New AI grasp planners handle unseen objects without object-specific training data

Two new robotic grasp planning systems, GOAG and CoToGrasp, can generate viable grasps for objects they have never encountered by learning only the geometry of a gripper's contact surface. GOAG achieved an 86.93% average success rate on the MultiDex benchmark while using no object-specific training data, outperforming existing baselines across multiple gripper types. Its dataset was created in roughly 1 GPU hour on a single Nvidia RTX 4090, compared to approximately 1,400 GPU hours required by a comparable object-specific pipeline. CoToGrasp similarly achieved state-of-the-art results on the large-scale DexGraspNet dataset, surpassing taxonomy-guided planners while remaining fully object-agnostic. Both systems currently rely on simulated inputs and static benchmarks, and their robustness in real-world, cluttered, or dynamic environments has yet to be thoroughly tested.