Dexterous grasp planning has long focused on physical stability, often overlooking how an object should be grasped to support downstream functional tasks. Conditioning grasp synthesis on human grasp taxonomies typically requires expensive, object-annotated datasets, limiting scalability.
To address this, researchers propose CoToGrasp, a novel generative framework that synthesizes diverse, stable grasps strictly conditioned on specific contact topologies. The key innovation is an object-agnostic training approach that bypasses the data collection bottleneck.
CoToGrasp introduces a feature-based canonical workspace that projects local object features into a unified gripper-centric domain. This decouples semantic functional intent from arbitrary object geometry, allowing the model to learn the intrinsic contact manifold of the gripper within this workspace.
Extensive evaluations on the large-scale DexGraspNet dataset demonstrate that CoToGrasp achieves state-of-the-art performance, outperforming existing taxonomy-guided planners. The researchers also validated the physical viability and kinematic feasibility of the synthesized contact topologies on a physical robot platform.
The code is available on the project website: https://cea-list.github.io/cotograspweb/.