Reproduce “Dex-Net 2.0: Deep Learning to Plan Robust Grasps with Synthetic Point Clouds and Analytic Grasp Metrics” on OpenResearch

About this paper

Jeffrey Mahler, Jacky Liang, Sherdil Niyaz, Michael Laskey, Richard Doan, Xinyu Liu, Juan Aparicio Ojea, Ken Goldberg

Dex-Net 2.0 introduces a deep learning framework for robust robotic grasping, trained entirely on 6.7 million synthetic point clouds and corresponding analytic grasp metrics. This approach achieves an 80% success rate on 10 novel objects and 94% on 40 novel household objects, demonstrating effective generalization and significantly reducing the need for real-world data collection.

Key insights

• Large-scale synthetic data, informed by robust analytic grasp metrics and realistic noise modeling, can effectively train deep learning models for robotic grasping without real-world physical interaction data.
• Aligning the grasp candidate to the image center and axis simplifies the learning task for the CNN by removing the need for explicit rotational invariance.
• A deep learning approach trained on synthetic data can achieve planning speeds 3 times faster than point-cloud registration methods while maintaining comparable or superior grasp success rates, particularly on novel objects.

Summary by alphaXiv.