Grasp-O: A Generative System for Object-Centric 6-DoF Grasping of Unknown Objects

Generative models are a promising avenue for learning generalizable robotic tasks from data. A fundamental task that remains a challenge to autonomous manipulation is the 6-DoF grasping of unknown objects. This work proposes Grasp-O: a simple, fast, and robust system for general-purpose vision-based 6-DoF grasping applications. Our system is built using a powerful and efficient Variational Autoencoder (VAE) that learns a distribution of SE(3) grasp poses conditioned on object point clouds. The generative model is complemented by a grasp classification network that discriminates between good and bad grasp. We conduct extensive evaluations in simulation and the real world and demonstrate that our system outperforms existing VAE-based methods.