Multifingered robotic grasping remains a challenge for deep learning models, which often struggle to generalize to new objects due to reliance on limited, object-specific datasets. A new approach, called GOAG (Generative and Object-Agnostic Grasp Planner), tackles this by focusing on the gripper's contact surface distribution rather than object geometry.
The key insight is that at the moment of a successful grasp, the gripper and object share identical surface geometry at their contact points. GOAG leverages this by training exclusively on gripper geometry, learning a compact latent representation that enables efficient sampling of valid grasps for unseen objects at inference time.
In experiments on the MultiDex dataset, GOAG achieved an average success rate of 86.93%, matching or exceeding leading methods that require object-specific training. It also offers significantly faster processing when generating many grasps, highlighting the advantages of object-agnostic learning.
The method was validated in both simulated and real-world scenarios across different grippers from the literature. Code and videos are available on the project website: https://cea-list.github.io/goagweb/.