Dong Liu

To address the challenge of recognition and grasping in space caused by the random postures and sparse point clouds of small-sized flexible components. This letter proposed a two-stage grasping framework (PCS-SE) that combined point cloud segmentation and sampling evaluation. In the first stage, a small-sized-component point-cloud segmentation network (SPS-Net) was proposed, which utilized PointNet as a spatial encoding module and enhanced the feature extraction ability for small object through width residual attention. In the second stage, a fast sampling evaluation strategy (FSES) was proposed, which selected the optimal grasping posture through multiple constraint principles after uniform sampling. The tests were conducted on the Shapenet dataset, FCDatas dataset, and real environment. The results showed that the mIoU reached 84.80% and 91.75% on the two datasets. The average grasping success rate of 87% and 96% were achieved in handover and desktop scenarios. That proved the effectiveness and robustness of the framework proposed in this letter for grasping small-sized objects.

As the customization of new energy vehicles accelerates, the quality requirements for die-cast components are becoming increasingly stringent. Accurate and robust defect prediction has thus become critical for enhancing the competitiveness of manufacturing enterprises. In response to challenges such as small sample sizes, severe label imbalance, and strong inter-label correlations in die-casting quality prediction, this paper proposes CastMoE, a Casting-specific Mixture of Experts Framework for multi-label classification. CastMoE is designed to integrate multiple heterogeneous models with structurally diverse architectures through a Gated Mixture-of-Experts Fusion mechanism, which dynamically assigns label-wise softmax weights to adaptively aggregate predictions from different sub-models based on their performance. The framework incorporates a combination-based structure search strategy to explore and evaluate various model subsets, and selects the optimal fusion configuration based on cross-validated metrics. To ensure temporal consistency among variable-length time-series samples from the die-casting process, the Fast Dynamic Time Warping (FastDTW) algorithm is utilized. All sub-models and the fusion network are trained with Focal Loss to address class imbalance and improve the sensitivity to minority labels. During inference, a label-wise adaptive thresholding strategy is applied to convert probabilistic outputs into final binary decisions, with thresholds optimized using validation F1 scores to enhance recall and classification quality. Through the CastMoE framework, the optimal model combination identified by structure search consists of BiLSTM and CNN-LSTM-Attention. This configuration achieved a Micro-F1 score of 96.50% and a Micro-Recall of 95.83% under five-fold cross-validation, demonstrating the framework’s robustness, effectiveness, and industrial applicability for multi-label defect prediction in die-casting production.

Current lower-limb exoskeleton joint actuators lack flexibility and cannot adapt to the variable stiffness characteristics of human joints. They also suffer from large mass and low assistive efficiency. This letter designs a variable stiffness joint actuator for lower-limb exoskeletons based on shape memory alloys (SMA). The memory effect of SMA is utilized to adjust the preload force of springs, realizing changes in equivalent stiffness. Based on this, an active stiffness adjustment unit is constructed using SMA, and the units are distributed circumferentially to enhance the range of stiffness and torque output. A cooling system is incorporated to accelerate stiffness switching. Experimental results show that the designed actuator offers a broad stiffness adjustment range. Under experimental conditions, the maximum output torque can reach 50.8098 N·m, and the switching frequency between stiffnesses can reach 1.85 Hz. The actuator weighs only 293.6 g, achieving the design of actuator for lower-limb exoskeletons with lightweight, highly maneuverable and a wide stiffness range.

Aluminum alloy die-castings hold a critical strategic position in industries such as automotive manufacturing, aerospace, and rail transportation. In their visual defect detection, defects exhibiting complex backgrounds, multiscale variations, and irregular shapes challenge existing detection models. To tackle these challenges, this article proposes the Multi-Level CNN-Transformer Interaction Network (MCTNet), which integrates the convolutional neural network’s (CNNs) capacity for local detail analysis with the Transformer’s strength in capturing global information. To address the distinct characteristics of the two branches, we propose two modules: channel attention down-sampling (CADS) and cross global attention (CGA). These modules enable multilevel interaction and fusion, overcoming single-module limitations. We then design a super-resolution-enhanced feature fusion structure to enhance fine defect feature extraction. Finally, an adaptive feature weighting (AFW) and task-aligned detection head is proposed to synergize defect localization and classification. The proposed method outperforms the baseline model by improving mAP₅₀ and mAP ${}_{50\text {:}90}$ by 4.8% and 5.7%, respectively, while surpassing state-of-the-art CNN-based and Transformer-based models in accuracy.

Deep reinforcement learning has significant advantages in the field of robot skill learning, however it usually cannot make good use of experience. In this letter, a Bi-phase Episodic Memory Guided (BEMG) method for speeding up the learning process of DRL and solving the problem of sparse reward is proposed, which brings a large learning speed increasing of robotic operation skills acquisition through memory guidance. The method constructs the relationship between robot states at different moments through the guidance of episodic memory, and automatically generates reward function through memory backtracking. In the framework of proposed method, each iteration is divided into two phases. In first phase, the association between state and action in memory is used to guide the action decision-making of DRL, thereby reducing unnecessary exploration for learning. In second phase, the memory module automatically generates reward values based on the relationship between different states, instead of artificially designing reward functions. The experimental verifications are carried out around robot operation skill learning tasks with different DRL algorithms. Experimental results show the proposed method can effectively improve the learning speed of robot operation skills and avoid the dependence on complex reward function.