Grasping by interconnection: robust manipulation with minimal object information

Dexterous grasping in unstructured environments remains a central open challenge in robotics, requiring simultaneous robustness to perceptual uncertainty and generalization across diverse object geometries. Existing approaches rely on accurate object models, either predefined or reconstructed online, making them brittle to the sensor noise inherent in real-world settings. We propose a grasp planning strategy that couples the robotic hand to minimal primitive-based object representations. Through virtual mechanical interconnections, each primitive shape corresponds to a grasp type and generates fully adaptive closing trajectories. Experimental evaluation across a wide set of objects demonstrates robustness to pose estimation errors of several centimeters in translation and several tens of degrees in orientation. This work demonstrates that robust, task-oriented dexterous grasping can emerge from minimal object information, opening a path toward reliable manipulation in unstructured real-world scenarios.