DeformableGrasp: Efficient Dexterous Grasping of Deformable Objects via Reinforcement Learning

Abstract

Achieving stable and diverse dexterous grasping of deformable objects remains a fundamental challenge in robotics, due to high-dimensional action spaces and complex material properties. In this paper, we propose DeformableGrasp, a multimodal perception-guided reinforcement learning framework that enables adaptive grasp strategies across various object categories, particularly for deformable objects. First, we introduce a unified multimodal representation that integrates multiple perceptions to support robust grasp planning for deformable objects. Second, we propose an asymmetric reinforcement learning architecture that leverages privileged information (e.g., deformation, motion) during training to optimize grasp policies, thereby enhancing both generalization capability and sample efficiency. Third, we develop a flexible training strategy that leverages a diverse set of objects along with contact-rich, penetration-free, and kinematically feasible grasp poses, naturally adapting to the various physical properties while proactively avoiding unnecessary contacts. Extensive evaluations demonstrate that DeformableGrasp achieves highly robust performance across a wide spectrum of object categories and significantly advances the state of the art in dexterous grasping for deformable and compliant objects—even under perceptual uncertainty and real-world disturbances. DeformableGrasp achieves an average success rate of 93.1%, outperforming existing methods on the deformable object benchmarks.