Grasp planners that hinge on object models are being eclipsed by approaches that synthesize contacts directly on the hand. A new generation of generative planners learns only the geometry of a gripper’s contact surface and can produce viable grasps for any unseen item in a single inference step.
Before these works, data‑driven grasp synthesis required large, object‑specific datasets collected either from painstaking annotation or from expensive simulation pipelines. The resulting models generalized poorly to novel shapes and incurred substantial engineering overhead to maintain per‑object training corpora.
GOAG reaches an average success rate of 86.93% on the MultiDex benchmark despite never seeing object‑specific data during training. While being trained in an object‑agnostic manner, GOAG achieved a higher average success rate for generating accurate grasps across all three grippers compared to the baselines [1].
The entire dataset creation for GOAG consumed only about 1 GPU hour on a single Nvidia RTX 4090, whereas comparable object‑specific pipelines reported generation times of roughly 1,400 GPU hours on an Nvidia A100. “The entire dataset creation required approximately 1 GPU hour on a single Nvidia RTX 4090.” “In contrast, a previous work [12] reported a much longer generation time of 1,400 GPU hours using Nvidia A100.” [1]
CoToGrasp attains state‑of‑the‑art performance on the large‑scale DexGraspNet dataset, surpassing existing taxonomy‑guided planners while remaining completely object‑agnostic. “Extensive evaluations on the large‑scale DexGraspNet dataset demonstrate that CoToGrasp achieves state-of-the-art performance, outperforming existing taxonomy-guided planners.” “By learning the intrinsic contact manifold of the gripper within this workspace, our model achieves zero-shot generalization to unseen objects at inference.” [2]
Both papers evaluate primarily on static benchmark collections and rely on simulated depth or point‑cloud inputs; real‑world clutter, occlusions, and dynamic interactions remain underexplored. This suggests that while contact‑surface learning removes the need for object‑specific datasets, additional research is required to confirm robustness under uncontrolled perception conditions.
Benchmarks such as MultiDex and DexGraspNet should incorporate an object‑agnostic baseline, because these results demonstrate that eliminating per‑object data does not sacrifice grasp success and dramatically cuts dataset generation time.
