Authors:Toon Van de Maele, Tim Verbelen, Ozan Çatal, Bart Dhoedt Abstract: Scene understanding and decomposition is a crucial challenge for intelligent systems, whether it is for object manipulation, navigation, or any other task. Although current machine and deep learning approaches for object detection and classification obtain high accuracy, they typically do not leverage interaction with the world and are limited to a set of objects seen during training. Humans on the other hand learn to recognize and classify different objects by actively engaging with them on first encounter. Moreover, recent theories in neuroscience suggest that cortical columns in the neocortex play an important role in this process, by building predictive models about objects in their reference frame. In this article, we present an enactive embodied agent that implements such a generative model for object interaction. For each object category, our system instantiates a deep neural network, called Cortical Column Network (CCN), that represents the object in its own reference frame by learning a generative model that predicts the expected transform in pixel space, given an action. The model parameters are optimized through the active inference paradigm, i.e., the minimization of variational free energy. When provided with a visual observation, an ensemble of CCNs each vote on their belief of observing that specific object category, yielding a potential object classification. In case the likelihood on the selected category is too low, the object is detected as an unknown category, and the agent has the ability to instantiate a novel CCN for this category. We validate our system in an simulated environment, where it needs to learn to discern multiple objects from the YCB dataset. We show that classification accuracy improves as an embodied agent can gather more evidence, and that it is able to learn about novel, previously unseen objects. Finally, we show that an agent driven through active inference can choose their actions to reach a preferred observation. PubDate: 2022-04-14T00:00:00Z
Authors:Gualtiero Piccinini Abstract: Situated approaches to cognition maintain that cognition is embodied, embedded, enactive, and affective (and extended, but that is not relevant here). Situated approaches are often pitched as alternatives to computational and representational approaches, according to which cognition is computation over representations. I argue that, far from being opposites, situatedness and neural representation are more deeply intertwined than anyone suspected. To show this, I introduce a neurocomputational account of cognition that relies on neural representations. I argue not only that this account is compatible with (non-question-begging) situated approaches, but also that it requires embodiment, embeddedness, enaction, and affect at its very core. That is, constructing neural representations and their semantic content, and learning computational processes appropriate for their content, requires a tight dynamic interaction between nervous system, body, and environment. Most importantly, I argue that situatedness is needed to give a satisfactory account of neural representation: neurocognitive systems that are embodied, embedded, affective, dynamically interact with their environment, and use feedback from their interaction to shape their own representations and computations (1) can construct neural representations with original semantic content, (2) their neural vehicles and the way they are processed are automatically coordinated with their content, (3) such content is causally efficacious, (4) is determinate enough for the system's purposes, (5) represents the distal stimulus, and (6) can misrepresent. This proposal hints at what is needed to build artifacts with some of the basic cognitive capacities possessed by neurocognitive systems. PubDate: 2022-04-14T00:00:00Z
Authors:Fei Wang, Wei Wang, Dan Wu, Guowang Gao Abstract: In obtaining color constancy, estimating the illumination of a scene is the most important task. However, due to unknown light sources and the influence of the external imaging environment, the estimated illumination is prone to color ambiguity. In this article, a learning-based multi-scale region-weighed network guided by semantic features is proposed to estimate the illuminated color of the light source in a scene. Cued by the human brain's processing of color constancy, we use image semantics and scale information to guide the process of illumination estimation. First, we put the image and its semantics into the network, and then obtain the region weights of the image at different scales. After that, through a special weight-pooling layer (WPL), the illumination on each scale is estimated. The final illumination is calculated by weighting each scale. The results of extensive experiments on Color Checker and NUS 8-Camera datasets show that the proposed approach is superior to the current state-of-the-art methods in both efficiency and effectiveness. PubDate: 2022-04-08T00:00:00Z
Authors:Oliver Heeb, Arnab Barua, Carlo Menon, Xianta Jiang Abstract: Ankle joint power is usually determined by a complex process that involves heavy equipment and complex biomechanical models. Instead of using heavy equipment, we proposed effective machine learning (ML) and deep learning (DL) models to estimate the ankle joint power using force myography (FMG) sensors. In this study, FMG signals were collected from nine young, healthy participants. The task was to walk on a special treadmill for five different velocities with a respective duration of 1 min. FMG signals were collected from an FMG strap that consists of 8 force resisting sensor (FSR) sensors. The strap was positioned around the lower leg. The ground truth value for ankle joint power was determined with the help of a complex biomechanical model. At first, the predictors' value was preprocessed using a rolling mean filter. Following, three sets of features were formed where the first set includes raw FMG signals, and the other two sets contained time-domain and frequency-domain features extracted using the first set. Cat Boost Regressor (CBR), Long-Short Term Memory (LSTM), and Convolutional Neural Network (CNN) were trained and tested using these three features sets. The results presented in this study showed a correlation coefficient of R = 0.91 ± 0.07 for intrasubject testing and were found acceptable when compared to other similar studies. The CNN on raw features and the LSTM on time-domain features outperformed the other variations. Aside from that, a performance gap between the slowest and fastest walking distance was observed. The results from this study showed that it was possible to achieve an acceptable correlation coefficient in the prediction of ankle joint power using FMG sensors with an appropriate combination of feature set and ML model. PubDate: 2022-04-01T00:00:00Z
Authors:Nianhao Xie, Yunpeng Hu, Lei Chen Abstract: Distributed control method plays an important role in the formation of a multi-agent system (MAS), which is the prerequisite for an MAS to complete its missions. However, the lack of considering the collision risk between agents makes many distributed formation control methods lose practicability. In this article, a distributed formation control method that takes collision avoidance into account is proposed. At first, the MAS formation control problem can be divided into pair-wise unit formation problems where each agent moves to the expected position and only needs to avoid one obstacle. Then, a deep Q network (DQN) is applied to model the agent's unit controller for this pair-wise unit formation. The DQN controller is trained by using reshaped reward function and prioritized experience replay. The agents in MAS formation share the same unit DQN controller but get different commands due to various observations. Finally, through the min-max fusion of value functions of the DQN controller, the agent can always respond to the most dangerous avoidance. In this way, we get an easy-to-train multi-agent collision avoidance formation control method. In the end, unit formation simulation and multi-agent formation simulation results are presented to verify our method. PubDate: 2022-03-31T00:00:00Z
Authors:Woolim Hong, Jinwon Lee, Pilwon Hur Abstract: Human gait phase estimation has been studied in the field of robotics due to its importance for controlling wearable devices (e.g., prostheses or exoskeletons) in a synchronized manner with the user. As data-driven approaches have recently risen in the field, researchers have attempted to estimate the user gait phase using a learning-based method. Thigh and torso information have been widely utilized in estimating the human gait phase for wearable devices. Torso information, however, is known to have high variability, specifically in slow walking, and its effect on gait phase estimation has not been studied. In this study, we quantified torso variability and investigated how the torso information affects the gait phase estimation result at various walking speeds. We obtained three different trained models (i.e., general, slow, and normal-fast models) using long short-term memory (LSTM). These models were compared to identify the effect of torso information at different walking speeds. In addition, the ablation study was performed to identify the isolated effect of the torso on the gait phase estimation. As a result, when the torso segment's angular velocity was used with thigh information, the accuracy of gait phase estimation was increased, while the torso segment's angular position had no apparent effect on the accuracy. This study suggests that the torso segment's angular velocity enhances human gait phase estimation when used together with the thigh information despite its known variability. PubDate: 2022-03-30T00:00:00Z
Authors:Le Wu, Xun Chen, Xiang Chen, Xu Zhang Abstract: The objective of this study is to develop a method for alleviating a novel pattern interference toward achieving a robust myoelectric pattern-recognition control system. To this end, a framework was presented for surface electromyogram (sEMG) pattern classification and novelty detection using hybrid neural networks, i.e., a convolutional neural network (CNN) and autoencoder networks. In the framework, the CNN was first used to extract spatio-temporal information conveyed in the sEMG data recorded via high-density (HD) 2-dimensional electrode arrays. Given the target motion patterns well-characterized by the CNN, autoencoder networks were applied to learn variable correlation in the spatio-temporal information, where samples from any novel pattern appeared to be significantly different from those from target patterns. Therefore, it was straightforward to discriminate and then reject the novel motion interferences identified as untargeted and unlearned patterns. The performance of the proposed method was evaluated with HD-sEMG data recorded by two 8 × 6 electrode arrays placed over the forearm extensors and flexors of 9 subjects performing seven target motion tasks and six novel motion tasks. The proposed method achieved high accuracies over 95% for identifying and rejecting novel motion tasks, and it outperformed conventional methods with statistical significance (p < 0.05). The proposed method is demonstrated to be a promising solution for rejecting novel motion interferences, which are ubiquitous in myoelectric control. This study will enhance the robustness of the myoelectric control system against novelty interference. PubDate: 2022-03-28T00:00:00Z
Authors:Ryuta Mizutani, Senta Noguchi, Rino Saiga, Yuichi Yamashita, Mitsuhiro Miyashita, Makoto Arai, Masanari Itokawa Abstract: We have reported nanometer-scale three-dimensional studies of brain networks of schizophrenia cases and found that their neurites are thin and tortuous when compared to healthy controls. This suggests that connections between distal neurons are suppressed in microcircuits of schizophrenia cases. In this study, we applied these biological findings to the design of a schizophrenia-mimicking artificial neural network to simulate the observed connection alteration in the disorder. Neural networks that have a “schizophrenia connection layer” in place of a fully connected layer were subjected to image classification tasks using the MNIST and CIFAR-10 datasets. The results revealed that the schizophrenia connection layer is tolerant to overfitting and outperforms a fully connected layer. The outperformance was observed only for networks using band matrices as weight windows, indicating that the shape of the weight matrix is relevant to the network performance. A schizophrenia convolution layer was also tested using the VGG configuration, showing that 60% of the kernel weights of the last three convolution layers can be eliminated without loss of accuracy. The schizophrenia layers can be used instead of conventional layers without any change in the network configuration and training procedures; hence, neural networks can easily take advantage of these layers. The results of this study suggest that the connection alteration found in schizophrenia is not a burden to the brain, but has functional roles in brain performance. PubDate: 2022-03-28T00:00:00Z
Authors:Miguel Nobre Castro, Strahinja Dosen Abstract: Modern myoelectric prostheses can perform multiple functions (e.g., several grasp types and wrist rotation) but their intuitive control by the user is still an open challenge. It has been recently demonstrated that semi-autonomous control can allow the subjects to operate complex prostheses effectively; however, this approach often requires placing sensors on the user. The present study proposes a system for semi-autonomous control of a myoelectric prosthesis that requires a single depth sensor placed on the dorsal side of the hand. The system automatically pre-shapes the hand (grasp type, size, and wrist rotation) and allows the user to grasp objects of different shapes, sizes and orientations, placed individually or within cluttered scenes. The system “reacts” to the side from which the object is approached, and enables the user to target not only the whole object but also an object part. Another unique aspect of the system is that it relies on online interaction between the user and the prosthesis; the system reacts continuously on the targets that are in its focus, while the user interprets the movement of the prosthesis to adjust aiming. Experimental assessment was conducted in ten able-bodied participants to evaluate the feasibility and the impact of training on prosthesis-user interaction. The subjects used the system to grasp a set of objects individually (Phase I) and in cluttered scenarios (Phase II), while the time to accomplish the task (TAT) was used as the performance metric. In both phases, the TAT improved significantly across blocks. Some targets (objects and/or their parts) were more challenging, requiring thus significantly more time to handle, but all objects and scenes were successfully accomplished by all subjects. The assessment therefore demonstrated that the system is indeed robust and effective, and that the subjects could successfully learn how to aim with the system after a brief training. This is an important step toward the development of a self-contained semi-autonomous system convenient for clinical applications. PubDate: 2022-03-25T00:00:00Z
Authors:Andrea Spyrantis, Tirza Woebbecke, Daniel Rueß, Anne Constantinescu, Andreas Gierich, Klaus Luyken, Veerle Visser-Vandewalle, Eva Herrmann, Florian Gessler, Marcus Czabanka, Harald Treuer, Maximilian Ruge, Thomas M. Freiman Abstract: BackgroundThe development of robotic systems has provided an alternative to frame-based stereotactic procedures. The aim of this experimental phantom study was to compare the mechanical accuracy of the Robotic Surgery Assistant (ROSA) and the Leksell stereotactic frame by reducing clinical and procedural factors to a minimum.MethodsTo precisely compare mechanical accuracy, a stereotactic system was chosen as reference for both methods. A thin layer CT scan with an acrylic phantom fixed to the frame and a localizer enabling the software to recognize the coordinate system was performed. For each of the five phantom targets, two different trajectories were planned, resulting in 10 trajectories. A series of five repetitions was performed, each time based on a new CT scan. Hence, 50 trajectories were analyzed for each method. X-rays of the final cannula position were fused with the planning data. The coordinates of the target point and the endpoint of the robot- or frame-guided probe were visually determined using the robotic software. The target point error (TPE) was calculated applying the Euclidian distance. The depth deviation along the trajectory and the lateral deviation were separately calculated.ResultsRobotics was significantly more accurate, with an arithmetic TPE mean of 0.53 mm (95% CI 0.41–0.55 mm) compared to 0.72 mm (95% CI 0.63–0.8 mm) in stereotaxy (p < 0.05). In robotics, the mean depth deviation along the trajectory was −0.22 mm (95% CI −0.25 to −0.14 mm). The mean lateral deviation was 0.43 mm (95% CI 0.32–0.49 mm). In frame-based stereotaxy, the mean depth deviation amounted to −0.20 mm (95% CI −0.26 to −0.14 mm), the mean lateral deviation to 0.65 mm (95% CI 0.55–0.74 mm).ConclusionBoth the robotic and frame-based approach proved accurate. The robotic procedure showed significantly higher accuracy. For both methods, procedural factors occurring during surgery might have a more relevant impact on overall accuracy. PubDate: 2022-03-25T00:00:00Z
Authors:Ping Jiang, Junji Oaki, Yoshiyuki Ishihara, Junichiro Ooga, Haifeng Han, Atsushi Sugahara, Seiji Tokura, Haruna Eto, Kazuma Komoda, Akihito Ogawa Abstract: Deep learning has been widely used for inferring robust grasps. Although human-labeled RGB-D datasets were initially used to learn grasp configurations, preparation of this kind of large dataset is expensive. To address this problem, images were generated by a physical simulator, and a physically inspired model (e.g., a contact model between a suction vacuum cup and object) was used as a grasp quality evaluation metric to annotate the synthesized images. However, this kind of contact model is complicated and requires parameter identification by experiments to ensure real world performance. In addition, previous studies have not considered manipulator reachability such as when a grasp configuration with high grasp quality is unable to reach the target due to collisions or the physical limitations of the robot. In this study, we propose an intuitive geometric analytic-based grasp quality evaluation metric. We further incorporate a reachability evaluation metric. We annotate the pixel-wise grasp quality and reachability by the proposed evaluation metric on synthesized images in a simulator to train an auto-encoder–decoder called suction graspability U-Net++ (SG-U-Net++). Experiment results show that our intuitive grasp quality evaluation metric is competitive with a physically-inspired metric. Learning the reachability helps to reduce motion planning computation time by removing obviously unreachable candidates. The system achieves an overall picking speed of 560 PPH (pieces per hour). PubDate: 2022-03-24T00:00:00Z
Authors:Xianmin Wang, Jing Li, Qi Liu, Wenpeng Zhao, Zuoyong Li, Wenhao Wang Abstract: Neural networks have played critical roles in many research fields. The recently proposed adversarial training (AT) can improve the generalization ability of neural networks by adding intentional perturbations in the training process, but sometimes still fail to generate worst-case perturbations, thus resulting in limited improvement. Instead of designing a specific smoothness function and seeking an approximate solution used in existing AT methods, we propose a new training methodology, named Generative AT (GAT) in this article, for supervised and semi-supervised learning. The key idea of GAT is to formulate the learning task as a minimax game, in which the perturbation generator aims to yield the worst-case perturbations that maximize the deviation of output distribution, while the target classifier is to minimize the impact of this perturbation and prediction error. To solve this minimax optimization problem, a new adversarial loss function is constructed based on the cross-entropy measure. As a result, the smoothness and confidence of the model are both greatly improved. Moreover, we develop a trajectory-preserving-based alternating update strategy to enable the stable training of GAT. Numerous experiments conducted on benchmark datasets clearly demonstrate that the proposed GAT significantly outperforms the state-of-the-art AT methods in terms of supervised and semi-supervised learning tasks, especially when the number of labeled examples is rather small in semi-supervised learning. PubDate: 2022-03-24T00:00:00Z
Authors:Xianfa Xue, Haohui Huang, Lei Zuo, Ning Wang Abstract: To meet the enormous demand for smart manufacturing, industrial robots are playing an increasingly important role. For industrial operations such as grinding 3C products, numerous demands are placed on the compliant interaction ability of industrial robots to interact in a compliant manner. In this article, an adaptive compliant control framework for robot interaction is proposed. The reference trajectory is obtained by single-point demonstration and DMP generalization. The adaptive feedforward and impedance force controller is derived in terms of position errors, and they are input into an admittance controller to obtain the updated amount of position deviation. The compliant interaction effect is achieved, which is shown that the grinding head fits on the curved surface of a computer mouse, and the interaction force is within a certain expected range in the grinding experiment based on the performance an Elite robot. A comparative experiment was conducted to demonstrate the effectiveness of the proposed framework in a more intuitive way. PubDate: 2022-03-23T00:00:00Z
Authors:Nan Lou, Yanan Diao, Qiangqiang Chen, Yunkun Ning, Gaoqiang Li, Shengyun Liang, Guanglin Li, Guoru Zhao Abstract: Knee osteoarthritis is a degenerative disease, which greatly affects the daily life of patients. Total knee replacement (TKR) is the most common method to treat knee joint disorders and relieve knee pain. Postoperative rehabilitation exercise is the key to restore knee joint function. However, there is a lack of a portable equipment for monitoring knee joint activity and a systematic assessment scheme. We have developed a portable rehabilitation monitoring and evaluation system based on the wearable inertial unit to estimate the knee range of motion (ROM). Ten TKR patients and ten healthy adults are recruited for the experiment, then the system performance is verified by professional rehabilitation equipment Baltimore Therapeutic Equipment (BTE) Primus RS. The average absolute difference between the knee ROM and BTE Primus RS of healthy subjects and patients ranges from 0.16° to 4.94°. In addition, the knee ROM of flexion-extension and gait activity between healthy subjects and patients showed significant differences. The proposed system is reliable and effective in monitoring and evaluating the rehabilitation progress of patients. The system proposed in this work is expected to be used for long-term effective supervision of patients in clinical and dwelling environments. PubDate: 2022-03-23T00:00:00Z
Authors:Xu Liang, Tingting Su, Zhonghai Zhang, Jie Zhang, Shengda Liu, Quanliang Zhao, Junjie Yuan, Can Huang, Lei Zhao, Guangping He Abstract: Aiming at the situation that the structural parameters of the general manipulators are uncertain, a time-varying impedance controller based on model reference adaptive control (MRAC) is proposed in this article. The proposed controller does not need to use acceleration-based feedback or to measure external loads and can tolerate considerable structure parameter errors. The global uniform asymptotic stability of the time-varying closed-loop system is analyzed, and a selection approach for control parameters is presented. It is demonstrated that, by using the proposed control parameter selection approach, the closed-loop system under the adaptive controller is equivalent to an existing result. The feasibility of the presented controller for the general manipulators is demonstrated by some numerical simulations. PubDate: 2022-03-18T00:00:00Z
Authors:Xuezhen Cheng, Chuannuo Xu, Xiaoqing Liu, Jiming Li, Junming Zhang Abstract: This article aims to address problems in the current clustering process of low-energy adaptive clustering hierarchy (LEACH) in the wireless sensor networks, such as strong randomness and local optimum in the path optimization. This article proposes an optimal combined weighting (OCW) and improved ant colony optimization (IACO) algorithm for the LEACH protocol optimization. First, cluster head nodes are updated via a dynamic replacement mechanism of the whole network cluster head nodes to reduce the network energy consumption. In order to improve the quality of the selected cluster head nodes, this article proposes the OCW method to dynamically change the weight according to the importance of the cluster head node in different regions, in accordance with the three impact factors of the node residual energy, density, and distance between the node and the sink node in different regions. Second, the network is partitioned and the transmission path among the clusters can be optimized by the transfer probability in IACO with combined local and global pheromone update mechanism. The efficacy of the proposed LEACH protocol optimization method has been verified with MATLAB simulation experiments. PubDate: 2022-03-18T00:00:00Z
Open Access journal
