Abstract To address the growing demand for rapid and precise prediction of projectile trajectories in modern digital artillery systems, this paper presents an innovative spatio-temporal prediction model based on deep learning. Unlike traditional methods that rely on simplified assumptions and step-by-step numerical integration, the proposed model explicitly learns spatial relationships and temporal evolution of projectile motion through a combination of a Spatial Transformer Encoder, a Temporal GRU Encoder, and a Temporal GRU Decoder. A graph-based attention mechanism and nonlinear fusion in the latent space further enhance the model's ability to capture complex physical interactions influenced by environmental factors such as wind speed and firing angle. The model is initially trained using extensive high-fidelity simulation data and then fine-tuned with a small amount of real firing data, which improves its adaptability to real-world conditions and reduces the simulation-to-reality bias. Experimental results show that the model accurately predicts key ballistic parameters under varying conditions while achieving real-time inference performance suitable for engineering deployment. Its lightweight structure and strong compatibility facilitate easy integration with platforms such as Unreal Engine for virtual testing, tactical training, and intelligent weapon system development. Comparative evaluations and ablation studies verify that the proposed spatio-temporal prediction approach consistently outperforms existing methods in both accuracy and efficiency.

Abstract Buckling failure is the dominant failure mode in strut-based or thin-walled lattice structures, limiting their effectiveness in structural applications requiring high stiffness and strength. This paper investigates a novel buckling-based three-dimensional topology optimization method to generate buckling-resistant strut-based lattice structure for the first time. A MATLAB code for three-dimensional topology optimization, based on buckling objective/constraint, is developed inspired by a two-dimensional code available in literature. A single cell from a simple cubic lattice with square cross-sectional struts, a body-centred cubic lattice with circular struts, and a multi-cell 2 × 2 simple cubic lattice are investigated as reference structures. Through analysing their buckling modes, the buckling-prone members are actively optimized via topology optimization exercise to maximize buckling load factors, whilst constraining the volume fraction and stiffness of the structure. The results demonstrate that topology-optimized designs exhibit higher buckling load factors than the reference structure. Numerical validation of linear buckling analysis using ABAQUS software is also performed. Few parametric studies with varying cell sizes and relative densities are also discussed. This methodology can be easily extended to enhance the buckling resistance of any lattice structure without altering the relative density of the corresponding reference structure. Future directions include numerical post-buckling studies with experimental validation.

Abstract The topological and geometric information in B-rep models is essential for various applications, including feature recognition, design verification, and automatic tool path generation. In particular, representing surfaces using an analytic format offers advantages for utilizing geometric information such as surface type, axis of rotation, and radius. Models composed solely of NURBS surfaces often have limitations in extracting this type of information. To address this issue, this study proposes a method for identifying and replacing NURBS surfaces that can be represented analytically. In the first step, the feasibility of substituting a NURBS surface with an analytic surface is determined by examining the characteristics of its isocurves. In the second step, an equivalent analytic surface is created and used to replace the existing NURBS surface. To validate the proposed method, surface replacement experiments were conducted on multiple test cases. The results showed that, among 10 176 NURBS surfaces identified for possible analytic replacement across 37 B-rep models, 9 163 surfaces were successfully substituted in a stable manner. In addition, an experiment comparing feature recognition in the pre- and post-replacement models was conducted. As a result, none of the 12 design features were recognized in the pre-replacement model, whereas all were correctly detected afterward, confirming that the replacement process significantly enhances feature recognition accuracy.

Abstract This study proposes a battery-aware transformer network (BATNet) for accurate and robust state-of-charge estimation in lithium-ion batteries for electric vehicles (EVs). BATNet combines feedforward neural networks with an encoder-only transformer enhanced by local attention to focus on recent signal trends and an adaptive forgetting mechanism to suppress transient anomalies. Simulations were conducted on a high-fidelity Hyundai KONA EV model in MATLAB Simulink, encompassing four driving cycles and five ambient temperatures. Results show that BATNet significantly outperforms traditional models, including gated recurrent units, recurrent neural networks, and long short-term memory networks, achieving a root mean square error as low as 0.0436%. An ablation study further verifies the contributions of the proposed architectural components. BATNet effectively captures nonlinear battery dynamics and remains stable under varying thermal and load conditions, making it well-suited for integration into real-time battery management systems in EV applications.

Abstract The increasing complexity of autonomous drone swarm operations, particularly in adversarial air combat scenarios, presents significant challenges in Multi-Agent Reinforcement Learning (MARL). Existing approaches often face issues such as poor coordination, suboptimal convergence, and overestimation bias,undermining their applicability to high-stakes environments. We propose the Sequential Decision Distribution Soft Actor-Critic (SD2AC), a novel off-policy MARL framework for drone swarm coordination to address these challenges.SD2AC introduces three key innovations:(1) a sequential decision-making framework in which agents adaptively condition their decisions on the policies of preceding agents, enhancing coordination and reducing conflicts; (2) a distributional Q-value critic that models the full return distribution to mitigate overestimation bias and improve policy robustness; and (3) an adaptive twin value distribution learning mechanism that leverages dual critics to dynamically select conservative value estimates, ensuring stable learning under uncertainty.Rigorous evaluation in diverse environments, including Close Air Combat, MECA, and MAMuJoCo, demonstrates the superiority of SD2AC over state-of-the-art MARL methods in terms of reward optimization, convergence speed, and coordination efficiency. Ablation studies further validate the individual contributions of each component.

Abstract In the fault diagnosis of rolling element bearings (REBs), spall size is a typical indicator of fault severity. Conventionally, spall size estimation relies on expert-knowledge-based or data-driven approaches. Expert-knowledge-based approaches require accurate assumptions about spall-excited events, making them challenging to apply in field environments. In contrast, data-driven approaches often struggle with insufficient training data and limited generalization across various operating conditions. To address this challenge, this paper proposes a frequency-enhanced neural network (FENN) with a hybrid spall-size estimator (HSSE). The proposed FENN employs both one-dimensional convolution in the time domain and Fourier convolution on the frequency magnitude, while preserving phase information in the frequency domain to enhance frequency components that are associated with spall in REBs. The novel HSSE proposed here integrates a data-driven spall-size estimator and an expert-knowledge-guided spall-size estimator to capture spall entry and exit events between rolling elements and raceways. Model validation results, which analyzed both simulation and experimental data from roller and ball bearings, demonstrate that the proposed approach provides accurate predictions of spall size, even with limited training data. Additionally, it is confirmed that the proposed model identifies the mechanical frequencies associated with spall events, providing interpretable results from raw vibration signals without requiring further processing.

Abstract In space operations, the requirements for control precision, stability, and robustness of free-flying space manipulators (FFSMs) are critical for mission success. However, due to unknown system dynamics and external disturbances, control performance may degrade. To address these challenges, this paper proposes a new sliding mode control method to achieve fast and accurate tracking for space manipulators, even in the presence of lumped disturbances. In this method, the sliding surface design incorporates the value of the sliding surface at the previous sampling time, and the resulting surface is referred to as the Time-Delay Logarithmic Sliding Surface (TD-LnSS). This design enables the proposed Time-Delay Logarithmic Sliding Mode Control (TD-LnSMC) method to improve the robustness of the closed-loop systems without the need for any approximation or compensation mechanisms, and improve tracking accuracy without introducing complex structures. Numerical simulations are conducted to demonstrate the effectiveness of the proposed approach in improving tracking performance for free-flying space manipulators.

Abstract Advancements in information and communications technology and artificial intelligence have led to a new era of remote autonomous operation, enhancing the ability to recognize and manipulate objects in 3D space. Building upon the basic chemical autonomous experimental system, this study integrates mechanical specimens and measurement devices into an interoperable laboratory platform with a mobile manipulator. This study introduces a fully open-source hierarchical computational laboratory automation platform that enables researchers in remote locations to operate experimental devices and obtain results through high-level commands. A novel legacy equipment retrofitting methodology is proposed, in which legacy equipment is classified into three categories with their respective internet-of-things-based communication-enabling methodologies. The effectiveness of the proposed system is demonstrated through a detailed case study involving international remote laboratory education and experiment automation to measure the physical properties of additively manufactured specimens. The demonstration showed the SmartLab's successful conversion of an existing engineering laboratory to a remote and automated laboratory. An intercontinental remote material testing experimental class was successfully performed between Tanzania and Korea providing hands on experience to students. In addition, material property of additively manufactured parts is automatically experimented without human intervention where 90 specimens are automatically manufactured and tested. Results show that by using the SmartLab system, the efficiency of the experimental procedure can be significantly increased by processing specimens 2.5 times more than human researchers per day. The experimental results obtained with SmartLab is well aligned with previous research demonstrating the potential for automated research.

Abstract With the rapid advancement of artificial intelligence, research on deep neural networks for estimating gaze from facial images captured by cameras has grown significantly. However, gaze estimation remains challenging due to the diversity of facial features, variability in head posture, partial occlusion, varying lighting conditions, eye image quality, complex eye movements, and a lack of large-scale, diverse data sets. Particularly, estimating a driver's gaze while driving a vehicle is difficult due to occlusion of the eyes when checking mirrors or turning the head to view blind spots and surroundings. To address these issues, we developed an appearance-based gaze estimation neural network based on attention, called GazeSymCAT (symmetric cross-attention transformer for gaze estimation). The proposed GazeSymCAT obtains contextual information that reflects the relationships among the extracted facial and eye image features through an encoder–decoder structure composed of self- and cross-attention layers of the transformer. This structure is followed by adding fully connected layers that estimate the gaze direction. The model proposed in this study was tested on the ETH-XGaze data set, a large-scale, state-of-the-art (SOTA) data set that includes extreme head postures, achieving SOTA performance. Additionally, it demonstrated SOTA-comparable performance on existing general gaze estimation-related data sets, MPIIFaceGaze and EYEDIAP.

Abstract Exoskeleton robots mimic the structure of the human body and are physically connected to the wearer through attachment parts such as cuffs or straps. Nevertheless, misalignment between the human body and the robot can occur due to improper wearing of the exoskeleton, the elasticity of human skin, and the geometric complexity of human joint movements. Such misalignment increases unnecessary physical interaction forces, causing discomfort and pain to the wearer. Therefore, these interaction forces should be considered when designing exoskeleton robots to ensure wearability. In this study, we propose a method for estimating human-exoskeleton interaction forces through posture prediction. The human-robot connection is modeled as an elastic element, and posture is predicted using an energy optimization algorithm. The predicted posture is then used to calculate the interaction forces. Since this method considers only the physical characteristics of the exoskeleton robot and wearer, it enables objective evaluation of the robot without the need to manufacture actual prototypes. We performed quantitative experiments using prototypes of a sensor-equipped dummy and an exoskeleton to confirm the effectiveness of the modeling method. Consequently, the proposed method is expected to reduce the time and costs associated with developing exoskeleton robots and obviate the need for human subject testing.