217,000 synthetic RGB-D images and 100,000 point clouds across 82 object categories for robotic manipulation. It contains ground truth for 6-DOF grasp poses, instance segmentation masks, and 3D object meshes.
Use Cases
- Train 6-DOF grasp detection models using the grasp_pose and success_score labels
- Develop instance segmentation algorithms using the segmentation_mask and object_id columns
- Benchmark synthetic-to-real transfer using the synthetic depth maps and real-world sensor data
Strengths
- 217,000 RGB-D images with corresponding depth maps and instance segmentation masks
- 82 unique 3D object models provided in .obj format
- Includes 6-DOF grasp labels with pose coordinates and success scores
- Features varying clutter levels ranging from isolated objects to highly occluded piles