GraspClutter6D: A Large-scale Real-world Dataset for Robust Perception and Grasping in Cluttered Scenes
IEEE RA-L 2025
[ArXiv] [Dataset] [GraspAPI] [Video] [PoseAnnoTool]
Seunghyeok Back1, Joosoon Lee2, Kangmin Kim2, Heeseon Rho2, Geonhyup Lee2, Raeyoung Kang2, Sangbeom Lee2, Sangjun Noh2, Youngjin Lee2, Taeyeop Lee3, Kyoobin Lee2
1Korea Institute of Machinery & Materials (KIMM)
2Gwangju Institute of Science and Technology (GIST)
3Korea Advanced Institute of Science and Technology (KAIST)
✅ 2025-08-11 GraspClutter has been accepted at IEEE Robotics and Automation Letters (RA-L).
✅ 2025-03-24 graspclutter6dAPI, a toolkit for 6-DoF grasping, is released on GitHub.
✅ 2025-03-21 gc6d-pose-anno, an annotation tool for 6D object pose is released on GitHub.
✅ 2025-03-10 GraspClutter6D Dataset is released at Huggingface.
⬜ Toolkit for 6D object pose estimation (compatible with BOP challenge format)
⬜ Train / Test codes for 6-DoF grasping networks in GraspClutter6D
GraspClutter6D is a large-scale, real-world dataset designed to push the boundaries of robust perception and grasping in highly cluttered environments—a critical challenge in robotics. Unlike existing benchmark datasets that often focus on simplistic scenes with minimal occlusion and limited diversity, GraspClutter6D offers a comprehensive resource tailored to practical, complex scenarios.
Key features include:
Scale and Complexity: The dataset comprises 1,000 highly cluttered scenes, averaging 14.1 objects per scene with 62.6% occlusion, captured across 52K RGB-D images. This scale is orders of magnitude more challenging than typical grasping datasets, bridging a significant gap in real-world applicability.
Diverse Real-World Scenarios: Featuring 200 objects across 75 varied environments (bins, shelves, and tables), GraspClutter6D is captured using four RGB-D cameras from multiple viewpoints, ensuring rich, multi-perspective data for training generalizable models.
Extensive Annotations: It provides 736K 6D object poses and 9.3B feasible robotic grasps, enabling research in instance segmentation, 6D object pose estimation, and 6-DoF grasp detection.
We provide benchmark results of state-of-the-art methods for segmentation, pose estimation, and grasping detection reveals critical insights into the challenges of such environments. Moreover, experiments demonstrate that grasping networks trained on GraspClutter6D significantly outperform those trained on existing datasets in both simulation and real-world settings, highlighting its value as a training resource.
