Adversarial Grasp Objects

摘要

Learning-based approaches to robust robot grasp planning can grasp a wide variety of objects, but may be prone to failure on some objects. Inspired by recent results in computer vision, we define a class of “adversarial grasp objects that are physically similar to a given object but significantly less ”graspable” in terms of a specified robot grasping policy. We present three algorithms for synthesizing adversarial grasp objects under the grasp reliability measure of Dex-Net 1.0 for parallel-jaw grippers: 1) two analytic algorithms that perturb vertices on antipodal faces (one that uses random perturbations and one that uses systematic perturbations), and 2) a deep-learning-based approach using a variation of the Cross-Entropy Method (CEM) augmented with a generative adversarial network (GAN) to synthesize classes of adversarial grasp objects represented by discrete Signed Distance Functions. The random perturbation algorithm reduces graspability by 32%, 12%, and 32% for intersected cylinders, intersected prisms, and ShapeNet bottles, respectively, while maintaining shape similarity using geometric constraints. The systematic perturbation algorithm reduces graspability by 32%, 11%, and 21%; and the GAN reduces graspability by 22%, 36%, and 17%, on the same objects. We use the algorithms to generate and 3D print adversarial grasp objects. Simulation and physical experiments confirm that all algorithms are effective at reducing graspability.