News

Prof. Yoon and PhD students (Minsung Yoon, Andrew Jeong, Taegeun Yang) led a tutorial session titled “The Evolution of Motion Generation Technologies for Intelligent Robots (지능형 로봇을 위한 모션 생성 기술의 진화)”.

Minsung introduced control methods for robotic arms and quadruped robots via deep reinforcement learning. Andrew presented diffusion and flow model-based robotic control and their future work. And Taegeun introduced recent VLA models and his research about compositional generalization for OOD tasks.

Jaemin Kim presented a paper at the KRoC 2025 Conference, held from February 4th to 6th, 2026. His presentation focused on the research topic “VLM-Guided Motion Planning for Automated Assembly”  This study proposes a method to accelerate the motion planning of parts in assembly tasks, by leveraging the spatial reasoning capabilities of pre-trained Vision-Language Models(VLMs).

홍콩에서 열린 SIGGRAPH Asia 2025에 참가하였습니다. 이번 홍콩 일정에는 홍콩대(HKU) Jia Pan 교수, 이화여대 김영준 교수와 함께 글로벌 로봇 산업의 중심지인 중국 선전을 방문했습니다. 특히 Limit X, UBtech Robotics, 그리고 Engine AI 등 주요 휴머노이드 로봇 기업을 방문하며 차세대 로봇 기술의 현주소를 직접 확인했습니다.

Jaeyoon presented his research at the AAAI2026 Conference, held from January 20 to 27, 2026, in Singapore EXPO.

His presentation introduced the paper, “Towards Test-time Efficient Visual Place Recognition via Asymmetric Query Processing,” which addresses the critical challenge of deploying high-performance Visual Place Recognition (VPR) on resource-constrained devices. The proposed framework bridges the gap between high-capacity foundation models and lightweight query networks by utilizing a novel geographical memory bank, achieving superior retrieval performance while significantly reducing computational overhead.

We are pleased to announce that Prof. Yoon has been selected as a Distinguished Member of the Association for Computing Machinery (ACM), the world’s largest computer science society. The ACM Distinguished Member grade is awarded to members with at least 15 years of professional experience who have achieved significant accomplishments or made a strong impact on the field of computing. This honor is reserved for the top 10% of ACM’s global membership (approximately 110,000 members), with 50-60 individuals selected annually.

He is recognized as an authority in the field of Physical AI and ACM acknowledged his research achievements in scalable rendering for high-speed processing of large-scale 3D models, physics-based simulation, and data generation for AI computer vision. His technologies are currently widely applied as core foundations in future industries ranging from 3D graphics to robotics.

News link

Woojae and Junsik attended 36th British Machine Vision Conference (BMVC) held during Nov. 24-27th in Sheffield, UK.

Woojae presented AegisRF: Adversarial Perturbations Guided with Sensitivity for Protecting Intellectual Property of Neural Radiance Fields, a novel framework that protects intellectual property of NeRF by employing a learnable sensitivity field to inject adversarial perturbations that disrupt unauthorized downstream tasks while preserving 3D rendering quality.

Junsik presented Learning Event-guided Exposure-agnostic Video Frame Interpolation via Adaptive Feature Blending, a novel event-guided framework for exposure-agnostic video frame interpolation that utilizes target-adaptive sampling and importance mapping to recover sharp frames from blurry inputs by optimizing temporal and spatial feature alignment.

Yoonki presented his paper, “Towards Robustness of Person Search against Corruptions,” at the ICCV 2025 Conference, held from October 19-23, 2025, in Honolulu, Hawaii.

The paper introduces new benchmarks (CUHK-SYSU-C and PRW-C) to evaluate how person search models fail when images are corrupted. Based on this analysis, it proposes a foreground-aware augmentation technique and a regularizer to substantially improve model robustness.

Jinhwan and Yoonki won 1st place out of 23 teams in the Grounded Video Question Answering (GVQA) track at the ICCV 2025 Perception Test Challenge. The framework introduces a three-stage pipeline (Reasoning, Grounding, Tracking) and a novel “trigger moment” derived from a CORTEX prompt for robust anchoring. This method achieved a HOTA score of 0.4968, substantially surpassing the previous year’s winning score of 0.2704. Congratulations!

Prof. Yoon visited Zhejiang University, one of China’s top universities, and gave an invited talk on our lab’s research.

The visit and seminar were arranged by Prof. Qi Ye, who is also an alumnus of Prof. TaeKun Kim. During the visit, we also had a great time with our former collaborators and SGVR alumni, Prof. Yuchi Huo (former postdoc) and Prof. Min Tang.

This was a valuable opportunity to share our research and reconnect with valued colleagues.

중국 항저우(Hangzhou)에서 열린 세계적인 로봇 학회 IROS 2025에서 SGVR Lab의 반가운 동문, Peng Du 교수님을 오랜만에 만났습니다. Peng Du 교수님은 2014년부터 2015년까지 우리 연구실에서 박사후연구원(Postdoc)로 함께 하셨으며, 당시 충돌 탐지(Collision Detection) 분야를 주로 연구하셨습니다. 현재는 Zhejiang 대학에서 연구 교수로 재직 중이시며, 최근에는 로보틱스 분야로 연구를 확장하며 큰 성과를 내고 계십니다. 특히 주목할 만한 소식은, Peng Du 교수님의 박사 제자 중 한 명이 세계적으로 유명한 로봇 기업인 Unitree의 창업 멤버라는 점입니다. 이를 바탕으로 교수님은 Unitree와 Zhejiang 대학 사이의 (Collaboration Center)를 운용하시며 산학 협력을 이끌고 계십니다.

Taegeun presented his paper at the IROS 2025 Conference, held from October 19 to 25, 2025.

His presentation focused on the research topic, “Efficient Navigation Among Movable Obstacles using a Mobile Manipulator via Hierarchical Policy Learning”. This research introduces a hierarchical reinforcement learning framework enabling a mobile manipulator to efficiently navigate among movable obstacles by estimating their properties and strategically pushing them.

Prof. Yoon, Taegeun, Jiwoo, and Xiangchen Liu attended the conference for the presentation, and it was a valuable opportunity to grasp the latest research trends.

We are thrilled to continue our lab’s tradition by announcing our second round of awards, celebrating the outstanding achievements of our members.

Research Impact Award Recipient: Woojae Kim

Continuing our tradition, the Research Impact Award goes to Woojae Kim for his highly-cited 2023 paper, “Feature separation and recalibration for adversarial robustness.” This award honors students whose research has demonstrated a significant and measurable impact on the academic community. His work exemplifies the profound and lasting influence that rigorous academic research has on the scientific community.

Open Source Contribution Award Recipients: Jumin Lee and Sebin Lee

This year, we are also thrilled to inaugurate the Open Source Contribution Award. This award recognizes students who have made outstanding contributions to open-source software, demonstrating technical excellence, community impact, and commitment to collaborative innovation.

The recipients are Jumin Lee and Sebin Lee. Their project for the 2024 paper, “Semcity: Semantic scene generation with triplane diffusion,” has become highly popular on GitHub, proving that open-source contributions can generate impact even faster than traditional citation metrics.

Congratulations to Woojae, Jumin, and Sebin for leading the way in both foundational research and modern innovation.

Jaeyoon’s paper, “Enhancing Visual Re-ranking through Denoising Nearest Neighbor Graph via Continuous CRF,” has been selected as a Spotlight Paper at ICIP 2025.

Out of 643 papers accepted at the conference, only the top 5.6% were chosen for a spotlight presentation.

The conference, the IEEE International Conference on Image Processing (ICIP), was held from September 14-18, 2025 in Anchorage, Alaska, USA.

The research addresses the noisy edge problem, a key limitation in Nearest Neighbor graph-based methods for improving visual retrieval. It introduces a novel denoising method that uses a statistical distance metric to robustly remove incorrect connections, leading to significant gains in retrieval accuracy.

Youngju and Taeyeon won the Best Student Paper Award at ICIP 2025 for their paper, “Pose-free 3D Gaussian splatting via shape-ray estimation.”

Out of 643 papers accepted at the conference, only one was selected for the Best Student Paper Award (Top 0.16%).

They presented their award-winning research at the conference, the IEEE International Conference on Image Processing (ICIP), which was held from September 14-18, 2025, in Anchorage, Alaska, USA.

The research proposes a framework that enables high-quality 3D scene reconstruction by jointly estimating shape and camera rays, thus removing the dependency on precise camera poses.

Congratulations!

Sangmin Lee presented a paper at the CKAIA2025 Conference, held from August 7 to 9, 2025. His presentation focused on the research topic, “RFDSI: Rectified Flow Distillation via 2nd-Order Inversion for Consistent Text-to-3D Generation”. This study proposes a method to mathematically reformulate RF distillation, to pinpoint why it succeeds, and to derive the optimal setting that brings the student’s 3D quality as close as possible to the teacher’s.

과학기술정보통신부가 추진하는 생성형 AI 인재 양성 사업에서 LG AI 연구원이 주관기관으로 선정되었고, 우리 연구실이 핵심 파트너로 함께합니다. 컨소시엄은 로봇·물리 환경 이해를 결합한 ‘피지컬 AI’ 분야의 대규모 생성모델 기술을 선도할 인재를 육성하는 것을 목표로 합니다. 이번 프로그램을 통해 연구실은 LG AI 연구원 및 참여 대학들과 공동 교육 과정을 설계하고, 실전 프로젝트 기반의 연구·실습을 지원하여, 피지컬 AI 분야의 거대 생성모델 기술을 선도할 인재를 양성할 것입니다. 관련 기사는 다음 기사 링크와 PDF에서 확인 가능합니다.