourages collaboration and innovation in the pursuit of enhanced performance assessment tools for athletes.

Sampling-based motion planning algorithms have been continuously developed for more than two decades. Apart from mobile robots, they are also widely used in manipulator motion planning. Hence, these methods play a key role in collaborative and shared workspaces. Despite numerous improvements, their performance can highly vary depending on the chosen parameter setting. The optimal parameters depend on numerous factors such as the start state, the goal state and the complexity of the environment. Practitioners usually choose these values using their experience and tedious trial and error experiments. To address this problem, recent works combine hyperparameter optimization methods with motion planning. They show that tuning the planner's parameters can lead to shorter planning times and lower costs. It is not clear, however, how well such approaches generalize to a diverse set of planning problems that include narrow passages as well as barely cluttered environments. In this work, we analyze optimized planner settings for a large set of diverse planning problems. We then provide insights into the connection between the characteristics of the planning problem and the optimal parameters. As a result, we provide a list of recommended parameters for various use-cases. Our experiments are based on a novel motion planning benchmark for manipulators which we provide at https://mytuc.org/rybj.