Object‑agnostic planners grasp unseen objects zero‑shot

Grasp planners using object models are being replaced by generative approaches that infer contact surfaces on-the-fly. These generative planners learn the geometry of a gripper's contact area and can generate successful grasps for unfamiliar objects in a single step. Previously, data-driven grasp synthesis needed large, specific datasets, often created through tedious annotation or costly simulations.

These models struggled to generalize to new shapes and required significant engineering effort to maintain per-object training data. GOAG, a generative and object-agnostic planner, achieved an 86.93% success rate on the MultiDex benchmark, even without any object-specific data during training. When trained in an object-agnostic way, GOAG outperformed prior baselines on all three grippers.

Creating GOAG's full dataset required only about 1 GPU hour on a single RTX 4090, while similar object-specific methods needed around 1,400 GPU hours on an A100. GOAG's entire dataset creation consumed roughly 1 GPU hour on a single Nvidia RTX 4090, whereas previous methods required around 1,400 GPU hours on an Nvidia A100. CoToGrasp sets a new standard on the large DexGraspNet dataset, surpassing existing taxonomy-guided methods while staying completely object-agnostic.

By learning the gripper's contact manifold in this workspace, CoToGrasp can generalize to unseen objects at inference time. Both studies primarily evaluated their methods on static benchmarks, using simulated depth or point-cloud inputs. In real-world settings with clutter, occlusions, and dynamic interactions, additional work is needed to confirm robustness.

The benchmarks MultiDex and DexGraspNet should include an object-agnostic baseline, as these results show that eliminating per-object data does not compromise grasp success and drastically reduces dataset generation time.