PhysicalAI Robotics GraspGen
Large-scale robotic simulation dataset for gripping objects, compatible with various industrial grippers.
57 million grip poses in WebDataSet format (.tar), JSON for splits
CC-BY 4.0
Description
GraspGen is a large-scale simulated robotic data set designed to improve the precision and robustness of gripping algorithms. It contains more than 57 million input attempts generated for 8515 3D objects from the Objaverse XL repository. Seizures are simulated for three types of grippers: the Franka Panda arm, the robotiQ-2f-140 gripper, and a single-contact suction cup.
What is this dataset for?
- Training AI models for robotic gripping in simulation
- Testing Sim2Real approaches in manipulating 3D objects
- Create benchmarks for different types of grippers
Can it be enriched or improved?
Yes. It is possible to add other types of tweezers, 3D objects, or environmental conditions to increase the diversity of scenarios. The integration of sensory feedback or real-time data would make it possible to go further in realism. New splits or annotations (such as object complexity) can also improve the granularity of analyses.
🔎 In summary
🧠 Recommended for
- Robotics laboratories
- AI manipulation engineers
- Sim2Real simulators
🔧 Compatible tools
- MeshCat
- PyBullet
- Isaac Sim
- WebDataSet
- Python
💡 Tip
For best results, filter objects by type or complexity before training or visualizing.
Frequently Asked Questions
Can this dataset be used with real robots?
Yes, it is ideal for Sim2Real transfer: the models trained on this data can be tested in real conditions.
Does the dataset contain images?
No, it contains grip poses and 3D objects, but no standard image visual rendering.
Can I visualize data without robotic hardware?
Yes, visualization scripts are provided to explore input poses using MeshCat and Python.
