Our lab is gaining attention for advances in Physical AI technologies. By connecting visual perception, physical-world understanding, prediction, and action planning, our lab is laying key groundwork for robots that can operate in real environments. These achievements, presented through four papers (two highlight papers and two oral papers) at ICLR 2026 and CVPR 2026, highlight the lab’s growing leadership in Physical AI.
KAIST School of Computing announced that Youngju‘s paper, titled “GLINT,” has been selected as one of just 74 Award Candidates out of over 4,000 submissions at CVPR 2026. This breakthrough technology solves a long-standing computer vision challenge by separating reflected and transmitted light to accurately reconstruct 3D environments containing transparent objects. See the article for details.
Our Lab attended the Korea Computer Congress (KCC) 2026 to participate in the Research Performance Exchange for the 2026 IITP “Human-Oriented Next-Generation Challenging AI Technology Development” project.
Out of 20 initial participating teams, our lab was one of three selected teams that proceeds to the 2nd phase of the project, after successfully passing the program’s phase evaluation. During the session, the final three teams shared their core research milestones and discussed strategic directions for the next phase of the project.
Andrew presented his paper, “CLaD: Planning with Grounded Foresight via Cross-Modal Latent Dynamics” at CVPR 2026, held from June 3rd – 7th, 2026, in Denver, USA.
He suggested CLaD, a framework that models the joint evolution of proprioceptive and semantic states under robotic actions through asymmetric cross-attention. By enforcing consistency over cross-modal transitions and using self-supervised objectives, the model predicts grounded latent foresights without experiencing representation collapse. These predicted foresights condition a diffusion policy for action generation, achieving a high success rate on long-horizon tasks with significantly fewer parameters than large vision-language-action (VLA) models.
Sebin and Jumin presented their paper, “Visual-RRT: Finding Paths toward Visual-Goals via Differentiable Rendering” at CVPR 2026, held from June 3rd – 7th, 2026, in Denver, USA.
Visual-RRT (vRRT) is a novel robot motion planning technology that enables robots to plan paths based solely on a target image or visual observations, eliminating the need for explicit numeric target joint values required by traditional RRT planners. By combining sampling-based RRT exploration with differentiable robot rendering, vRRT searches for paths that minimize the visual discrepancy between the target image and the current rendered robot state. The method enhances exploration efficiency through a frontier-based strategy that prioritizes visually promising nodes and an inertial gradient tree expansion technique that leverages prior optimization data.
Youngju presented his paper, “GLINT: Modeling Scene-Scale Transparency via Gaussian Radiance Transport” at CVPR 2026, held from June 3rd – 7th, 2026, in Denver, USA. The paper was presented as an oral at CVPR 2026 and selected as a Best Paper Award Candidate, corresponding to roughly the top 3.4% and 1.8% of accepted papers, respectively.
GLINT addresses the challenge of reconstructing real-world scenes containing transparent and reflective surfaces, where conventional 3D Gaussian Splatting often fails to preserve the physical presence of glass. By disentangling the visible interface from transmitted and reflected radiance, GLINT enables more faithful and physically consistent reconstruction of complex scenes with glass, windows, and specular objects.
Chanmi presented her paper, “Beyond the Patch: Exploring Vulnerabilities of Visuomotor Policies via Viewpoint-Consistent 3D Adversarial Object” at ICRA 2026 Conference, held from June 1st – 5th, 2026, in Vienna, Austria.
She presented a study on 3D adversarial attacks to identify the vulnerability of policies in continuously changing camera settings, such as visuomotor policies with wrist-cam configurations.
Jiwoo presented his paper, “Uncertainty-Aware Non-Prehensile Manipulation with Mobile Manipulator under Object-Induced Occlusion” at ICRA 2026 Conference, held from June 1st – 5th, 2026, in Vienna, Austria.
He suggested CURA-PPO, a reinforcement learning framework for safe non-prehensile manipulation under occlusions, jointly estimating collision risk and uncertainty to operate reliably in partially observable environments.
Minsung presented his paper, “Phase-Aware Policy Learning for Skateboard Riding of Quadruped Robots” at ICRA 2026 Conference, held from June 1st – 5th, 2026, in Vienna, Austria.
He suggested PAPL (Phase-Aware Policy Learning), a reinforcement learning framework that enables quadruped robots to ride skateboards by leveraging phase-aware policy modulation while maintaining a single unified control policy.
Our research group has been selected for two major InnoCORE research initiatives, marking an important step toward advancing next-generation AI and robotics research.
1. AI Meta-Scientist Initiative
Vision: Toward AI systems that autonomously generate new knowledge
The AI Meta-Scientist Initiative aims to develop a new class of AI systems capable of scientific discovery and knowledge creation.
The project focuses on integrating:
– Advanced foundation models
– Reasoning capabilities
– System-level orchestration
to realize an AI system operating at the level of a research institute director (meta-scientist-level intelligence).
2. 5D AI Robotics Initiative (5D AI-RI)
Vision: Next-generation multimodal robot AI for real-world adaptation
The 5D AI Robotics Initiative (5D AI-RI) aims to develop:
– World-leading 5D multimodal robot foundation models
– Real-world adaptive robotic platforms
with the goal of advancing embodied AI and fostering national-level
top-tier young researchers in robotics and AI.
_____________________________________________________
Program Highlights
Duration: 5 years
Support: Up to 2 postdoctoral researchers per year
_____________________________________________________
These initiatives will enable our group to push the frontier of physical intelligence, embodied AI, and AI-driven scientific discovery, while contributing to the development of next-generation AI talent and technologies.
Kyubeom‘s paper was selected as Oral paper at the ICLR 2026 conference, held from April 23rd – 27th, 2026, in Rio de Janeiro, Brazil.
He introduced his paper “Radiometrically Consistent Gaussian Surfels for Inverse Rendering”, suggesting RadioGS, an inverse rendering framework that outperforms existing Gaussian-based methods in scene property decomposition by introducing a novel radiometric consistency loss for accurate indirect illumination modeling. This physically-based constraint provides supervision for unobserved directions by minimizing the residual between learned surfel radiance and its physically-rendered counterpart through differentiable 2D Gaussian ray tracing. The framework features an efficient relighting strategy that adapts Gaussian radiances to new illumination within minutes, achieving high-fidelity, real-time rendering at under 10ms per frame.
Joonsung presented his paper, “No Caption, No Problem: Caption-Free Membership Inference via Model-Fitted Embeddings,” at the ICLR 2026 Conference, held from April 23rd – 27th, 2026, in Rio de Janeiro, Brazil.
He introduced MoFit, the first caption-free membership inference attack for diffusion models, which determines whether an image was used during training without access to its original caption. The key idea is that training samples are more sensitive to mismatched conditioning than non-members, so the method optimizes a model-fitted surrogate image and embedding to amplify this difference during denoising. Experiments on multiple Stable Diffusion models show that MoFit consistently outperforms prior VLM-caption-based attacks and can even surpass some methods that use ground-truth captions.
올해 윤성의 교수님이 석좌교수로 임명되어 축하 회식 자리를 가졌습니다.
교수님과 연구실 학생 모두의 밝은 앞날을 응원합니다!
윤성의 교수님이 로봇학제 책임교수로서 로봇학제 석박사 입시 설명회 및 삼성전자-로봇학제 프로그램, REPS 입시 설명회를 진행 하였습니다.
SGVR 연구실에서는 앞으로 지속적으로 로봇학제 학생을 선발할 예정입니다.
윤성의 교수님이 연구책임자로서 최근 정보통신기획평가원 (IITP) 연구과제에 선정되어, 앞으로 3년간 연간 약 10억원 규모의 지원을 받아 물리 통합 지능 기반 차세대 AI 연구를 수행하게 됩니다.
공동연구자인 전산학부 김태균 교수님, 성민혁 교수님, 오태현 교수님과 전기및전자공학부 명현 교수님과 더불어, (주)홀리데이로보틱스와의 협력을 통해, 원천 기술 개발에서 실제 환경에 대한 적용까지 계획하고 있습니다.
본 과제를 수행하며 물리적 이해에 기반한 디지털 세계 구축 및 에이전트의 자율 인지·판단 기술을 고도화하여, 피지컬 AI의 핵심 기술 확보에 주력할 예정입니다.