Hands-on Experience

Methodology: To address this, we propose a geometric-invariant space embedding method. As illustrated in the video, our approach enables pedestrians captured at different intersections, with varying CCTV angles and crosswalk orientations, to be embedded into a unified spatial representation. This standardization allows robust learning of crossing behaviors. Furthermore, we adopt a Transformer-based encoder, achieving accuracy of 94.10% and an F1-score of 92.35%.

Impact: The proposed method demonstrates strong potential for real-world deployment. Its robustness to intersection variability makes it suitable for integration into city-wide intelligent traffic management systems to proactively ensure pedestrian safety at signalized intersections.

Methodology: We calibrate top-down drone views with CCTV footage to construct vehicle bounding boxes in both front and bottom perspectives. By calibrating vehicles captured in drone views with those in CCTV footage, we develop a cross-view dataset that includes bottom bounding boxes. A model is then trained to estimate the both bounding boxes from CCTV-only views, enabling accurate spatial localization without relying on drone footage.

Impact: This work enables real-time vehicle safety monitoring using only existing CCTV infrastructure. By bridging the gap between aerial and CCTV footage, it allows accurate estimation of surrogate safety measures at intersections, contributing to proactive vehicle conflict detection and urban traffic safety management.

Impact: Our approach facilitates accurate multi-agent trajectory forecasting at intersections using observed movement patterns. By modeling pedestrian–vehicle interactions, it enhances the reliability of safety assessment and supports the development of smarter traffic control systems that account for both pedestrians and vehicles.

Methodology: We developed 3D-ASCN that recognizes objects using 3D point cloud data from LiDAR sensors. To address the challenges caused by variations in LiDAR sensor configurations, our model is designed to maintain reliable performance even when the quality or resolution of LiDAR data changes. Specifically, we introduce a distance-based kernel and a direction-based kernel that learn the structural feature representations of objects from point clouds. These kernels enable the model to focus on the intrinsic geometry of objects rather than the density of the point cloud itself, allowing for robust classification and object detection across different LiDAR resolutions.

Impact: The proposed 3D-ASCN can contribute to the real-world deployment of AVs by enabling reliable object recognition regardless of the LiDAR sensor used. By overcoming performance degradation caused by LiDAR channel shifts, our approach reduces hardware dependency and cost, allowing AV systems to be both scalable and economically feasible.

Benchmark Statistics: The VRU-Accident benchmark consists of 1K real-world dashcam accident videos involving VRUs. It features 6K multiple-choice video question answering (VQA) questions covering six categories: weather & lighting, traffic environment, road configuration, accident type, accident cause, and prevention measure. Each question comes with four answer choices, one correct and three counterfactual distractors, resulting in 24K candidate options, including 3.4K unique answers to encourage diverse and nuanced reasoning. The benchmark also provides 1K densely annotated scene-level captions that narrate the spatiotemporal dynamics of accident scenarios.

Impact: VRU-Accident serves as the first large-scale benchmark to systematically evaluate MLLMs' understanding of VRU-related accident scenes through both VQA and dense captioning tasks. By providing diverse and semantically rich annotations grounded in real-world scenarios, the benchmark enables quantitative evaluation of MLLMs' reasoning, grounding, and narrative capabilities in high-risk traffic contexts. It lays the foundation for developing safer and more interpretable AI systems for AVs and transportation safety research.

Impact: Safe-LLaVA offers the first comprehensive solution for training and evaluating privacy-aware MLLMs. It enables researchers and practitioners to develop models that are both biometrically safe and informative, which is critical for real-world deployment in domains requiring strict privacy compliance. By preventing both explicit and implicit biometric inference, Safe-LLaVA represents a key step toward building responsible and ethically aligned multimodal AI systems.

Motivation: Robotic grippers are widely used in industrial and research settings for object grasping. Conventional soft grippers such as the Fin Ray Effect (FRE) grippers can achieve stable adaptive grasping without complex control, but they have limitations in manipulating the pose or orientation of grasped objects.

Methodology: To address the limitations of soft grippers, we designed a novel gripper mechanism by adding an additional degree of freedom (DOF) to a standard FRE-inspired structure, enabling controlled deformation for manipulation tasks. We performed mathematical modeling and parametric analysis to determine the optimal force application points and directions for effective object manipulation. Then, we conducted Finite Element Analysis (FEA) simulations to validate deformation behavior under different load conditions.

Impact: The proposed gripper design enables simultaneous stable grasping and in-hand manipulation, overcoming the main limitations of conventional adaptive grippers. The gripper also provide a foundation for further development of adaptive robotic end-effectors capable of handling complex manipulation tasks in real-world environments.

Overview: This project was carried out as part of my undergraduate capstone experience, with the goal of integrating the fundamental principles of mechanical engineering, including kinematics, dynamics, and motor design, into a comprehensive robotic system. I was particularly inspired by the Omega Gripper, published in IEEE/ASME Transactions on Mechatronics, which was specifically designed to grasp challenging objects such as thin cards. Recognizing this as an excellent example of how core engineering knowledge could be combined into a solution, I adopted the Omega Gripper as the foundation of my work. To deepen my understanding, I visited the POSTECH March Lab and consulted directly with the first author of the original paper, which helped resolve technical questions and further motivated me to pursue this project.

Gripper Kinematic Design: The gripper mechanism features an underactuated structure implemented using a parallel four-bar mechanism. I derived and analyzed both the closed-loop forward and inverse kinematic equations to model the motion mathematically and ensure precise control over the gripper’s configuration during operation.

Kinematic Simulation: Using MATLAB, I simulated the kinematics of the mechanism to optimize key parameters such as link lengths and ranges of motion, ensuring that the gripper could reliably adapt to a variety of object shapes and sizes.

Dynamic Simulation and Prototyping: Dynamic simulations were conducted to estimate the loads applied during grasping and to select appropriate material properties for the gripper body. Based on these analyses, I designed the gripper structure and fabricated the components via 3D printing. I also selected motors considering factors such as required grasping force and system responsiveness.

Demonstration and Real-time Tracking: For demonstration, I integrated an encoder at the lower part of the gripper and an IMU sensor at the center. The gripper was operated based on the inverse kinematic equations to achieve adaptive grasping of objects. Real-time tracking and visualization of the gripper’s state were performed using forward kinematics, allowing continuous monitoring of its configuration during operation.