Reproduce “GraspFactory: A Large Object-Centric Grasping Dataset” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Srinidhi Kalgundi Srinivas, Yash Shukla, Adam Arnold, Sachin Chitta
GraspFactory, a new dataset from Autodesk Research, offers over 109 million 6-DoF parallel-jaw grasps for more than 33,000 geometrically diverse objects, addressing a data scarcity in robotic grasping. Models trained on GraspFactory demonstrated substantially improved generalization, achieving success rates between 0.85-1.00 on various novel industrial objects in simulation, outperforming models trained on existing datasets.
Key insights
- Greater geometric diversity in training data leads to significantly improved generalization capabilities for robotic grasping models on unseen objects.
- Scaling up object-centric grasping datasets through robust generation pipelines can overcome limitations of previous datasets, fostering more adaptable robotic systems.
- Physics-based simulation is effective for evaluating and selecting robust grasp candidates, enhancing the quality of large-scale datasets.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.