A Labeled GM-PHD Filter with Target Amplitude Information

Shunyi Luo, Linfeng Xu, and Peijie Yang

Abstract: Radar multi-target tracking faces significant challenges in complex environments characterized by heavy clutter and high target maneuvers. The traditional Gaussian mixture probability hypothesis density (GM-PHD) filter offers efficient multiple target tracking through bypassing data association, but it struggles to achieve target identities and often fails in scenarios with dense clutter or closely spaced targets. To address these issues, this paper proposes an improved GM-PHD filter by incorporating amplitude information into the weight update process to improve discrimination between targets and clutter in data association. In addition, a threshold-based clustering strategy based on kinematic similarity is introduced to enhance the separation of closely spaced targets and maintain label consistency. Simulation results show that the proposed method achieves improved identity recognition, tracking accuracy, and trajectory continuity in cluttered and dense-target environments compared with conventional GM-PHD and labeled GM-PHD filters.

Track-Before-Detect with Stereo Microphone Arrays

Jiazheng Wang, Mohammad Fard, Amirali Khodadadian Gostar, and Reza Hoseinnezhad

Abstract: Situational awareness of autonomous vehicles in urban environments heavily relies on line-of-sight sensors such as cameras, lidar, and radar. However, these visual sensors may fail to detect targets when they are occluded. In contrast, audio signals have stronger penetration capabilities, enabling human to perceive the direction-of-arrival (DoA) of sound. This directional information can also be extracted using beamforming techniques applied to multi-channel microphone arrays. In this work, we utilize a stereo microphone array to track the 2D position of a single vehicle. Such a minimal array configuration is susceptible to ambient noise and vehicle sounds leading to high variance in position estimation. To address this issue, we propose a track-before-detect (TBD) based particle filter algorithm that incorporates stereo microphone measurements for 2D vehicle tracking. The proposed method is validated using a custom dataset that includes scenarios where vehicles pass at various distances from the microphone array. Using bird's-eye view camera footage as ground truth, experimental results demonstrate that our method achieves accurate and low-variance 2D sound source tracking.

Covariance Intersection-Based Distributed Fusion for Heterogeneous Sensor Networks with Multi-Target Tracking in Clutter

Jiazheng Li, Ye Yuan, and Yuhuan Xiong

Abstract: With increasing demands on multi-target tracking (MTT) accuracy in complex environments, this paper proposes a distributed heterogeneous sensor network based on the covariance intersection (CI) fusion criterion, integrating radar and infrared (IR) sensors. First, we establish target motion and sensor measurement models, and unify the measurement dimensions and formats for cross-sensor compatibility. Unlike single-target tracking, MTT under cluttered conditions requires resolving measurement-to-track associations. To address this, the joint probabilistic data association (JPDA) algorithm is implemented locally at each sensor node for target state estimation and update. To fuse these local estimates without assuming inter-node correlation, the CI criterion is employed. Monte Carlo simulation results validate the effectiveness and accuracy of the proposed method in cluttered MTT scenarios.

MD-IPDA Combined with GP for ETT

Taek Lyul Song, Hoseok Sul, Dongsheng Yang, Yunfei Guo, and Jee Woong Choi

Abstract: Target tracking using high-resolution sensors often suffers from multiple detections generated by a single target, leading to a typical multiple detection (MD) problem that requires distinguishing between true target detections and clutter. To address this issue, a Multiple Detection-Integrated Probabilistic Data Association (MD-IPDA) algorithm is combined with a Gaussian Process (GP) for robust single-target tracking in cluttered environments. The MD-IPDA framework utilizes the target existence probability (TEP) as a track scoring mechanism to ensure reliable track maintenance. Meanwhile, GP is employed to construct measurement model for multiple scattering points, based on predefined contour points determined by the target's shape. The performance of the proposed extended target tracking (ETT) method is validated through simulation experiments and further assessed using the information reduction factor (IRF), allowing direct comparison with existing ETT algorithms.

Auxiliary-Guided Subspace-Aligned Tensor Completion for Radar Data Recovery

Lijun Liang, Tiancheng Li, Yan Song, and Pablo Chamoso

Abstract: To address the critical data loss problem in multichannel radar systems in complex electromagnetic environments, we propose a subspace-aligned tensor completion framework that effectively incorporates auxiliary information. Conventional matrix models flatten high-dimensional data into two-dimensional representations, disregarding the intrinsic low-rank structures of multidimensional data. Furthermore, existing completion algorithms predominantly depend on the primary radar's observations, leading to information scarcity that compromises reconstruction accuracy under severe data loss. To overcome these limitations, we develop a Tucker decomposition-based joint optimization framework that simultaneously enforces low-rankness of the three-dimensional observation tensor via weighted nuclear norm regularization, integrates preprocessed auxiliary radar tensors as reconstruction priors, and establishes a factor matrix alignment mechanism to enhance structural consistency. Theoretically, we establish an upper bound on the recovery error with four components: low-rank optimization error, primary auxiliary tensor deviation, subspace alignment error, and observation noise. Simulations verify that our method consistently outperforms both standard low-rank completion techniques and naive fusion approaches across various observation ratios. Notably, the empirical error strictly complies with the derived theoretical bound, validating the efficacy of our integrated framework, which combines auxiliary information guidance with subspace alignment.

Joint Data Transmission Mode Selection and Power Allocation for Multi-Target Tracking in Radar Network

Hao Jiao, Peng Zhang, Junkun Yan, Chang Gao, Hongwei Liu, and Yun Zhu

Abstract: To address the issues of high communication overhead and information degradation in conventional radar networks (RNs) relying on static fusion architectures, this paper proposes a joint data transmission mode selection and power allocation scheme. This scheme aims to maximize multi-target tracking performance under constraints on communication bandwidth at the fusion center and node transmission power. Firstly, we derives a Bayesian Cramér-Rao Lower Bound of mixed fusion (BCRLB-MF) to quantify the joint impact of power allocation and data transmission mode selection on tracking accuracy under detection uncertainty conditions. Secondly, based on the normalized BCRLB-MF, a mixed-integer nonlinear programming model is established to achieve coordinated optimization of data transmission modes and power. Lastly, a fast solver is designed using the L2-box alternating direction method of multipliers (ADMM) algorithm. Simulation results indicate that the proposed scheme can effectively dynamically balance resource utilization and enhance the tracking performance of RN for targets.

Scanning Sector Selection Scheme in Multi-radar System under Dynamic Interference Environment

Yuan Tang, Peng Zhang, Junkun Yan, Rongrong Wang, Tianyi Jia, Bo Jiu, and Hongwei Liu

Abstract: This paper proposes a scanning sector selection scheme in multi-radar systems to collaboratively enhance target area coverage in the presence of dynamic jamming. Specifically, a scanning sector selection method is developed within a reinforcement learning (RL) framework, modeled as a Markov Decision Process (MDP). By employing a Deep Q-Network (DQN), the system learns an optimal strategy to guide the radar in selecting appropriate scanning sectors at each time step. Simulation results demonstrate that the proposed method effectively enables the radar to adapt its scanning sectors in response to dynamic interference, thereby significantly improving target area coverage.

Cartesian and Polar Formulations of the NCV Model in Radar Tracking

Mahendra Mallic, Linfeng Xu, and Xiaoqing Tian

Abstract: The nearly constant velocity (NCV) model in 2-D is widely used in radar tracking using the range and azimuth measurements. The NCV model is linear and the measurement model is nonlinear in the Cartesian coordinates. In contrast, if a polar state vector is used with components range, azimuth, range-rate, and azimuth-rate, then the dynamic model for the NCV motion becomes nonlinear, and the measurement model is linear. Some existing publications adopt a polar state vector for the NCV motion but assume a linear dynamic model. This formulation actually represents a motion with a nearly constant range-rate and nearly constant azimuth-rate, rather than the NCV motion. In this paper, we examine the accuracy of this approximate model using a maritime target tracking scenario. We use a cubature Kalman filter for the filtering problem and calculate the root mean square errors and posterior Cramér-Rao lower bound for filter evaluation. Monte Carlo simulation results demonstrate that the approximate linear model using a polar state vector yields less accurate state estimates compared with the linear dynamic model and nonlinear measurement model using a Cartesian state vector.

Harmonic-Model-Based Wiggling Correction for Robust Indirect Time-of-Flight Imaging

Jaehyeok Kim and Jaehwa Jeong

Abstract: Three-dimensional spatial information is widely used across many fields, and the need for measurement techniques that provide high accuracy, precision, and robustness to environmental variation is increasingly critical. This paper addresses wiggling caused by harmonics of the modulation waveform in Indirect Time-of-Flight (iToF) and proposes a method that simultaneously corrects distance and amplitude without hardware modification. We employ a harmonic model designed for real signals and estimate residual harmonic components via prior calibration. We then use Particle Swarm Optimization (PSO) to estimate the model parameters from the calibration data. In the online stage, we apply Newton's method to a single frame, thereby removing wiggling in both distance and amplitude. Experiments at modulation frequencies of 12, 18, and 24 MHz confirm consistency between the model and measurements. At 12 MHz the distance Mean Absolute Error (MAE) decreases from 7.25 cm to 0.85 cm after a single iteration and converges to 0.82 cm after two to three additional iterations, demonstrating convergence speed and reproducibility. Because only four parameters per pixel are stored, memory demand is lower than prior-data-based corrections, and convergence within two or three iterations per frame supports practicality. The approach is also expected to be effective when combined with Multi-Path Interference (MPI) mitigation.

A Multi-rotor Drone Trajectory Planning Method Based on Energy Optimization

Jiangting Wang, Weifeng Liu, and Linqing He

Abstract: In the optimization of complex environmental trajectory, traditional methods are mostly based on time optimization, ignoring the constraints of energy consumption on flight endurance, which makes it difficult to balance endurance and mission efficiency in actual applications. In response to this, this paper proposes a drone trajectory planning method based on energy optimization: in the framework of traditional time optimization objective function, refine energy consumption modeling is integrated, adaptive weight adjustment strategies are innovatively introduced, and the game relationship between flight time and energy consumption is dynamically balanced. The experiment was developed based on the ROS simulation environment, and multiple sets of differentiated time-energy weight ratio parameters were set for comparison and testing, and the correlation curve between time and energy consumption was visualized with the help of Matplotlib. The results show that while ensuring the rationality of flight time, this method can effectively reduce the total energy consumption of drones, providing an optimization paradigm that takes into account timeliness and economy for drone mission planning in actual scenarios such as logistics distribution, patrol and monitoring, and expanding the application ideas of drone trajectory planning in complex environments.

Adaptive Tracking Control for Wheeled Mobile Robots Using Sliding Mode Control and Disturbance Observer

Thi Mai Do; Nam Hoai Nguyen; Phuoc Doan Nguyen; Hoa Dinh Nguyen

Abstract: This research introduces an adaptive tracking controller for the kinetic loop to solve the trajectory tracking problem for non-holonomic wheeled mobile robots subject to undesired factors including parameters uncertainty and unknown external disturbance. The proposed controller is based on the integral sliding mode control and disturbance observer. It is combined with an existing adaptive controller for the kinematic loop. The proof for stability of the overall system is performed based on the Lyapunov theory. Simulations with comparisons are carried out to prove the effectiveness of the designed controller. Based on the simulation results, it can be concluded that the proposed control scheme has high capacity of anti-disturbance and anti-uncertainty, significantly reduces the chattering magnitude and improves the tracking performance.

A Low-Code Methodology for Developing AI Kiosks: a Case Study with the DIZEST Platform

SunMin Moon, Jangwon Gim, Chaerin Kim, Yeeun Kim, YoungJoo Kim, and Kang Choi

Abstract: This paper presents a comprehensive study on enhancing kiosk systems through a low-code architecture, with a focus on AI-based implementations. Modern kiosk systems are confronted with significant challenges, including a lack of integration, structural rigidity, performance bottlenecks, and the absence of collaborative frameworks. To overcome these limitations, we propose a DIZEST-based approach methodology, a specialized low-code platform that enables intuitive workflow design and seamless AI integration. Through a comparative analysis with existing platforms, including Jupyter Notebook, ComfyUI, and Orange3, we demonstrate that DIZEST delivers superior performance across key evaluation criteria. Our photo kiosk case study further validates the effectiveness of this approach in improving interoperability, enhancing user experience, and increasing deployment flexibility.

Low-Overlap Point Cloud Registration via Pre-Trained Implicit Autoencoders

Francisco Molina and Martin Adams

Abstract: Point cloud registration is a fundamental task in computer vision and robotics, aimed at estimating the rigid transformation that aligns two 3D scans of the same scene. In real-world scenarios, this task becomes challenging due to low overlap between scans, which can result from occlusions, viewpoint changes, or inherent sensor limitations. Recent deeplearning-based methods, such as PREDATOR, CoFiNet, and GeoTransformer, have made significant progress on this problem by incorporating attention mechanisms, spatial hierarchies, and geometric priors. However, all these approaches share a common limitation: they learn point-wise descriptors from scratch without explicit guidance about global geometry. This article studies the feasibility of applying the Implicit AutoEncoders (IAE) approach as an unsupervised pre-training stage to obtain more robust geometric representations. By training an encoder to reconstruct implicit surfaces from point clouds, descriptors capturing both local and global information are obtained. The results show that, contrary to expectations, pre-training does not improve and even degrades the quality of correspondences and alignment accuracy. A thorough analysis leads to six reasons why IAEs are not suitable candidates for pre-training registration models.

Behavior Recognition Based on Multimodal Information of Unmanned Aerial Vehicle Swarm of DeepSeek V3

Xuanyuan Shi, Weifeng Liu, and Wanyu Li

Abstract: This study proposes a method for behavior recognition of UAV swarms based on multimodal information using the DeepSeek-V3 large model, aiming to address the issues of accuracy and robustness in UAV swarm behavior recognition in complex dynamic scenarios. By fusing static attributes with dynamic radar measurement data, a unified input representation framework is constructed, and the self-attention mechanism of DeepSeek-V3 is used to model spatiotemporal dependencies. Innovatively, the DeepSeek-V3 large model is proposed for recognizing the behavior of UAV swarms, and a prompt optimization strategy is adopted instead of traditional fine-tuning methods to guide the model to output results of behavior classification, formation matching, and anomaly detection. Experiments show that in a swarm composed of 4 Matrice 300 RTK UAVs, the average accuracy of this method in behavior classification tasks reaches 89.3%, the formation pattern matching degree is 84.7%, the comprehensive F1 score for anomaly detection is 0.84, and the mean absolute error (MAE) for trajectory prediction is 1.15 meters, providing an efficient solution for intelligent collaborative control of UAV swarms.

Multi-Agent Planning for Pursuing Multiple Objects via Proximal Policy Optimisation

Hoa Van Nguyen, Diluka Moratuwage, Tran Thien Dat Nguyen, Changbeom Shim, and Ba-Ngu Vo

Abstract: This paper addresses the challenge of efficient multi-agent planning for capturing multiple dynamic objects using deep reinforcement learning (DRL). We propose a scalable solution based on the multi-agent proximal policy optimisation (MAPPO) with three key innovations. First, we employ a dynamic state representation that adapts to a variable number of objects, ensuring scalability. Second, we introduce an optimal task allocation mechanism using the Hungarian algorithm to provide a cooperative signal for agent assignments. Finally, this is coupled with a multi-component reward function designed to incentivise both rapid and coordinated pursuit. Experiments demonstrate that our method enables agents to efficiently and collaboratively pursue all objects across various scenarios, significantly outperforming strong multi-agent reinforcement learning (MARL) baselines in both coordination and efficiency. The results highlight a practical pathway toward developing effective, adaptive multi-agent systems for complex, real-world tasks.

Design and Implementation of an AI-Powered Robotic System for Laser Cutting Application in Nuclear Power Plant

Rajat Jayantilal Rathod, Himanshu K. Patel, Priyank Jayantilal Rathod, and Meet Parmar

Abstract: This paper describes the functional design and implementation of an AI-powered robotic system to perform the critical laser cut operation in a nuclear power plant. There are a number of laser-cut applications for nuclear power plant operations, but our prime focus is to develop a robotic system for nuclear reactor maintenance, specifically en masse coolant channel replacement work in a pressurised heavy water reactor. Earlier, laser technology was used in various manufacturing industries and some non-working nuclear reactor core operations. Over the last few years, it has been observed that their applications are now being extended to nuclear power plant maintenance, decommissioning, and other plant operations. Recent developments have been made in robotic systems as well, still integration of laser cutting applications still needs integrated autonomous robotic-based solutions for nuclear power plants maintenance work. The traditional system still uses large semi-automated power manipulator systems that are operated manually by radiation workers. The paper has outlined the gap and implemented an Artificial Intelligence and Machine Learning algorithms-based robotic system that can be used for precise movement, path planning, and decision-making capabilities. This research will provide the design and implementation of an AI-powered robotic system focused on laser cutting applications to be used for nuclear power plant maintenance work in a hazardous nuclear environment. Moreover, it will also help to demonstrate experimental capability to collect the data for AI-powered nuclear waste management and disposal technology.

SafeDoc Agent For Facility Safety Report Evaluation

Jaehyun Kim, Gayeong Kim, Jaemoon Lee, and Woosuk Choi

Abstract: Following two major large-scale structural collapse incidents in the 1990s, there emerged an urgent need to establish a systematic framework to ensure the safety of infrastructure and buildings. As a result, the detailed safety inspection system was introduced, and since 2008, the scope of evaluation has expanded to include comprehensive safety assessments. Although reports are generally prepared and submitted in accordance with official evaluation criteria, there are frequent instances where the content or formatting fails to align with the required standards, yet such issues often go unnoticed during submission. To mitigate these risks, the necessity of introducing AI-powered automated verification systems for report evaluation has become increasingly apparent. In this study, we propose an evaluation-guided large language model (LLM) architecture designed to verify whether safety inspection reports are properly structured and complete prior to submission. The proposed method achieved approximately 55% improvement in detecting omitted elements compared to existing approaches, highlighting its effectiveness and demonstrating its strong potential for future application in fine-grained report evaluation through specialized sLLM.

Robust Small Methane Plume Segmentation in Satellite Imagery

Khai Duc Minh Tran, Hoa Van Nguyen, Aimuni Binti Muhammad Rawi, Hareeshrao Athinarayanarao, and Ba-Ngu Vo

Abstract: This paper tackles the challenging problem of detecting methane plumes, a potent greenhouse gas, using Sentinel-2 imagery. This contributes to the mitigation of rapid climate change. We propose a novel deep learning solution based on U-Net with a ResNet34 encoder, integrating dual spectral enhancement techniques (Varon ratio and Sanchez regression) to optimise input features for heightened sensitivity. A key achievement is the ability to detect small plumes down to 400 m² (i.e., for a single pixel at 20 m resolution), surpassing traditional methods limited to larger plumes. Experiments show our approach achieves a 78.39% F1-score on the validation set, demonstrating superior performance in sensitivity and precision over existing remote sensing techniques for automated methane monitoring, especially for small plumes.

3D Ground Truth Reconstruction from Multi-Camera Annotations Using UKF

Linh Van Ma, Unse Fatima, Tepy Sokun Chriv, Haroon Imran, and Moongu Jeon

Abstract: Accurate 3D ground truth estimation is critical for applications such as autonomous navigation, surveillance, and robotics. This paper introduces a novel method that uses an Unscented Kalman Filter (UKF) to fuse 2D bounding box or pose keypoint ground truth annotations from multiple calibrated cameras into accurate 3D ground truth. By leveraging human-annotated ground-truth 2D, our proposed method, a multi-camera single-object tracking algorithm, transforms 2D image coordinates into robust 3D world coordinates through homography-based projection and UKF-based fusion. Our proposed algorithm processes multi-view data to estimate object positions and shapes while effectively handling challenges such as occlusion. We evaluate our method on the CMC, Wildtrack, and Panoptic datasets, demonstrating high accuracy in 3D localization compared to the available 3D ground truth. Unlike existing approaches that provide only ground-plane information, our method also outputs the full 3D shape of each object. Additionally, the algorithm offers a scalable and fully automatic solution for multi-camera systems using only 2D image annotations.

Performance Evaluation of Large Language Models for Humidity Forecasting in Shipyards

Hyunji Kim, JaeSum Lee, Daekyeom Lee, Jaemoon Lee, and Kang Choi

Abstract: Shipyard painting operations are highly sensitive to ambient humidity, necessitating accurate forecasting for effective process control. Traditional forecasting methods, such as recurrent neural networks (RNNs) and long short-term memory, typically perform well under stable conditions but struggle with fluctuating environmental dynamics. This work studies large language models (LLMs), specifically Gemma variants, for hourly humidity forecasting using real-time sensor data from multiple shipyard zones. The experimental results show that Gemma-12B significantly improves forecasting accuracy under more variable autumn conditions, while the RNN-based method generally outperforms all Gemma variants during stable early-summer conditions. Specifically, Gemma-12B achieves substantially lower mean squared error, demonstrating its superior capability in capturing non-linear temporal dependencies inherent in dynamic humidity patterns. These results indicate the robustness of LLMs in such contexts and their promising applicability to industrial forecasting tasks.

Doppler-Assisted Automotive Radar Data Clustering Under Micro-Doppler Effect

Jing Zeng, Aranee Balachandran, and Ratnasingham Tharmarasa

Abstract: Modern radars have high resolution and produce multiple measurements from each target in a measurement cycle. Clustering radar measurements is a significant and challenging problem. Utilizing Doppler information is an effective method to improve clustering accuracy, while the presence of the micro-Doppler effect could significantly affect the clustering accuracy. In this paper, a two-step clustering algorithm to group micro-Doppler measurements with the measurements from the target body is proposed by accounting for the tire position constraints and maximizing range rate likelihood. Matlab simulated results demonstrate that our proposed method significantly reduces computation time by 99% while improving overall clustering accuracy by 2% compared to our previous clustering approach.

CADA: Confidence-Aware Dynamic Association for Multi-Object Tracking in Complex Environments

Hao Wang, Yun Zhu, Shuang Liang, Wanying Zhang, and Li Zhao

Abstract: Existing online visual Multi-Object Tracking (MOT) in complex scenes requires the effective fusion of different features. However, existing methods often employ static feature fusion strategies, which adapt to dynamic challenges with difficulty such as occlusions or erratic movements, leading to frequent identity switches. To address this limitation, we propose a novel Confidence-Aware Dynamic Association strategy built upon the Generalized Labeled Multi-Bernoulli (GLMB) filtering framework. The core of our approach is a dynamic fusion module that adaptively adjusts the weights between motion, appearance and spatial costs based on tracking confidence and motion predictability. We further introduce an Enhanced Re-activation Mechanism to reliably recover targets after prolonged occlusions. Evaluations on the MOT16 and MOT20 datasets validate the effectiveness of our approach. Our tracker achieves a significant performance boost over the baseline GLMB method, demonstrating superior identity preservation capabilities and a marked reduction in identity switches.

Multi-Sensor Multi-Object Trajectory Estimation with Lineage Inference

Tran Thien Dat Nguyen, Hoa Van Nguyen, and Changbeom Shim

Abstract: In this paper, we propose an algorithm to estimate object trajectories and lineage information in the presence of spawning events using multiple sensors. For efficient computation, we approximate the predicted multi-object density with object spawning by a generalized labeled multi-Bernoulli (GLMB). To handle multi-sensor observation, we apply the multi-sensor GLMB filter update using Gibbs sampling with an approximate proposal to efficiently sample for significant GLMB components. The experimental results with a 3D tracking scenario demonstrate accurate estimation performance both in terms of object states and lineage information.

A Lightning-Fast Sensor Control Algorithm for Swarm Tracking

Tharani Rajapaksha, Amirali Khodadadian Gostar, Aidan Blair, and Reza Hoseinnezhad

Abstract: This paper introduces a novel sensor control approach tailored for tracking swarms of autonomous agents, such as Unmanned Aerial Vehicles (UAVs). The aim of this proposed method is to track swarms in the presence of a controllable sensor with limited sensor resolution, improving the computational time. A key contribution of this work is the incorporation of swarm information into the formulation of the sensor control problem, thereby improving the computational performance in real-time execution. To address this, the Sequential Monte Carlo (SMC) swarm tracking-based sensor control method is proposed, along with a new likelihood function that integrates swarm characteristics into the pseudo-update step. Our proposed method combines both the SMC-based swarm tracking filter with the Labeled Multi-Bernoulli filter (LMB), enabling efficient and accurate estimation of swarm cardinality, overall dynamics, shape-related information, and individual target trajectories using controllable sensors. Numerical evaluations demonstrate that the proposed approach achieves a significantly improved computational efficiency compared to the LMB filter-based sensor control method, while providing marginal improvements in tracking accuracy.

Real-Time Intrusion Detection of Ghost Attacks in Distributed Multi-Object Tracking Without Offline Pre-Training

Vahid Ghorbani, Amirali Khodadadian Gostar, Amir Ghorbani, Zahir Tari, Aidan Blair, and Reza Hoseinnezhad

Abstract: Leveraging the recently proposed Average Likelihood for Attack Resilient Multi-object (ALARM) filtering method, this study develops an intrusion detection scheme to counter stealthy, measurement-oriented adversarial interventions known as ghost attacks in distributed multi-object tracking systems. In this scheme, each node executes both the ALARM filter and a conventional multi-object filter in parallel; a mismatch in the cardinalities of their respective multi-object densities signals a potential attack. Once an anomaly is detected, the estimated ghost cardinality is obtained by comparing the two filter outputs. Estimated ghost counts are then evaluated against ground-truth using the ℓ2 error (Euclidean norm of residuals), mean squared error, root mean squared error, mean absolute error, and coefficient of determination (𝑅2). This approach provides a unified framework for both detecting stealthy adversarial activity and estimating the number of false targets in a distributed multi-object tracking environment.

Finite-Time Control Barrier Functions for Safety-Critical Control of Discrete Systems with Application to Robot Navigation

Liang Gao, Xiongchang Li, Zongyang Lv, and Yuhu Wu

Abstract: This paper proposes a unified finite-time control framework for safety-critical control of nonlinear discrete-time systems. First, a novel concept of finite-time discrete control barrier function (FT-DCBF) is introduced to ensure finite-time safety recovery, along with the definition of a strong safety set whose forward invariance is rigorously proven. Subsequently, the discrete control Lyapunov function (DCLF) is integrated with the proposed FT-DCBF within a quadratic programming (QP) framework to construct a unified finite-time safety controller. To demonstrate the effectiveness of the proposed method, it is applied to obstacle avoidance navigation for wheeled mobile robots (WMRs) in complex environments. Under disturbance conditions, the proposed control strategy significantly enhances the navigation safety and robustness of the WMRs. Finally, simulation results validate the efficacy of the proposed control framework.

On Finite Bayesian Potential Games

Shuting Le, Yuhu Wu, Benlian Xu, and Zhengtian Wu

Abstract: In order to effectively solve the equilibrium problem of Bayesian game with finite policy set, the transformation of incomplete information and the establishment of Bayesian potential equation in Bayesian game are proposed in our previous reslut. On this basis, combined with the characteristics that human utility in Bayesian game is not affected by the same type, this paper establishes an improved potential equation and designs the corresponding improved algorithm which can reduce the computational complexity. Secondly, using the characteristics of matrix semi tensor product, an alternative method to solve the potential function of Bayesian potential game is given. Finally, in the framework of Bayesian game, the theoretical results are applied to the deployment of Network Functions Virtualization (NFV) system service function chain with incomplete information, and then provide the best resource allocation scheme and joint resource optimization algorithm in the sense of game equilibrium.

Discrete Position Estimation of Mini 4WD AI Vehicle on a Course by Dempster-Shafer Theory Based State Space Modeling

Norikazu Ikoma

Abstract: In a context of "Mini 4WD AI", which equips a microcomputer board to control single DC motor powered by two AA batteries to drive its four wheels simultaneously, vehicle's position estimation over the circuit course is crucial to win "Mini 4WD AI" tournament held by an academic community in Japan. Simplified modeling is mandatory to cope with to the actual Mini 4WD AI vehicle system due to high-speed driving of the vehicle and limited data transfer rate between on-board microcomputer and host PC. We propose to estimate the position in discrete space to specify a certain course-block among all the possible blocks consisting of the course. Dempster-Shafer theory based state space modeling and its state estimation have been employed to deal with uncertainty of the vehicle's position on the course more efficiently. Formulation of the problem in a state space model based on Dempster-Shafer theory, state estimation procedure being possible to work in real-time, and evaluation in simulation experiments have been demonstrated in this paper.

A Riemannian Geometric Approach to Space Debris Orbit Determination

Leonardo Cament, Pablo Barrios, and Martin Adams

Abstract: Tracking and cataloging orbiting Space Objects (SOs) is a critical component of modern Space Situational Awareness (SSA). The unpredictable nature of orbital dynamics, driven by numerous external forces, motivates the use of probabilistic filtering techniques for accurate state estimation. Much of the literature focuses on Gaussian Mixture (GM) or computationally expensive particle based filtering in SSA, however GM representations often fall short in capturing the complex, non-Euclidean structure of orbital motion. Therefore in this work, the Probabilistic Admissible Region (PAR) approach is used for initial orbit determination and, to overcome the limitations of classical GM approaches, we introduce a manifold-based approach that models the SO state on a Riemannian manifold M, allowing for more accurate prediction while maintaining computational tractability. We implement a Cubature Kalman Filter (CKF) predictor on M and validate the approach using orbital data from four CubeSats operated by the Universidad de Chile. Initial results show that the proposed method can be comparable to linear Gaussian prediction methods, particularly in short observation scenarios, and often superior in long-observation scenarios.

C3-2 · 세션 제목

내용을 채워 넣으세요.

C3-3 · 세션 제목

내용을 채워 넣으세요.

C3-4 · 세션 제목

내용을 채워 넣으세요.

C3-5 · 세션 제목

내용을 채워 넣으세요.

A Two-Stage Deep Learning Network for Super-Resolution Direction-of-Arrival Estimation

Xiaolong Li, Qinyao Chang, Zhen Cui, Wujun Li, and Wei Yi

Abstract: Achieving high angular resolution is essential in modern radar systems to distinguish closely spaced targets. Super-resolution direction-of-arrival (DOA) estimation addresses this need by overcoming the Rayleigh resolution limit. Traditional sparse reconstruction methods suffer from high mutual coherence among dictionary columns when the angular grid is densified, resulting in degraded reconstruction accuracy. Although unrolled architectures such as learned iterative shrinkage-thresholding algorithm (LISTA) mitigate this issue to some extent, they remain burdened by a large number of parameters and slow inference. To address these issues, this paper proposes a two-stage network named TS-LM Net, which integrates LISTA and multilayer perceptron (MLP). The proposed network decomposes the DOA estimation task into two sequential stages: a coarse integer-level estimation followed by a fine fractional-level refinement. By avoiding a single-stage mapping over the full angular grid, the proposed network significantly reduces the number of parameters. Simulation results demonstrate that TS-LM Net achieves comparable high estimation accuracy while realizing lower parameter complexity and faster runtime performance.

Soft Kinematics-Constrained Smooth Trajectory Prediction using a Diffusion Transformer-based Architecture

Vishnu Dev Tripathi, Vaibhav Malviya, and Rahul Kala

Abstract: Trajectory prediction is widely used in the navigation of social robotic systems. Model-driven approaches (or Physics-based algorithms) cannot capture the complexity of social compliance, and thus, data-driven approaches are widely used. We study a diffusion-based trajectory prediction algorithm that generates samples of trajectories from a learned distribution. We observe that while the generator keeps the trajectory close to the ground truth trajectory for at least one sample and thus has a small error, most of the samples have zig-zag shapes, sharp turns, are kinematically infeasible, and are unrealistic. These attributes are absent in the model-driven approaches. As the first objective, we force the trajectory generator to output kinematically feasible and smooth trajectories. Since humans can kinematically take sharp turns, our notion of kinematic constraints is soft; smoother turns are preferable, while the model should allow sharp turns if absolutely necessary. This results in more realistic trajectory prediction. We also compare the approach to several state-of-the-art methods. As another objective, we also critically analyse the public datasets for algorithmic ease of handling corner cases like sharp turns, high densities, high/low travel speeds, a large number of people, etc. and thus characterize datasets by complexity in corner cases. Not all publicly available datasets have high social interaction and mostly feature agents going nearly straight. High-interaction datasets have corner cases that are hard to learn, as they represent the heavy-tail problem. We evaluate the datasets specifically on those corner cases. We show that the inclusion of soft kinematics smoothness constraints is particularly valuable for high complexity datasets.

GenRadar: Radar Data Generation via Diffusion Network Conditioned on Multi-Radar Fused Scenes

Zuyuan Guo, Wujun Li, Hongfu Li, Yuxuan Huang, and Wei Yi

Abstract: High-fidelity radar data is essential for the robust validation of radar signal processing, yet the physical collection of diverse, large-scale datasets is often prohibitively expensive, and fails to capture rare, critical scenarios. Furthermore, even simulated datasets are often static and computationally expensive to expand. To address this data scarcity and inflexibility, we propose GenRadar, a data augmentation framework for generative radar data synthesis. Our approach re-frames the challenge as a conditional view synthesis problem, learning to generate realistic Range-Azimuth (RA) radar maps. The core of our method is a two-stage architecture that decouples scene understanding from image generation. First, a Transformer-based encoder fuses contemporaneous data from multiple radars into a single, unified, and viewpoint-agnostic scene representation. This holistic latent vector then serves as a strong condition for a powerful diffusion-based decoder, which synthesizes a sharp, physically plausible RA map. Extensive experiments demonstrate that GenRadar achieves state-of-the-art performance, significantly outperforming existing generative baselines in both quantitative metrics and qualitative realism. By providing a scalable method for transforming fixed datasets into flexible, on-demand generative tools, GenRadar offers a powerful solution for data augmentation and the creation of comprehensive test scenarios.

Korean Speech Recognition via Retrieval-Augmented Generation-Based Intent-Aware Keyword Correction

Chaerin Kim, Yeonghun Chae, Hyoungki Ahn, and Changbeom Shim

Abstract: This study proposes a novel Korean speech recognition system in Kiosks that integrates domain-specific keyword correction and visualization report generation based on Retrieval-Augmented Generation (RAG). Conventional speech recognition for Korean, characterized by its agglutinative structure, numerous homophones, and significant pronunciation variations, often exhibits high error rates, especially when recognizing specialized terminology in the cultural heritage domain. To address these challenges, the proposed system utilizes a keyword correction mechanism that combines a topic-specific keyword dictionary with semantic similarity analysis, thereby refining the recognized speech input. The system automatically generates structured HTML visualization reports based on the corrected queries. Experimental results using 70 speech data samples across seven cultural heritage topics demonstrate that the keyword correction method improved speech recognition accuracy from 45.71% to 91.43%, while the visualization report system significantly enhanced user satisfaction and learning effectiveness compared to traditional text-based responses. This work overcomes structural limitations of Korean speech recognition and demonstrates its practical applicability in cultural heritage education and guidance scenarios.

Automated Viability Classification of Human Lymphocytes via Machine Learning-Assisted Digital In-Line Holography

Kang Choi, SunMin Moon, Huijin Rim, Sanghoon Shin, Du Yong Kim, Yeonghun Chae, and Sungkyu Seo

Abstract: Digital In-line Holography (DIH) is a compact and label-free imaging technique that captures diffraction-based "shadow" patterns of biological cells without complex optical setups or staining. However, manual analysis of DIH datasets is challenging due to speckle noise, multiple scattering, and twin-image artifacts. To address these challenges, this study introduces an automated machine learning framework that integrates YOLOv8-based cell detection, extraction of physically meaningful diffraction features, i.e., Shadow Parameters, and viability classification through regression models and threshold-based decisions. Experimental validation with human lymphocyte samples identified the width of secondary diffraction maxima (WSM) as the key predictor for cell viability. Among tested regression algorithms, the random forest model performed best (R² = 0.8517), and using only WSM achieved a classification accuracy of 92.8%, highlighting its potential for real-time, label-free viability assessments. The complete analytical process was implemented using the dizest low-code platform, allowing rapid visual workflow creation without programming, thus offering a practical and scalable solution suitable for clinical diagnostics and industrial bioprocess monitoring.

Haptic Feedback for Telerobotic Ultrasound Procedures: Enhancing Safety Through Force Limiting

Meet Parmar, Himanshu K. Patel, and Rajat Jayantilal Rathod

Abstract: The development of telerobotic systems for medical diagnostics has made it easier for people in remote areas and areas that don't get enough healthcare to get it. Ultrasound imaging is an important part of non-invasive diagnostics that depends on precise control of probe pressure to make sure the images are clear and the patient is safe. But because telerobotic ultrasound systems don't give direct tactile feedback, they can apply too much or too little contact force, which can make the diagnosis less effective and the patient less comfortable. This paper talks about a telerobotic ultrasound system with haptic feedback that uses a real-time force-limiting algorithm to give the operator more control and lower the risks that come with using too much force. The system has a master-slave structure. On the patient side, there is a UR5 robotic arm with an ultrasound probe and a force sensor. On the operator side, there is a haptic interface that lets the operator feel the force in both directions. Using tissue phantoms to test the system shows that adding haptic feedback to force limitation greatly cuts down on force overshoot and makes tasks more precise. This makes the system safer and better at doing remote diagnostics.

Real-Time Colonoscopy Simulation and Expert Data Acquisition for Inverse Reinforcement Learning

Noheun Myeong and Pileun Kim

Abstract: This study proposes a real-time simulation system that replicates the physical interactions during colonoscopy using colon and endoscope models implemented in the SOFA framework. The system records expert manipulation data in real time to construct a dataset for future applications in Inverse Reinforcement Learning (IRL). Joystick inputs and endoscopic video frames are synchronized at the frame level, allowing for precise alignment of states and actions. SuperPoint-based feature extraction is used for real-time preprocessing to ensure robust state estimation, even in environments constrained to an RGB camera and lighting. This methodology facilitates the generation of reliable training datasets within anatomically realistic and real-time simulated environments, thereby establishing a foundation for learning autonomous control policies.

Digital Twin-based Design and Simulation of a Self-sensing Gripper Finger for Adaptive Grasping

Geonjun Kim, Hyun-Jin Baek, Sukhwa Lee, SungKu Kang, and Hae-Sung Yoon

Abstract: This paper presents the digital twin-based design and simulation of a self-sensing gripper finger capable of adapting its grasping mode according to the size and stiffness of target objects. Unlike conventional grippers that rely on electric sensors and active control, the proposed design utilizes torsion spring-based mechanical sensing to detect objects’ mechanical properties and adjust the internal structure of the finger. Multiple grasping modes are defined to handle various object conditions, and the self-sensing mechanism enables immediate responses. In addition, finite element simulations demonstrate that the finger achieves both higher rigidity and greater compliance compared to conventional compliant fingers. To complement the FEA results and optimize the new rib structure, a digital twin framework was implemented. The developed digital twin framework provides a foundation for future optimization of self-sensing grippers.

Experimental Study on Switching Control Scheme of Folding-Wing UAV for Flight Mode Transition

Seunghyun Choi and Dongwon Jung

Abstract: Recent advances in vertical take-off and landing (VTOL) technologies have led to the development of morphing configurations such as tilt-rotor, tilt-wing, and folding-wing vehicles. These variable-geometry systems can achieve both efficient cruise and vertical mobility, but their nonlinear dynamics and drastic aerodynamic coupling make stable transition control challenging.
This paper presents a practical switch control methodology for a variable folding-wing VTOL UAV implemented on a Pixhawk autopilot. The proposed system synchronizes the wing morphing mechanism with attitude and thrust control loops to enable smooth and stable transition between multicopter and fixed-wing modes. An analytical blending function is formulated to gradually redistribute control authority across roll, pitch, and yaw axes throughout the transition phase.
The control logic is integrated into the PX4 flight stack and validated through both high-fidelity simulation and real flight- test data. The proposed method achieved attitude tracking errors below 2◦ and demonstrated consistent switching behavior under nonlinear aerodynamic interactions. The results confirm that the proposed approach provides a reliable and reproducible methodology for real-time transition control of variable-geometry UAVs, contributing to the broader application of morphing aerial platforms.

Flight Load Derivation of a Lift & Cruise eVTOL Aircraft Based on V–n Diagram Analysis

Myeon-Gi Kim, Jinho Shin, Yun-Sang Lee, Sang-Woo Kim, and Moonsung Cho

Abstract: This study presents the derivation of flight loads for an electric vertical take-off and landing (eVTOL) aircraft based on a V–n diagram analysis in accordance with FAA 14 CFR Part 23 Appendix A and ASTM F3396/F3396M standards. Using the KUAM-301 Rev. 02 Lift-and-Cruise configuration as a reference model, the minimum design speeds and load factors were obtained to construct the flight envelope. The derived positive and negative limit load factors were 3.0 and −1.5, and the corresponding gust limit load factors were 3.41 and −1.86. From these parameters, the maximum flight loads acting on the main structural components were calculated as 45,160 N for the main wing, 8,249.7 N for the horizontal stabilizer, and 5,579.3 N for the vertical stabilizer. The results verify the applicability of fixed-wing load estimation procedures to eVTOL cruise conditions and provide fundamental reference data for the preliminary structural design and safety evaluation of future urban air mobility vehicles.

Finite Element Analysis of the Thermal Response of Glass Fiber/Vinyl Ester for Drone’s Outer Skin

Seong-Hun Jeon, Myeong-Gi Kim, and Sang-Woo Kim

Abstract: Drones that perform missions in high-temperature environments such as structural fires, high-temperature facilities, and volcanic activity are greatly affected by the thermal environment; therefore, selecting thermal-protection materials for the drone outer skin is essential. Glass fiber/vinyl ester composite has been reported in studies evaluated on the premise of application to drone structures, making it a valid candidate material for the outer skin. This study implemented a heat-transfer analysis code by coupling a Finite Element Analysis (FEA) program with user subroutines and performed code verification against literature data for a glass fiber/vinyl ester plate. The verification was conducted under conditions where a 100 mm × 100 mm × 9 mm model was subjected to a through-thickness incident heat flux of 50 kW/m² for 3,000 s. As a result, at 3,000 s the surface temperature approached the literature-reported value of 615.89 °C, and the back surface temperature converged to approximately 412.55 °C. The density began to decrease due to pyrolysis at approximately 70 s at 0.0 mm and immediately reached the char density; at 4.5 mm it started at approximately 570 s and reached full char at about 890 s, which is similar to the depth-wise delay trend reported in the literature. Thus, it was confirmed that the coupled analysis reproduces the heat-transfer behavior of the glass fiber/vinyl ester composite at the literature level and provides a verification basis applicable to evaluating the thermal response of thermal-protection materials for drone outer skins for high-temperature missions.

Ga₂O₃ Schottky Barrier Diodes as Dual Platforms for X-ray Detection and Natural Material Shielding

Wan Sik Hwang

Abstract: This work demonstrates the versatility of β-Ga₂O₃ Schottky barrier diode (SBD) detectors through two complementary applications: (i) discrimination of stacked A4 paper sheets under low-dose X-ray irradiation, and (ii) evaluation of soft X-ray shielding performance of natural wood films with different densities. The fabricated Ga₂O₃ SBDs exhibited stable diode characteristics with low leakage current and high rectification ratios (>10⁶). In the A4 sheet study, the detectors achieved a maximum sensitivity of 820 μC·Gy⁻¹·cm⁻² at −10 V and successfully distinguished up to seven stacked sheets, with attenuation behavior following the Beer–Lambert law25apl_Xray A4 Ga2O3. In the wood study, shielding efficiency increased monotonically with density, with the densest sample (0.91 g/cm³) achieving >97% attenuation at room temperature and retaining >85% efficiency at 423 K. These results highlight Ga₂O₃ SBDs not only as high-sensitivity detectors but also as robust platforms for evaluating eco-friendly, lightweight shielding materials suitable for aerospace and radiation-sensitive environments.

Keyframe Selection Strategies for Balancing Global Accuracy and Local Stability in Endoscopic SLAM

Hyeon-Seo Kim, Seokmin Hong, and Byungjeon Kang

Abstract: This study evaluates the effect of keyframe selection strategies on endoscopic simultaneous localization and mapping (SLAM) using the trans t2 b sequence of the C3VD dataset. Six strategies, including Fixed-interval, Adaptive baseline, Adaptive kf25/50/100, and Sharp-photo only, were compared. The Fixed-interval method achieved the highest global accuracy, reporting the lowest errors in ATE Sim(3) (11.43 mm) and ATE SE(3) (16.30 mm). In contrast, Adaptive kf50 exhibited the largest global error (18.38 mm and 18.63 mm, respectively). For local stability, the Sharp-photo only strategy demonstrated the best rotational precision with Rotation RPE Δ = 1: 0.36° and Δ = 5: 1.14°, whereas Adaptive kf25 showed the largest rotational deviation at Δ = 5 (1.30°). In terms of translational stability, Adaptive Kf50 achieved the lowest Translation RPE Δ = 1 (2.95 mm), while Sharp-photo only maintained the best performance at Δ = 5 (5.53 mm). Overall, these results highlight a trade-off between global trajectory accuracy and local motion stability, suggesting that while adaptive scheduling enhances frame efficiency, Fixed-interval and Sharp-photo only approaches provide more consistent and robust performance under endoscopic imaging conditions.

A PDMS-Based Acoustic Fresnel Lens for Vortex Trapping of Microparticles

Viet Hoang Nguyen; Hiep Xuan Cao, Daewon Jung, Jong-oh Park; Byungjeon Kang

Abstract: Acoustic tweezers have gained increasing attention as a contactless method for particle manipulation, yet most existing implementations rely on bulky standing-wave devices or complex phased transducer arrays that restrict their practicality. In this study, we propose and demonstrate a PDMS-based acoustic lens inspired by the principle of a Fresnel zone plate. The lens was fabricated using a simple and low-cost mold-casting process, highlighting its scalability and suitability for compact acoustic systems. Numerical simulations and experimental validations confirmed the effectiveness of the proposed lens, demonstrating its ability to generate a vortex trap and to stably confine a 3D-printed spherical particle with a diameter of 750 μm. These findings establish the PDMS acoustic lens as a practical and cost-effective alternative for acoustic manipulation, while also opening opportunities for its future integration into biomedical, diagnostic, and microfluidic platforms.

Mode-Switching-Based Control Algorithm in Acoustic Actuator for 3D Microrobot Manipulation in Nonhomogeneous Medium

Hiep Xuan Cao, Daewon Jung, Hoang Viet Nguyen, Seokmin Hong, Jong-oh Park, and Byungjeon Kang

Abstract: Acoustic technologies offer a promising and versatile toolkit for tracking and manipulating both synthetic and biological objects, ranging from nanometer to millimeter sizes, within complex biological environments. This work reports a Mode-Switching-Based Control Algorithm was developed to precisely control a microrobot. We developed an integrated system employing a single acoustic actuator capable of both transmitting and receiving signals, thereby streamlining the hardware architecture. Experimental results demonstrated a 22.7% enhancement in peak acoustic pressure. Positional accuracy was validated with mean deviations of 80 ± 75 micrometers, 71 ± 70 micrometers, and 204 ± 200 micrometers along the X, Y, and Z axes, respectively well below the microrobot's 600 micrometers body length indicating high-fidelity control. This proof-of-concept highlights the potential for dual-function acoustic devices to enable simultaneous tracking and manipulation, paving the way for advanced biomedical applications.

On-Device Vision Language Model Based Natural Language Mission Execution Framework

Seunggyu Lee, Seungjun Oh, Sangwoo Jeon, Youngju Park, Siwon Lee, Ghilmo Choi, and Dae-Sung Jang

Abstract: This work presents a framework that enables robots to assist with everyday tasks by leveraging Vision-Language Models (VLMs). Our framework takes both environmental information and natural language commands from users as inputs. An on-device VLM generates and validates a behavioral plan, represented as a sequence of pseudo-functions, using an actor-critic structure. These pseudo-functions are then connected to the low-level control system of a UGV to demonstrate the feasibility of the approach. Experimental results show that the framework can reliably interpret and execute user commands, highlighting its potential as a step toward practical robot assistance in daily life.

Situation-Aware Risk Quantification for UAV-Satellite Communications using Semantic Anomaly Detection

Junbeom Park, Zizung Yoon, and Jongsou Park

Abstract: Unmanned Aerial Vehicle (UAV)–satellite communications are increasingly critical for mission operations yet remain vulnerable to diverse packet-level and protocol-level threats, including jamming and spoofing. Existing intrusion detection systems (IDSs), however, lack situational awareness, often assessing anomalies with static severity regardless of mission context. This study introduces a semantic detection approach that integrates DistilBERT-based packet classification with a mission-phase–aware risk score to provide situationally adaptive anomaly assessment. The proposed framework is evaluated on a custom dataset extended to six traffic classes, embedding packet-level symptoms of threats across both the packet and protocol layers. Experimental results demonstrate that DistilBERT achieves 99.0% accuracy and a 98.9% macro-F1 score, while maintaining an inference latency of 26.2ms per packet—making it suitable for on-board deployment. More importantly, the proposed Situational Risk Score effectively differentiates threat criticality across mission phases. It reveals that spoofing and privilege abuse are most critical during takeoff and landing, whereas jamming poses the greatest risk during mission execution. These findings confirm that combining semantic anomaly detection with situational context not only ensures high accuracy but also delivers operationally relevant risk quantification for UAV–satellite systems.

Design, fabrication and evaluation of 6R-Pantograph Antenna mechanism

Hyeongseok Kang and Byungkyu Kim

Abstract: Research on the development of satellite antenna technologies for rapidly transmitting large volumes of data has been actively conducted. Since the performance of an antenna is determined by the size of its aperture, the development of large antenna deployment mechanisms is essential. Accordingly, the 6R-Pantograph Antenna, which combines the conventional 6R Ring-truss antenna structure with a pantograph module is proposed. To verify the performance and feasibility of the proposed antenna mechanism, synchronization and deployment performance were evaluated with three different deployment methods (1. free deployment using torsion springs; 2. deployment speed control using a wire; 3. deployment using the pantograph module). Conclusively, experimental results showed that the pantograph module driven deployment method enables stable synchronized deployment.

A Perceptive Fingerless Soft Gripper via Inductive Sensing

Haneul Kim, Changbeom Shim, and Seonggun Joe

Abstract: Task-oriented grippers have shown reliable object manipulation through predefined strategies, particularly in structured environments. However, when they engage unknown objects (e.g., having irregular shapes, stiffness) in unstructured environments, accomplishing stable and precise prehension often requires complex perception and control. To address these open challenges, this study proposed a pneumatic adaptive fingerless gripper that can achieve reliable object grasping without prior knowledge of object. The morphology of the pneumatic adaptive fingerless gripper was made of zig-zag patterns, which plays a pivotal role in omnidirectional deformation of the gripper. To enhance the overall holding strength and lifting payload of the gripper, parametric optimization was performed. Furthermore, inductive sensing solution was seamlessly integrated with the soft gripper, enabling perceptive sensory feedback. With these in mind, we demonstrated the effectiveness of our design by successfully grasping a diverse range of objects showing potential to real-world scenarios.

A Preliminary Study on Programmable Soft Architectures via Surface Tessellations

Hayeon Shin, Nhan Huu Nguyen, and Seonggun Joe

Abstract: Kinematic synthesis enabling to multi-dimensional movements of soft bodies have been performed, in a way that either considering actuation policies (e.g., multiple actuators) or seamlessly merging different regimes that exhibit a single modal movement. This work introduces a design strategy that allows soft bodies to exhibit desired kinematic movements while ensuring reliability upon cyclic loading/unlading conditions. To this aim, parametric modeling for unit cells designed by surface tessellation is performed, identifying design parameters (including width, height, ligament, and thickness) of the unit cell. The entire soft body mechanics is analyzed by employing constitutive hyperelastic modeling. As a result, it is identified that deformability is strongly influenced by the width-to-heigh ratio, yet the relevant stiffness of the entire soft body is dependent on both thickness of ligament and of unit cell. In light of these findings, programmable deformation of soft bodies is achievable by distributing different unit cells (having different unit cell density), which reveals that inherent stiffness of conventional soft bodies can be enhanced.

Design and Fabrication of Metamaterial Based Capacitive Sensor via 3D Printing

Seik Park, Eunji Kim, and Seonggun Joe

Abstract: This paper presents the design, fabrication, and experiment of a miniaturized capacitive tactile sensor fabricated using Stereolithography (SLA) 3D printing. Capacitive sensors rely on the distance of two electrodes, and the deformation of the dielectric layer occurs the capacitance changes, which makes sensors possible to detect changes of displacement, force, or pressure. Conventional fabrication approaches have limitation of flexibility and difficulties in achieving microscale structures, limiting their integration to wearable, minimized applications. To address these challenges, a lattice structure was embedded within the dielectric layer to enable linear compression and mechanical compliance. Finite Element Modeling (FEM) was conducted to analyze electromechanical behaviors, and experimental measurements validated the predictions. The proposed sensor demonstrates strong potential for tactile applications such as Braille detection, wearable human–machine interfaces, and immersive VR/AR systems.

Multichannel Ultrasound Driving Module-Based Drug Delivery System: Inducing Targeted Release of Doxorubicin for Liver Cancer Therapy

Daewon Jung, Hiep Xuan Cao, Hoang Viet Nguyen, You Hee Choi, Tae Il Kang, Seokmin Hong, Jong-Oh Park, and Byungjeon Kang

Abstract: Acoustic approaches for non-invasive manipulation of microparticles have emerged as a promising strategy to enhance the efficiency of drug delivery in biomedical applications. Ultrasound-driven systems, in particular, offer distinct advantages by generating well-defined pressure nodes at precise spatial locations, thereby enabling controlled particle manipulation. In this study, we present an ultrasound driving system composed of 86 transducers arranged in a hemispherical configuration. The acoustic field distribution generated by the system was characterized using an acoustic scanning device to ensure accurate performance evaluation. The primary objective of this work is the development of an ultrasound-based acoustic driving platform capable of precise microparticle control in clinically relevant scenarios. To assess the feasibility of the proposed system under biologically relevant conditions, validation experiments were conducted using agarose phantoms embedded with doxorubicin. Experimental results demonstrate that the system can achieve an enhancement of more than 30% in targeted drug release at the focal region, thereby highlighting its potential as a non-invasive and effective tool for particle control and drug delivery in medical applications.

Machine Learning-Based Estimation of Root-Zone Soil Moisture from Heterogeneous Data Sources

Kamilla Rakhymbek and Markhaba Karmenova

Abstract: Reliable Root-Zone Soil Moisture (RZSM) estimation is a persistent challenge, particularly in data-sparse regions. This study introduces an input-efficient framework to estimate RZSM at 50 cm and 100 cm depths using station meteorological data and a surface soil moisture from AMSR2 and ERA5-Land. An XGBoost model was trained and tested on sites in Kazakhstan and the United States to evaluate both local performance and cross-country generalizability. The proposed models significantly outperformed a baseline derived from reanalysis data. For example, at the Shemonaiha site (0-50 cm), the model reduced RMSE from 0.273 to 0.048. Results showed that using AMSR2 surface moisture predictor was superior for the shallower 0-50 cm layer, while using ERA5-Land surface moisture predictor was more stable for the 0-100 cm depth. Although the models generalized well within regions, their performance degraded significantly in cross-country applications, highlighting the necessity of regional adaptation. This study provides a practical framework for RZSM estimation, with future work aimed at enhancing model efficiency by incorporating static predictors such as soil properties.

Bearing-based Distributed Formation Control for Unmanned Aerial Vehicles

Jaewan Choi and Younghoon Choi

Abstract: The utilization of unmanned aerial vehicles (UAVs) for reconnaissance and surveillance has been rapidly increasing in recent years. To effectively monitor wide areas, significant research efforts have focused on the simultaneous operation of multiple UAVs, including formation flight and task allocation. This study proposes a distributed formation-flight system in which UAVs maintain formation through inter-vehicle information exchange. The system utilizes bearing-based data to preserve the formation geometry and addresses a key limitation of conventional bearing-based approaches, which require two or more leaders. The proposed method enables formation maintenance with a single leader. Furthermore, a collision-avoidance algorithm is designed, allowing each UAV to autonomously avoid obstacles while maintaining the overall formation as much as possible. The proposed system is validated through numerical simulations in MATLAB, demonstrating its effectiveness in both formation maintenance and safe maneuvering.

Reinforcement Learning Based Controller for High-Precision Dual-Servo Stages

Seung Woo Seo, Yongho Jeon, and Moon Gu Lee

Abstract: In semiconductor manufacturing, ultra-precision stages that transfer wafer or reticle are widely used not only in the frontend process such as photolithography but also in backend process such as packaging and inspection. Design requirements of these stages are the ability to move long stroke with high speed and high positioning precision. To meet these demands, dual-servo stages with a coarse and fine structure are widely employed. However, the traditional control methods such as PID (Proportional - Integral - Differential) controller and MPC (Model Predictive Controller) have shown limitations in achieving fast and precise control due to systems' nonlinearities, disturbances, and structural uncertainties. To improve the control performance, this work proposes a reinforcement learning based controller for the aforementioned dual-servo stage. This enabled faster and more precise control despite the various disturbances, nonlinearities, and friction in the dual servo stage mechanism. It can contribute to the control system design of next-generation ultra-precision process equipment such as lithography scanners.

Registration Between Physical Model and Virtual Model via Image Marker in MR Device

Jingyu Kim, Seokmin Hong, and Byungjeon Kang

Abstract: Mixed Reality(MR) is a technology that combines Virtual Reality(VR) and Augmented Reality(AR). The feature of MR is the ability to interact with both physical and virtual models in real time. This capability has led to extensive research on MR simulations in fields such as architecture, engineering, healthcare, and education. Meanwhile, in the case of effective interaction, it is very important to register between virtual and physical models accurately. Traditional registration methods involve manual registration by users or the use of image markers with patterns to obtain physical model's coordinates. However, there is an issue to acquire specific position of image markers, which is affected by given image pattern, number, their distance, and angle from the MR device. To find the markers accurately, we performed several registration experiments with various image marker-based patterns and numbers. Specifically, we conducted experiments with simple and complex image marker patterns, while varying the number of markers, distance, angle between the image markers and the MR device. Our experimental results show our effective method for the registration between image markers and MR device.

Discriminative Correlation Filter-Based Object Tracking: Recent Research Trends and Challenges

Jeongwoo Lee and Pileun Kim

Abstract: Tracking in unmanned aerial vehicles (UAVs) is an essential technology in various applications such as defense and surveillance. Discriminative Correlation Filters (DCF)-based trackers have been extensively researched due to their high computational efficiency and real-time applications. Research in DCF-based trackers has evolved through various approaches, such as enhancing feature representations and improving the filter learning process. This paper compares and analyzes representative DCF-based trackers, examining how they have overcome the limitations of earlier models.

Entity-Aware Trajectory Refinement : Next Best View Based 3D Gaussian Splatting

Namjun Park and Pileun Kim

Abstract: With the rapid growth of drone applications, efficient 3D rendering has become increasingly important. This article focuses on entity-aware importance weighting to guide optimal view selection when photographing high-complex objects such as buildings. To reduce the required image set and improve reconstruction quality, we propose 3D Gaussian Splatting (3DGS) instead of conventional photogrammetry and additionally investigate how to maximize coverage at each drone waypoint. Images collected from an arbitrary flight trajectory were processed using segmentation and Next-Best-View (NBV) analysis to adaptively generate optimized waypoints for 3DGS. Through this process, 3DGS confirmed significant improvements in rendering quality compared to arbitrary trajectories. These results demonstrate that object-aware trajectory refinement enables more efficient and higher-quality reconstruction. Future work includes enabling real-time drone posture control with advanced UAV platforms.

Investigating the Accuracy-Latency Trade-off for Real-Time Open-Vocabulary Models on an Edge Device

Jongyoon Park, Daeil Ko, and Pileun Kim

Abstract: This study systematically analyzes the trade-Offs involved in optimizing real-time robotic perception pipelines on a resource-constrained edge device, the NVIDIA Jetson AGX Orin 64GB platform. We measured the performance of pipelines by evaluating combinations of two open-vocabulary detection models (NanoOWL and YOLO-World) and two prompt-based segmentation models (NanoSAM and EfficientViT-SAM). Each pipeline was optimized using NVIDIA TensorRT across various precision levels (FP32, FP16, and Best). We present a quantitative comparison of the latency for each configuration, providing practical insights and benchmarks for researchers and developers deploying these advanced models on embedded systems.

Robust Path Navigation via Reinforcement Learning

Daeyoel Kang, Jongyoon Park, and Pileun Kim

Abstract: Deep Reinforcement Learning (DRL) for mobile robot autonomous navigation confronts several inherent limitations that hinder its real-world applicability. The stochastic nature of actions generated by DRL policies can undermine performance consistency, while inefficient exploration often delays the learning process or prevents the discovery of an optimal solution. Furthermore, high-dimensional sensor inputs, such as raw LiDAR point clouds, lead to the "curse of dimensionality," making it computationally infeasible for an agent to learn effectively. This research aims to enhance the robustness of path planning by addressing these challenges. To achieve this, a hybrid approach is proposed, integrating the flexible decision-making capabilities of DRL with the stability of traditional path planning. The proposed model adopts the Twin Delayed Deep Deterministic Policy Gradient (TD3) network as its base, incorporates a low-dimensional state representation from LiDAR data, applies a reward-driven prioritized experience replay (PER), and integrates a classical path planner to guide the learning process, thereby enhancing both stability and efficiency. Experimental results in simulation environments demonstrate that the proposed model achieves faster convergence and higher mission success rates, confirming a significant improvement in path planning robustness.