Learning Multi-Object Dense Descriptor for Autonomous Goal-Conditioned Grasping
School of Control Science and Engineering, Shandong University
Abstract
In a goal-conditioned grasping task, a robot is asked to grasp the objects designated by a user. Existing methods for goal-conditioned grasping either can only handle relatively simple scenes or require extra user annotations. This paper proposes an autonomous method to enable the grasping of target object in a challenging yet general scene that contains multiple objects of different classes. It can effectively learn a dense descriptor and integrate it with a newly designed grasp affordance model. The proposed method is a self-supervised pipeline trained without any human supervision or robotic sampling. We validate our method via both simulated and real-world experiments while the training relies only on a variety of synthetic data, demonstrating a good generalization capability.
See details about the generation process of synthetic scene data in here or view PDF.
All the data and code will be released soon!