Fast and lightweight 6D object pose estimation for efficient robotic grasping

Efficient object recognition and pose estimation constitute critical challenges in advancing service robots toward practical applications. Despite considerable progress, achieving real-time detection and reliable 6D pose estimation under stringent precision requirements remains an open problem. To tackle this issue, we propose a lightweight framework that integrates graph matching with geometric feature methods, enabling real time recognition and 6D pose estimation of target objects characterized by visually distinctive planar surfaces in complex and dynamic environments. Firstly, a target detection module is developed based on the lightweight graph matcher LightGlue, with adaptations designed for target objects exhibiting diverse textures, appearances, and surface geometries. To enhance recognition stability, a combination of Kalman filtering (KF) and exponential moving average (EMA) optimization strategies is employed, ensuring consistent detection and tracking of targets. Secondly, a 6D pose estimation method based on geometric vector mapping is introduced. By computing vectors from the object’s center to key geometric points, an orthogonal coordinate system is constructed to represent the pose, from which the orientation angles at the center are derived. This method achieves high frame-rate processing while preserving estimation accuracy, thereby satisfying real-time requirements in dynamic environments. Finally, building on the proposed recognition and pose estimation methods, a complete robotic grasping system is developed, enabling adaptive 6D grasping even under steep inclination angles. Experiments show that our scheme outperforms traditional methods in detection speed and exhibits stronger robustness and adaptability under wide viewing angles ( \(36^\circ \, \pm \,2^\circ\) ).