Publications

Attention-based Intrinsic Reward Mixing Network for Credit Assignment in Multi-Agent Reinforcement Learning

Credit assignment is a critical problem in cooperative Multi-Agent Reinforcement Learning (MARL). To address this problem, current studies mainly rely on the intrinsic reward, which is directly summed with the global reward to generate a total reward. However, such kinds of intrinsic reward functions ignore the dependence among agents and inevitably limit the adaptivity and effectiveness of MARL methods. In this paper, we propose a novel method, Attention-based Intrinsic Reward Mixing Network (AIRMN), for credit assignment in MARL. Specifically, we design a new intrinsic reward network on the basis of the attention mechanism, in order to enhance the effectiveness of teamwork. Besides, we devise a new mixing network that combines the intrinsic and extrinsic rewards in a nonlinear and dynamic manner, so as to adapt the total reward to the variation of the environment. Experimental results on the battle games of StarCraft II demonstrate that AIRMN outperforms the state-of-the-art methods in terms of the average test win rate, and also validate that AIRMN can dynamically return the precise intrinsic reward to each agent based on their contributions to the team cooperation, thereby better dealing with the credit assignment problem.

Robot Indoor Scenes Recognition Based on Autonomous Developmental Neural Network

A method of robot indoor scene recognition based on autonomous developmental neural network is proposed.3-layer adaptive developmental neural network is used to build the brain-mind model.In developing phase, top-k competition is utilized to simulate the lateral inhibition of neurons, and the winner updates the synapse weight vector with the lobe component analysis (LCA) algorithm.The strengthened neurons can get thinking results according to current environment information, and indoor scenes can be recognized autonomously by mobile robots.Through human-like thinking, the learning results are stored as "knowledge", and the thinking results are derived from experience.Experimental results show that the model of autonomous developmental neural network proposed as the carrier of "knowledge", fully meets the need of indoor scenes recognition task, and realizes the autonomous learning, understanding and growth of robots based on vision.

MS-FRAN: A Novel Multi-Source Domain Adaptation Method for EEG-Based Emotion Recognition

Electroencephalogram (EEG)-based emotion recognition has gradually become a research hotspot. However, the large distribution differences of EEG signals across subjects make the current research stuck in a dilemma. To resolve this problem, in this article, we propose a novel and effective method, Multi-Source Feature Representation and Alignment Network (MS-FRAN). The effectiveness of proposed method mainly comes from three new modules: Wide Feature Extractor (WFE) for feature learning, Random Matching Operation (RMO) for model training, and Top- h ranked domain classifier selection (TOP) for emotion classification. MS-FRAN is not only effective in aligning the distributions of each pair of source and target domains, but also capable of reducing the distributional differences among the multiple source domains. Experimental results on the public benchmark datasets SEED and DEAP have demonstrated the advantage of our method over the related competitive approaches for cross-subject EEG-based emotion recognition.

Point cloud augmented virtual reality environment with haptic constraints for teleoperation

Remote manipulation of a robot without assistance in an unstructured environment is a challenging task for operators. In this paper, a novel methodology for haptic constraints in a point cloud augmented virtual reality environment is proposed to address this human operation limitation. The proposed method generates haptic constraints in real time for an unstructured environment, including regional constraints and guidance constraints. A modified implicit surface method is applied for regional constraint generation for the entire point cloud. Additionally, the isosurface derived from the implicit surface is proposed for real-time three-dimensional artificial force field estimation. For guidance constraint generation, a new incremental prediction and local artificial force field generation method based on the modified sigmoid model is proposed in an unstructured point cloud virtual reality environment. With the generated haptic constraints, the operator can control the robot to realize obstacle avoidance and easily reach the target tasks. System evaluation is conducted, and the result demonstrates the effectiveness of the proposed method. In addition, a 10-participant study with users who control the robot to three specific targets shows that the system can enhance human operation efficiency and reduce time costs by at least 59% compared with no-haptic-constraint operations. Additionally, the designed questionnaire also demonstrates that the proposed methodology can reduce the workload during human operations.

Noncontact Detection Method for Food Texture Assessment Based on Air Puff Combined With Structured Light Imaging

Given the shortcomings of subjectivity and destructiveness associated with traditional methods of measuring food texture parameters (such as fruit firmness and meat tenderness) when using a texture meter, a non-contact measurement method utilizing an Air-puff combined with Structured Light Imaging (ASLI) was proposed in this study. First, the finite element simulation method was used to simulate the airflow impact model and optimize the ranges of the angle and distance parameters. The detection device was then developed, and binocular 3D phenotype acquisition and 3D feature extraction algorithms were investigated. Denoising, point cloud segmentation, greedy projection triangulation, Delaunay triangulation, and surface fitting algorithms were used to process the point cloud, and concave parameters such as depth, mapping area, surface area, and volume of the concave area on the beef surface were obtained. Finally, four typical samples with different meat (beef and chicken) and fruit (kiwifruit and peach) parameters were tested. Five machine learning methods were used to model and analyze the correlation among texture parameters. The results revealed that the model established by the ensemble learning modeling method with extracted concave parameters possessed the highest accuracy. The prediction accuracy of the shear force values of the beef and chicken samples were 0.92 and 0.91. The accuracy of meat and fruit grading was 0.942 and 1.0. Therefore, the non-contact detection method proposed in this study to determine food texture characteristics based on airflow can be used to replace the traditional texture analyzer in food engineering and exhibits good application potential.

Experimental Analysis on the Effectiveness of Kinematic Error Compensation Methods for Serial Industrial Robots

Based on the established serial 6-DOF robot calibration experiment platform, this paper aims to analyze and compare the effects of four error compensation methods, which are pseudotarget iteration-based error compensation method with three different forms and the Newton–Raphson-based error compensation method. Firstly, the pose error model of the serial robot is established based on the M-DH model in this paper. The calibration results show that the accuracy of the Staubli TX60 robot has been greatly improved. The average comprehensive position accuracy is increased by 88.7%, and the average comprehensive attitude accuracy is increased by 56.6%. Secondly, the principles of the four error compensation methods are discussed, and the effectiveness of the four error compensation methods are compared through experiments. The results show that the four error compensation methods can achieve error compensation well. The compensation accuracy is consistent with the identification accuracy of the kinematic model. The pseudotarget iteration with differential form has the best performance by the comprehensive consideration of accuracy and computational efficiency. Error compensation determines whether the accuracy of the identified model can be achieved. This paper presents a systematic experimental validation research on the effectiveness of four error compensation methods, which provides a reliable reference for the kinematic error compensation of industrial robots.

Early stroke diagnosis and evaluation based on pathological voice classification using speech enhancement

The proposed ESD method combining SE with SVs and SS can serve as an assistive diagnostic tool to help medical professionals and individuals detect and prevent strokes at an earlier stage, reduce workload, and improve identification objectivity. The code and experimental protocol of this paper are available at https://github.com/LiuYingchenseu/ESPVC.

Multi-dimensional Texture Haptic Cross-modal Generation and Display Method based on Bi-Mamba Network

Design, Control and Simulation of a Leg-Wheel Robot: STransleg

A versatile leg-wheel robot called STransleg is proposed in this paper. STransleg is propelled by four identical leg-wheel mechanisms, which can transform between the legged and wheeled modes. The leg has three active joints in the legged mode, and the wheel is active in the wheeled mode. The waist yaw and pitch joints are introduced to STransleg to improve the locomotion flexibility and stability. In this paper, the design and kinematics modeling methods of STransleg are described. Additionally, the control method is proposed for STransleg to make it do the four-wheeled locomotion, walking locomotion, and carriage-like locomotion. The simulations are carried out to make the dynamical model of STransleg do the different locomotion, and the results verify the feasibility of the robot design method and the proposed control method.

Multi-Dimensional Force Sensor Design for Haptic Human-Computer Interaction

Haptic human-computer interaction (HapHCI) is interaction between a human and a computer with realistic sense of touch. Haptic interaction between human and computer involves solving challenging problems in mechanical design, sensor, actuator, computer graphics, physical-based modelling and rendering algorithm, human capabilities, and other areas. With the increasing applications of HapHCI in virtual reality, teleoperation, rehabilitation, tele-surgery, entertainment, etc, the importance of the sense of touch for human-computer interaction has been widely acknowledged [Gabriel 2006]. For example, in virtual surgery training system, the surgeon controls the surgical tools and characterizes virtual tissues as normal or abnormal through the sense of touch provided by the HapHCI device. Another example is HapHCI based rehabilitation system for post-stroke patient exercise. During active rehabilitation exercise process, it requires accurate damping force control, and during passive rehabilitation exercise process, relatively accurate traction force is necessary. HapHCI technique usually consists of three fundamental parts: force/tactile measuring, haptic modelling, and haptic display device. Haptic modelling as well as haptic display hardware has been discussed a lot and exploited for ten years particularly in the area of virtual reality. However, so far little attention has been paid to the design of multidimensional force sensor for HapHCI, and the existing commercial six degree-of-freedom (DOF) force sensors are designed mainly for industrial robot control, which are too expensive and often over designed for HapHCI in axis and in bandwidth. As an important component in the HapHCI system, multi-dimensional force sensor not only measures the human hand force/torque acted on the interactive hardware device, such as handcontroller, master-manipulator, joystick etc, as a command input the computer, but also provides force/torque information for close-loop control of precise haptic display. A number of multi-dimensional force sensors have been developed during the past decades, which are intended for use at the end effector of a robot to monitor assembly or machine force. Most of them are six axes force/torque sensors [Watson, Drake, 1975] [Lord Corporation, 1985] [Nakamura et al., 1987] [Kaneko, Nishihara, 1993] [Kim, 2001], which measure three axes forces Fx, Fy, Fz, and three axes torques Mx, My, Mz. And some of them

Continuous Wrist Angle Estimation Under Different Resistance Based on Dynamic EMG Decomposition

Estimating wrist movements through neural drives is crucial in human-machine interface (HMI). However, studies on wrist movements mostly focused on isometric contractions, while research on dynamic EMG decomposition during non-stationary movements is notably scarce. Moreover, the impact of different resistance on the motor unit (MU) decomposition and wrist angle estimation remains unexplored. To address these gaps, this paper proposed a novel framework to decode neural drives from EMG signals during dynamic wrist movements. Specifically, the EMG signals were divided into short segments firstly. Next, progressive FastICA peel-off (PFP) algorithm was utilized to decompose each EMG segment into motor unit spike trains (MUST). Then, a linear window function was applied to track the motor units (MU) to obtain complete MUSTs. Three resistance levels were investigated during wrist flexion and extension: 20%, 40%, and 60% maximum voluntary contraction (MVC). Multiple linear regression (LR) and convolutional neural network (CNN) were used to estimate wrist angles within a range of ± 20° based on neural drives. Results showed the proposed framework could effectively identify MUs at these three resistance levels, with an average global pulse-to-noise ratio (PNR) above 20 dB. The determination coefficients of LR model were 0.92 ± 0.06, 0.91 ± 0.07, and 0.85 ± 0.13 at the three resistance levels, respectively, while those of CNN were 0.88 ± 0.10, 0.88 ± 0.11, and 0.81 ± 0.17. This study demonstrates it is feasible to estimate wrist angles based on decomposed neural drives at different resistance levels, and has significant implications for HMI development.

Spiking Spatio-Temporal Neural Architecture Search for EEG-Based Emotion Recognition

Evolving Gaussian RBF network for nonlinear time series modelling and prediction

A genetic algorithm and recursive least squares (RLS) learning algorithm for a Gaussian radial basis function network is described, for modelling and predicting nonlinear time series. Better generalisation performance can be achieved than that of the usual clustering and RLS method.

On sliding mode control of single input Markovian jump systems

Functional Assessment of A Flexible Thermal Stimulation Device for Hand Edema Elimination in Stroke Survivors

Stroke survivors often suffer from severe discomfort caused by persistent hand edema, including hand pain, restricted knuckle mobility, and negative emotions, which significantly hinders normal rehabilitation progress. Whether physical or psychological in impact, the prevention and treatment of hand edema warrant significant attention throughout all stages of stroke rehabilitation. However, traditional rehabilitation therapies generally neglect edema management, focusing primarily on motor function recovery. Consequently, this study introduces the TESSR (ThErmal Stimulation device for Stroke Rehabilitation), a novel apparatus designed to promote hand rehabilitation for stroke survivors. TESSR is equipped with a flexible stimulation array comprising 50 independently controlled stimulation units, each capable of rapid heating and cooling transitions as well as delivering sustained and stable thermal outputs. This alternating thermal stimulation (TS) is designed to alleviate edema and activate the sensory cortex, thereby facilitating neuroplasticity. A controlled clinical study was conducted with 12 stroke survivors to assess the effectiveness of the TESSR device. The participants were divided into two groups: one receiving conventional rehabilitation (CR group) and the other receiving supplementary TS conducted by the TESSR device (TS group). Following a two-week intervention, the TS group exhibited a notable reduction in hand edema, with a mean decrease of 38.6%, compared to the minimal improvement observed in the CR group. Furthermore, participants in the TS group demonstrated improved hand function, including enhanced sensory perception and diminished pain. The findings demonstrate the potential of TESSR as a comprehensive and effective tool for hand rehabilitation.

3D-point-cloud registration and real-world dynamic modelling-based virtual environment building method for teleoperation- CORRIGENDUM

The Figure 4 in the original version has a stylistic error. It ought to use the figure as follows: The authors apologise for this error.

Force-regulated Elastic Linear Objects Tracking for Virtual and Augmented Reality

Elastic rods are commonly seen in our daily life. Although humans are sensitive to the shape change of rods, it is not intuitive to estimate the external forces applied to generate the deformation. We propose a method to interactively track the elastic linear objects by using the Cosserat rod model to regulate the captured noisy points. We develop a framework based on particle filters to work with the physics-based model, turning the inverse physics problem into a forward simulation and search problem. We show that with the proposed method, we can simultaneously digitalize the shape as well as the external forces on real-world elastic rods. With these capabilities, we demonstrate virtual and augmented reality applications to facilitate the interaction with elastic linear objects. The tracking performance is also validated with experiments.

Research on gear fault diagnosis based on feature fusion optimization and improved two hidden layer extreme learning machine