The current underwater tracking control methods based on optical image priori data have the problems of low reliability, high training cost, and poor adaptability to different environments. This paper proposes an underwater dynamic tracking control method based on complex environment simulation and deep reinforcement learning, which can solve out the optimal real-time yaw angle and propelling force of Autonomous Underwater Vehicle (AUV) for target tracking. Firstly, we design an underwater simulation environment generator which supports parallel training. This generator is able to provide a large number of complex stochastic environments containing random current effects and sound speed uncertainty for the deep reinforcement learning training. Then, a control algorithm based on Attention Conv-LSTM (ACL) and Soft Actor-Critic (SAC) networks is proposed, which has fast convergence speed and adaptation capability to complex environment. In addition, we design a composite reward function to avoid premature local optimal solutions. Experimental results show that the tracking detachment and collision probabilities under the simulation environment with random ocean currents and acoustic velocity uncertainty interference are as low as 0.12% and 0.82%, respectively. The AUV testing under real environment presents good dynamic tracking performance both for linearly and randomly moving targets.

Semi-supervised semantic segmentation relieves the reliance on large-scale labeled data by leveraging unlabeled data. Recent semi-supervised semantic segmentation approaches mainly resort to pseudo-labeling methods to exploit unlabeled data. However, unreliable pseudo-labeling can undermine the semi-supervision processes. In this paper, we propose an algorithm called Multi-Level Label Correction (MLLC), which aims to use graph neural networks to capture structural relationships in Semantic-Level Graphs (SLGs) and Class-Level Graphs (CLGs) to rectify erroneous pseudo-labels. Specifically, SLGs represent semantic affinities between pairs of pixel features, and CLGs describe classification consistencies between pairs of pixel labels. With the support of proximate pattern information from graphs, MLLC can rectify incorrectly predicted pseudo-labels and can facilitate discriminative feature representations. We design an end-to-end network to train and perform this effective label corrections mechanism. Experiments demonstrate that MLLC can significantly improve supervised baselines and outperforms state-of-the-art approaches in different scenarios on Cityscapes and PASCAL VOC 2012 datasets. Specifically, MLLC improves the supervised baseline by at least 5% and 2% with DeepLabV2 and DeepLabV3+ respectively under different partition protocols.

Active rehabilitation can use electro-encephalogram (EEG) signals to identify the patient's left and right leg movement intentions for rehabilitation training, which helps stroke patients recover better and faster. However, the lower limb rehabilitation robot based on EEG has low recognition accuracy so far. A classification method based on EEG signals of motor imagery is proposed to enable patients to accurately control their left and right legs. Firstly, aiming at the unstable characteristics of EEG signals, an experimental protocol of motor imagery was constructed based on multijoint trajectory planning motion of left and right legs. The signals with time-frequency analysis and event-related desynchrony/synchronization (ERD/S) analysis have proved the reliability and validity of the collected EEG signals. Then, the EEG signals generated by the protocol were preprocessed and common space pattern (CSP) was used to extract their features. Support vector machine (SVM) and linear discriminant analysis (LDA) are adapted and their accuracy of classification results are compared. Finally, on the basis of the proposed classifier with excellent performance, the classifier is used in the active control strategy of the lower limb rehabilitation robot, and the average accuracy of the left leg and right leg controlled by two healthy volunteers was 95.7%, 97.3%, 94.9%, and 94.6%, respectively, by using the ten-fold cross test. This research provides a good theoretical basis for the realization and application of brain-computer interfaces in rehabilitation training.

Vision-based tactile sensor (VBTS) has gained great traction within embodied tactile community in recent years. However, to reproduce a high-performance one, the contact elastomer is the key and its processing method largely depends on equipment and sophisticated operating skills, which sets obstacles for labs lacking hardware expertise and financial support. In this work, we present a low-cost and easy-to-fabricate processing scheme for the contact elastomer. Our scheme adopts a novel structure of reflective membrane by chemically bonding two different reflective layers, and utilizes laser stencils and bearing balls to make marker layers with high dimension consistency. The whole manufacturing process is completed manually and costs only 4.34 USD per elastomer. Several experiments are conduct to evaluate the performance of the elastomer. Results show that the finished elastomer possesses microlevel tactile mapping resolution and reliable force perception, as well as favorable durability. Additionally, the versatility of the method is shown by successfully designing and making other two different VBTSs. Compared to other advanced manufacturing methods, our proposed method not only enables the finished elastomer with comparable tactile performance, but also achieves a trade-off between simplicity and high sensitivity, making high-performance VBTSs more accessible to robotic community.