Supermarket-6DoF: A Real-World Grasping Dataset and Grasp Pose Representation Analysis
Jason Toskov, Akansel Cosgun
TL;DR
The paper tackles the lack of real-world benchmarking for 6-DoF robotic grasping by introducing Supermarket-6DoF, a dataset of 1,500 real-world grasp attempts across 20 supermarket objects with ground-truth success and stability labels. It provides rich sensory data (RGB, depth, and point clouds) and exact 6-DoF grasp poses, enabling analysis of grasp representations beyond traditional top-down or analytic metrics. The authors compare three gripper-pose representations and demonstrate that modeling the gripper as a point cloud yields the best grasp-success prediction performance, with stability prediction remaining more challenging, especially for heavier or more rigid objects. The dataset and accompanying tools support reproducible benchmarking and are poised to advance learning-based grasping methods in real-world manipulation tasks.
Abstract
We present Supermarket-6DoF, a real-world dataset of 1500 grasp attempts across 20 supermarket objects with publicly available 3D models. Unlike most existing grasping datasets that rely on analytical metrics or simulation for grasp labeling, our dataset provides ground-truth outcomes from physical robot executions. Among the few real-world grasping datasets, wile more modest in size, Supermarket-6DoF uniquely features full 6-DoF grasp poses annotated with both initial grasp success and post-grasp stability under external perturbation. We demonstrate the dataset's utility by analyzing three grasp pose representations for grasp success prediction from point clouds. Our results show that representing the gripper geometry explicitly as a point cloud achieves higher prediction accuracy compared to conventional quaternion-based grasp pose encoding.