A Generalizable Gripper Interface for Hardware-Agnostic Policy Learning in Robotic Manipulation

Alejandro Gonzalez Hernandez, Mario Cavero-Vidal, Alberto San-Miguel-Tello, Oscar Palacin

Keywords: Hand-held grippers, Learning from Demonstration

TL;DR: This paper presents a new hardware interface enabling easy integration of various robotic grippers, including switch-actuated ones. It enhances demonstration capture by directly recording gripper states, overcoming limitations of vision-only methods.

Abstract: Leveraging on the recent advances on the Universal Manipulation Interface (UMI) solution for cost-effective and easy human demonstration acquisition for robot learning, the paper at hand introduces a generalizable hand-held gripper implementation that broadens its usage to any type of robotic gripper. Proposed solution consists on a hardware implementa- tion that directly captures gripper state together with an interface that ensures synchronization by embedding into video acquisition. Besides tackling the issues from the default vision-based approach for extracting gripper width, i.e. occlusions and additional computation load, this approach allows to consider different gripper operation modes such that any gripper configuration can be integrated, which has been exemplified in the paper for switch-actuated ones. For this purpose, a mechanical design has been also proposed to quickly change between different grippers adaptations using the standard UMI design. Solution performance results are presented together with its application on a manufacturing use case introducing switch-actuated grippers.

Submission Number: 7