News

On November 4th, our lab visited the Chrysanthemum Festival at Yurim Park. We strolled among the fully bloomed chrysanthemums, enjoying relaxed conversations and a refreshing break. Afterward, we shared lunch together, strengthening our team bonds. It was a meaningful outing, made even more special by the chance to connect and enjoy time with everyone in the lab.

Our PhD student SeongJoo Moon presented his work titled “LiDAR-camera Online Calibration by Representing Local Feature and Global Spatial Context” at IROS2024. This research proposes a Transformer-based framework for LiDAR-camera online calibration, addressing degradation in calibration accuracy caused by physical vibrations and environmental changes. By learning feature correspondences and integrating global spatial context, the method achieves superior calibration performance compared to existing benchmarks, enhancing perception systems for autonomous driving applications.

Our PhD student Mincheul Kim presented his work titled “CCTV-Informed Human-Aware Robot Navigation in Crowded Indoor Environments” at IROS 2024. This study, which has also been accepted for publication in RAL 2024, proposes a novel navigation framework that integrates CCTV data with deep reinforcement learning to overcome the limitations of onboard sensors, enabling robots to navigate crowded indoor spaces safely and efficiently. By leveraging external CCTV insights on human movements, the framework anticipates interactions, minimizes collisions, and generates socially acceptable paths, significantly improving navigation compared to methods reliant solely on local sensing.

Our PhD student Minsung Yoon presented his work titled “Learning-based Adaptive Control of Quadruped Robots for Active Stabilization on Moving Platforms” at IROS2024. This research introduces Learning-based Active Stabilization on Moving Platforms (LAS-MP), a novel system that enables quadruped robots to balance effectively on six-DOF moving platforms by leveraging adaptive posture adjustment policies and state estimators trained on proprioceptive sensor data. The framework systematically addresses challenges caused by platform-induced inertia forces and showcases superior performance over baselines through trajectory generation, ablation studies, and detailed evaluations of its components.

Our Ph.D. students Seongjoo Moon, Mincheul Kim, and Minsung Yoon, along with our M.S. student Jiwoo Hwang, attended the IROS2024 conference held at ADNEC in Abu Dhabi, UAE, from October 14th to October 18th, 2024. The conference offered a great opportunity for the team to connect, participate in various programs, and exchange ideas. It was a valuable experience for discussing future research directions and staying updated on the latest developments in the field.

Our lab alumnus, Prof. Bochang Moon, conducted a seminar on October 14, 2024, in the undergraduate graphics class, presenting his paper “Target-Aware Image Denoising for Inverse Monte Carlo Rendering.” This study develops a target-aware image denoiser to address the noise issue in scene parameter inference through physically based differentiable rendering. The seminar was highly beneficial for students, helping them understand the challenges in inverse rendering and learn about the latest advancements in noise reduction techniques.

Professor Yoon and his M.S. students attended the KAIA2024 conference in Busan from August 15th to 17th, 2024. They presented their research papers and explored the latest advancements in artificial intelligence.

Seokryun Choi presented his paper, “Event-based Trajectory Estimation of a Ball,” which proposes synthesizing event data in Isaac Gym by emulating the mechanism of event cameras.

Taeyeon Kim presented her paper, “Articulated Motion Retargeting with 4D Gaussian Splatting,” which addresses the limitations of motion retargeting between reconstructed objects in a template-free domain.

The conference offered valuable opportunities that significantly enriched their academic experience.

Our lab alumnus, Prof. Bochang Moon, participated in SIGGRAPH 2024, held in Denver from July 28 to August 1, and presented his research titled “Target-Aware Image Denoising for Inverse Monte Carlo Rendering.” This study addresses the noise issue that arises when inferring scene parameters through physically based differentiable rendering by developing a target-aware image denoiser. The research points out that using conventional image denoisers can lead to undesirable local minima due to denoising bias and proposes a new denoiser that incorporates target image information to overcome this limitation. By determining denoising weights through linear regression based on target images, this method enables more robust scene parameter inference in inverse rendering optimization.

Chanmi Lee presented a paper at the KCC2024 Conference in the Artificial Intelligence Applications section, held from June 26th to 28th, 2024. Her presentation focused on the research topic “Adversarial Attack on Visuomotor Policy.” This research reveals a vulnerability in vision-based visuomotor policies, particularly those trained with Soft Actor-Critic (SAC). To address this, she proposed a method to attack the policy, analyzing the robot’s movement path to evaluate how adversarial perturbations impact the robot’s decision-making process.

Professor Yoon and his Ph.D. students, Woobin Im, Heechan Shin, and Guoyuan An, participated in the PhD Forum session at KCC2024. They presented their research and engaged in discussions with peers about their work. Additionally, they had the opportunity to dine with professionals from companies related to their research in Korea, allowing for valuable networking and in-depth conversations. This experience provided them with a meaningful opportunity to exchange ideas and deepen their understanding of their respective fields.

Phd-Forum-KCC24_WBDownload

Professor Yoon and his Ph.D. students, Sebin Lee, Minsung Yoon, Mincheul Kim, and M.S. student Jeil Jeong, conducted a tutorial session at the KCC2024 conference. The topic of this tutorial was “Advancements in Robot Motion Generation Techniques: From Sampling-Based to Reinforcement Learning and Applications (로봇 모션 생성 기법의 발전: 샘플링 기반에서 강화 학습 및 응용까지).”

Sebin introduced sensor technology for robot environmental perception. Minsung covered reinforcement learning techniques and their applications to robotic arms. Mincheul presented techniques for generating social paths for robots considering crowds through the integration of external sensors. Jeil explained recent advances in quadrupedal robot motion planning.

Professor Yoon and some of his Ph.D. and M.S. students attended the KCC2024 conference in Jeju from June 26th to June 28th, 2024. At the conference, they participated in paper presentations, tutorials, Ph.D. forums, and top conference sessions. Additionally, we had the opportunity to have a meal and chat with people from LG Electronics’ VS Business Division. We also visited the Arte Museum and enjoyed delicious meals, spending a fun and enjoyable time with lab members.

Jumin Lee and Sebin Lee presented their recent research, “SemCity: Semantic Scene Generation with Triplane Diffusion,” at CVPR 2024 in Seattle. This study introduces SemCity, a 3D diffusion model for generating realistic outdoor scenes using a triplane representation, enabling tasks like scene inpainting and semantic completion. Their method achieves notable results on the SemanticKITTI dataset, supporting flexible object manipulation and large-scale scene generation.