Publication | RCL Group

Flying Co-Stereo Accepted by T-RO

Mon, 12 Jan 2026 00:00:00 +0000

Our work has been accepted by IEEE Transactions on Robotics (T-RO)! Flying Co-Stereo is a cross-agent collaborative stereo vision system that leverages the wide-baseline spatial configuration of two UAVs for long-range dense 3D mapping. By developing a dual-spectrum visual-inertial-ranging estimator and a hybrid feature association strategy, our system achieves dense mapping at distances of up to 70 meters, corresponding to up to a 350% improvement in maximum perception range compared to conventional stereo vision systems.

PG-SLAM Accepted by T-RO

Wed, 03 Sep 2025 00:00:00 +0000

Our work has been accepted by IEEE Transactions on Robotics (T-RO)! PG-SLAM is a photo-realistic and geometry-aware RGB-D SLAM method for dynamic environments built upon Gaussian Splatting. It jointly maps the dynamic foreground (including non-rigid humans and rigid items), reconstructs the static background, and localizes the camera by leveraging both geometric and appearance constraints. Experiments on various real-world datasets demonstrate state-of-the-art performance in both camera localization and scene representation.

ROEVO Accepted by T-RO

Fri, 27 Jun 2025 00:00:00 +0000

Our work has been accepted by IEEE Transactions on Robotics (T-RO)! ROEVO introduces a robust visual odometry system using a novel “organized edges” feature representation for RGB-D cameras. By transforming disjoint edge pixels into sequentialized clusters, the method effectively retains textural and structural information and enables edge-level association across multiple frames via a co-visibility graph, achieving robust performance in diverse environments.

Gaussian-LIC Accepted at ICRA 2025

Mon, 27 Jan 2025 00:00:00 +0000

Our work has been accepted by ICRA 2025! Gaussian-LIC is a real-time photo-realistic SLAM system that marries Gaussian Splatting with LiDAR-Inertial-Camera fusion for robust pose estimation and photo-realistic online mapping in general unbounded scenarios. Implemented purely in C++ and CUDA with carefully designed acceleration strategies, our system outperforms existing methods while maintaining real-time capability. Code is available on .