Geometry based 6D pose estimation method works without pretrained models
G6D: Geometric Learning-Free RGB-D 6D Pose Solver for Robotic Manipulation
Computer Vision and Pattern RecognitionRobotics
Summary
Estimating the exact position and orientation of objects in 3D (6D pose) is important for robots to pick up things correctly. Many recent methods need large pretrained AI models requiring heavy computation and lots of GPU memory, which makes running them on robots harder. The authors present G6D, a technique that uses pure geometry and depth information from cameras to find object poses without any learning or prior training. This approach works well on several standard datasets and real robot tasks, even running efficiently on CPUs without needing GPUs.
What this means in practice
- •For robotics engineers: Use G6D to estimate object poses in pick-and-place robots without relying on trained AI models or GPUs, enabling more interpretable and resource-efficient perception.
- •For manufacturing automation teams: Deploy pose estimation on factory robots using only CPUs, reducing hardware costs and simplifying system integration for object handling tasks.
Authors
Yixuan Liang, William Chen, Yunan Wang, Jizhou Yan, Zhao Jin, Changling Liu, Chuxiong Hu
Abstract
6D object pose estimation is fundamental to robotic manipulation and automation. Recent zero-shot methods have significantly improved generalization to unseen objects, but most still rely on large-scale pretrained models with substantial GPU computation and memory demands. These requirements complicate deployment on robotic platforms where perception, planning, and control share limited computational resources, while learned intermediate representations offer limited geometric interpretability for task-specific adaptation. To address these limitations, we propose G6D, a learning-free, geometry-driven RGB-D 6D pose solver. Given an RGB-D observation, an object instance mask, camera intrinsics, and a CAD model, G6D generates pose hypotheses through template-based geometric matching and refines them using silhouette and depth consistency, forming a purely geometry-driven pose estimation paradigm. This paradigm requires neither pretrained visual models nor target-specific training and preserves interpretable geometric representations throughout pose estimation. Moreover, adjustable hypothesis counts provide flexible accuracy-computation trade-offs, while a CPU-only configuration supports deployment without GPU resources. Experiments on LineMOD and five BOP19 datasets demonstrate advanced performance. Real-world pick-and-place experiments further demonstrate G6D's applicability to robotic manipulation. The complete project is publicly available at https://ai4control.github.io/G6D-Project-Page .