Publications

Abstract Chinese mitten crab has unique limb structures composed of a hard exoskeleton and flexible muscles. They enable the crab to locomote adaptively and safely on various terrains. In this work, we investigated the limb structures, motion principle, and gaits of the crab using a high-speed camera and a press machine. Then, a novel compliant robot leg design method is proposed, inspired by the crab limb. The leg comprises six hard scleromeres and a flexible thin-wall spring steel sheet (FSSS) mimicking the exoskeleton and muscle. The scleromeres connected one by one with rotational joints are designed with slots. The front end of the FSSS is fixed on the scleromere close to the ground. The rear end crosses the slots and is mounted at the shaft of a linear actuator installed at the rear scleromere. The leg bends and stretches when the actuator pushes and pulls the FSSS, respectively. The kinematic modeling, rigid-flexible coupling dynamic simulations, and leg prototype tests are conducted, which verify the leg design approach. Thirdly, we put forward a multi-legged robot with eight compliant legs and design its gait using the gaits of the crab. Finally, the robot’s performance is evaluated, including the capabilities of walking on different terrains at adjustable speeds and body heights, traversing low channels, walking on slopes, and carrying loads. The results prove that the single-motor-actuated compliant legs and their dynamic coupling with the rigid robot body frame can enable them to have the ground clearance ability and realize the adaptive walking of the robot. The leg design methodology can be used to design multi-legged robots with the merits of compact, light, low mechanical complexity, high safety, and easy to control, for many applications, such as environmental monitoring, search and rescue.

Lightweight CNNs and Transformers have shown great promise in sensor-based human activity recognition (HAR), yet their structural synergies remain underexplored. This paper bridges this gap by integrating the MetaFormer paradigm—a general architecture abstracted from Transformers that structurally separates token mixing (i.e., self-attention) and channel mixing (i.e., feed-forward networks)—into efficient CNN design. While MetaFormer offers a powerful inductive bias, its standard self-attention mechanism is often computationally intensive for resource-constrained HAR. To address this, we revolutionize the classic MobileNetV3 architecture from a MetaFormer perspective, introducing Rep-MMB, a new family of pure lightweight CNNs. By leveraging structural reparameterization, Rep-MMB decouples multi-branch training-time complexity from efficient single-branch inference, enabling high accuracy with low latency. Evaluations on four public HAR benchmarks show that Rep-MMB outperforms state-of-the-art lightweight models in accuracy and efficiency, with practical validation on embedded devices. We hope that Rep-MMB may serve as a strong baseline to inspire future edge-deployed HAR research.

It requires terrain classification for unmanned Mars Rover to identify the safe areas. The current deep learning-based semantic segmentation and object recognition suffer from a large number of parameters and long training time. In this paper, a lightweight segmentation framework called Mobile-DeepRFB is proposed for the Martian terrain classification. It improves from the DeepLabV3 <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$+$</tex-math> </inline-formula> by taking the MobileNetV3 as the backbone module to decrease the parameters and the Receptive Field Block (RFB) module to strengthen the feature extraction capability as well as to enlarge the receptive field. Experimental results on the NASA Mars terrain dataset AI4MARS show that the presented method reduces 94% about the parameter number and improves the mean pixel accuracy by 2% compared to the existing ResNet101 and Xception backbone networks. The deployment of this framework on a low-computing power embedded platform (NVIDIA Jetson Xavier) demonstrates its great potential to apply to Mars rovers. <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Note to Practitioners</i> —This paper was motivated by the problem of terrain classification of planetary rovers. Existing methods are typically based on semantic segmentation technology to recognize various terrains while suffering from the drawback of a large number of parameters. We propose a lightweight segmentation framework to address this issue. In particular, the lightweight backbone network is applied to significantly reduce the number of parameters. The receptive field module is substantially improved to enhance the feature extraction capability. Eventually, we deploy the framework on a low-computing platform. Experimental tests show that the framework can significantly reduce the number of network parameters and it can be used for planetary rovers with limited computational resources.

This work proposes a scalable gamma non-negative matrix network (SGNMN), which uses a Poisson randomized Gamma factor analysis to obtain the neurons of the first layer of a network. These neurons obey Gamma distribution whose shape parameter infers the neurons of the next layer of the network and their related weights. Upsampling the connection weights follows a Dirichlet distribution. Downsampling hidden units obey Gamma distribution. This work performs up-down sampling on each layer to learn the parameters of SGNMN. Experimental results indicate that the width and depth of SGNMN are closely related, and a reasonable network structure for accurately detecting brain fatigue through functional near-infrared spectroscopy can be obtained by considering network width, depth, and parameters.

Electroencephalogram (EEG)-based emotion recognition is of great significance for aiding in clinical diagnosis, treatment, nursing and rehabilitation. Current research on this issue mainly focuses on utilizing various network architectures with different types of neurons to exploit the temporal, spectral, or spatial information from EEG for classification. However, most studies fail to take full advantage of the useful Temporal-Spectral-Spatial (TSS) information of EEG signals. In this paper, we propose a novel and effective Fractal Spike Neural Network (Fractal-SNN) scheme, which can exploit the multi-scale TSS information from EEG, for emotion recognition. Our designed Fractal-SNN block in the proposed scheme approximately simulates the biological neural connection structures based on spiking neurons and a new fractal rule, allowing for the extraction of discriminative multi-scale TSS features from the signals. Our designed training technique, inverted drop-path, can enhance the generalization ability of the Fractal-SNN scheme. Sufficient experiments on four public benchmark databases, DREAMER, DEAP, SEED-IV and MPED, under the subject-dependent protocols demonstrate the superiority of the proposed scheme over the related advanced methods. In summary, the proposed scheme provides a promising solution for EEG-based emotion recognition.

Capturing finger joint angle information has important applications in human–computer interaction and hand function evaluation. In this paper, a novel wearable data glove is proposed for capturing finger joint angles. A sensing unit based on a grating strip and an optical detector is specially designed for finger joint angle measurement. To measure the angles of finger joints, 14 sensing units are arranged on the back of the glove. There is a sensing unit on the back of each of the middle phalange, proximal phalange, and metacarpal of each finger, except for the thumb. For the thumb, two sensing units are distributed on the back of the proximal phalange and metacarpal, respectively. Sensing unit response tests and calibration experiments are conducted to evaluate the feasibility of using the designed sensing unit for finger joint measurement. Experimental results of calibration show that the comprehensive precision of measuring the joint angle of a wooden finger model is 1.67%. Grasping tests and static digital gesture recognition experiments are conducted to evaluate the performance of the designed glove. We achieve a recognition accuracy of 99% by using the designed glove and a generalized regression neural network (GRNN). These preliminary experimental results indicate that the designed data glove is effective in capturing finger joint angles.

Self-supervised Human Activity Recognition (HAR) has been gradually gaining a lot of attention in ubiquitous computing community. Its current focus primarily lies in how to overcome the challenge of manually labeling complicated and intricate sensor data from wearable devices, which is often hard to interpret. However, current self-supervised algorithms encounter three main challenges: performance variability caused by data augmentations in contrastive learning paradigm, limitations imposed by traditional self-supervised models, and the computational load deployed on wearable devices by current mainstream transformer encoders. To comprehensively tackle these challenges, this paper proposes a powerful self-supervised approach for HAR from a novel perspective of denoising autoencoder, the first of its kind to explore how to reconstruct masked sensor data built on a commonly employed, well-designed, and computationally efficient fully convolutional network. Extensive experiments demonstrate that our proposed Masked Convolutional AutoEncoder (MaskCAE) outperforms current state-of-the-art algorithms in self-supervised, fully supervised, and semi-supervised situations without relying on any data augmentations, which fills the gap of masked sensor data modeling in HAR area. Visualization analyses show that our MaskCAE could effectively capture temporal semantics in time series sensor data, indicating its great potential in modeling abstracted sensor data. An actual implementation is evaluated on an embedded platform.

Currently, there is a growing interesting in emotion recognition. Representation of emotional states is a very challenging issue. Considering the calculation cost and generalization capability for practical application, a series of features which contain common time and frequency domain are extracted from physiological signals to represent different emotional states. To reduce feature dimensionality and improve the emotion recognition accuracy, a two-stage feature optimization method based on feature correlation analysis (FCA) and ReliefF algorithm is proposed to select critical features. Firstly, FCA is employed to analyze the redundancy between features, then ReliefF is adopted to analyze the correlation between features and categories, and the optimal feature subset is obtained using the two-stage FCA-ReliefF feature optimization method. Support vector machine is employed as the classifier to evaluate classification performance in this investigation. The effectiveness of the method which is proposed is validated by testing on two publicly available multimodal emotion datasets, Augsburg Biosignal Toolbox (AuBT) and Database for Emotion Analysis Using Physiological Signals (DEAP). Compared with recent similar reported studies, the method developed in this research for emotion recognition is stable and competitive, and its accuracy reaches to 98.40% (AuBT) and 92.34% (DEAP).