Grasping by interconnection : can robust and adaptive grasping emerge from minimal object information?

This paper investigates whether robust and adaptive dexterous grasping can emerge from minimal object information by modeling grasping as a dynamic interconnection between the robot and a simplified object representation. Instead of relying on precise object models or large-scale data-driven training, objects are approximated using coarse geometric primitives (sphere, cylinder, and box), each associated with a canonical grasp type from human grasp taxonomies. Grasp execution is formulated within the Virtual Model Control (VMC) framework, where virtual springs and dampers mechanically couple a virtual hand to a virtual object, allowing grasp motions to emerge naturally from the coupled dynamics rather than from predefined trajectories. Structured stiffness distributions and adaptive damping profiles promote coordinated, human-like finger closure while mitigating object ejection during transient phases. The approach is implemented on a Shadow Dexterous Hand and evaluated through robustness and generalization experiments under geometric and pose uncertainties, as well as on a diverse set of everyday objects. Results support the hypothesis that reliable dexterous grasping can arise from minimal yet structured object representations combined with dynamic object–robot interconnection.