GraspGen: 57 Million Simulated Grasps for Three Robot Grippers

Over 57 million simulated grasps computed for 8,515 objects from the Objaverse XL dataset. The dataset is specific to three gripper types: the Franka Panda, the Robotiq-2f-140, and a single-contact suction gripper. It was created by NVIDIA and released on Hugging Face in June 2025.

Use Cases

• Training machine learning models for robotic grasping based on simulated grasp poses.
• Benchmarking sim-to-real transfer performance for different gripper types mentioned in the description.
• Analyzing grasp success factors across a large set of 3D objects from Objaverse XL.
• Developing and evaluating grasp planning algorithms for industrial or research robot arms.

Strengths

• Contains over 57 million individual grasp data points.
• Covers 8,515 distinct 3D objects from the large-scale Objaverse XL collection.
• Provides data for three specific, commonly used robot gripper models.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Data is simulated, which may not perfectly reflect real-world physics and performance.

Provenance

Source

NVIDIA.

Collection Method

Grasps were computed in simulation for objects sourced from the Objaverse XL (LVIS) dataset.

Freshness

Last updated 2025-06-21 18:43:53.