CoToGrasp: Contact-Topology-Conditioned Dexterous Grasp Synthesis via Canonical Workspace Learning
CoToGrasp trains without object-annotated grasp data, then generates task-shaped dexterous grasps for unseen objects.
The paper frames the problem as more than grip stability: robots also need the right contact pattern for the intended function. Its method learns contact topologies in a gripper-centered canonical workspace, separating functional intent from object geometry. Evaluations on DexGraspNet report state-of-the-art results against taxonomy-guided planners. The authors also report physical robot tests showing the synthesized contacts are viable and kinematically feasible. HF Daily Papers' note
The paper frames the problem as more than grip stability: robots also need the right contact pattern for the intended function. Its method learns contact topologies in a gripper-centered canonical workspace, separating functional intent from object geometry. Evaluations on DexGraspNet report state-of-the-art results against taxonomy-guided planners. The authors also report physical robot tests showing the synthesized contacts are viable and kinematically feasible. HF Daily Papers' note
score 4