TLDR: GraspFactory is a new, large-scale dataset featuring over 109 million 6-DoF grasps for Franka Panda and Robotiq 2F-85 grippers across tens of thousands of geometrically diverse objects. Developed by Autodesk Research, it addresses the challenge of robot grasping models struggling with novel objects due to limited training data diversity. The dataset was created through grasp sampling, collision checks, and physics-based simulation evaluations. Experiments show that models trained on GraspFactory significantly outperform those trained on existing datasets in generalizing to unseen objects, both in simulation and real-world robotic tasks.
Robots are becoming increasingly common in industrial settings, handling a vast array of objects. However, a significant challenge for robotic grasping models is their ability to adapt and generalize when encountering new, unfamiliar objects. Traditional training datasets often lack the geometric diversity needed for robots to reliably grasp a wide range of items found in real-world environments like warehouses or manufacturing plants.
Addressing this critical need, a new research paper introduces GraspFactory: A Large Object-Centric Grasping Dataset. This innovative dataset is designed to empower the next generation of robot grasping models with enhanced generalization capabilities. The paper, authored by Srinidhi Kalgundi Srinivas, Yash Shukla, Adam Arnold, and Sachin Chitta from Autodesk Research, presents a comprehensive solution to the problem of limited geometric diversity in existing datasets.
What is GraspFactory?
GraspFactory is a massive dataset containing over 109 million 6-Degrees-of-Freedom (6-DoF) grasps. These grasps are specifically tailored for two popular robotic grippers: the Franka Panda, covering 14,690 distinct objects, and the Robotiq 2F-85, encompassing an even larger set of 33,710 objects. The key to GraspFactory’s effectiveness lies in its object diversity, drawing from the ABC-Dataset, which features over a million high-quality geometric models. This ensures that the dataset includes objects with a wide variety of shapes and complexities, unlike many prior datasets that focused on a limited number of semantic classes.
How was the dataset created?
The creation of GraspFactory involved a sophisticated three-step process:
- Grasp Sampling: Candidate grasps were initially generated for each object’s 3D CAD model using an antipodal sampling method. This technique identifies potential contact points on an object’s surface where a gripper could achieve a stable hold.
- Collision Checks: To ensure practicality, these candidate grasps underwent rigorous collision checks. Any grasp where the gripper’s fingers would collide with the object itself was filtered out.
- Physics-Based Evaluation: The remaining grasps were then evaluated for robustness in the Isaac Sim simulator, a highly accurate physics simulation environment. This step determined whether a grasp could firmly hold an object and withstand external forces, mimicking real-world conditions. For instance, the simulation tested if the object remained securely held as the robot moved it through predefined poses.
Demonstrating Generalization
To validate the dataset’s impact, the researchers trained an existing diffusion-based grasping model, SE(3)-DiffusionFields, using a subset of GraspFactory for the Franka Panda gripper. They then compared its performance against the same model trained on the ACRONYM dataset, a widely used existing dataset.
The results were compelling. In simulated experiments, the model trained on GraspFactory significantly outperformed the ACRONYM-trained model across a range of novel industrial objects. For many complex objects, the GraspFactory-trained model achieved success rates close to 100%, while the ACRONYM-trained model often failed completely or had very low success rates. This highlights GraspFactory’s ability to improve a model’s generalization to objects it has never seen before.
Beyond simulation, the team conducted extensive hardware experiments using a UR10e robot equipped with a Robotiq 2F-85 gripper. These real-world tests evaluated both the feasibility and robustness of the grasps generated by the GraspFactory-trained model. With over 900 grasps evaluated across 18 diverse objects, the model demonstrated high success rates, proving its practical utility in physical settings. The experiments confirmed that the dataset enables models to generate grasps that are reliable and executable in the real world.
Also Read:
- How Robots Learn Human-Like Object Manipulation from Video
- ImaginationPolicy: A New Approach for Versatile Robotic Manipulation
Looking Ahead
The introduction of GraspFactory marks a significant step forward in robotic grasping. By providing an unprecedented scale and geometric diversity of 6-DoF parallel-jaw grasps, it offers a powerful resource for training data-intensive models. The researchers plan to further enhance the dataset by integrating finger geometry into training, extending it to an even larger number of objects from the ABC dataset, and including grasps for other types of end-effectors, such as suction-cup grippers. This ongoing work promises to further expand the versatility and applicability of robotic grasping systems.
For more technical details, you can read the full research paper here.
