DEXOP: Hand Exoskeleton System for Teaching Robot Dexterous Manipulation In-The-Wild

Hao-Shu Fang, Branden Romero, Arthur Hu, Edward H Adelson, Pulkit Agrawal

Keywords: Dexterous Manipulation, Exoskeleton, Data Collection

TL;DR: A passive hand exoskeleton to collect dexterous manipulation data with rich tactile and accurate action info, 0 embodiment gap with real robot.

Abstract: We introduce DEXOP, a novel passive hand exoskeleton system designed to collect robots dexterous manipulation in-the-wild, without needing a real robot. Traditional teleoperation systems for high DoF dexterous hands are usually expensive and limited by the lack of intuitive feedback to human teleoperator. DEXOP allows human users to directly control a robot’s dexterous hand through a passive exoskeleton system, where the human fingers are mechanically connected to the robot fingers, controlling the robot hand while also feeling the force applied to the robot hand seamlessly. By optimizing the kinematic configuration and providing high force transparency, human users can control a robot's hand just like controlling their own hand. Equipped with precise position encoders and tactile sensors, DEXOP captures high-fidelity dexterous manipulation data, facilitating manipulation learning without the need for costly hardware or careful engineering. We evaluate the system across multiple dexterous tasks, demonstrating its capability to accomplish highly dexterous, contact-rich tasks and its potential to scale the collection of high-quality demonstration data. Learning experiments show significant improvements in the performance-time ratio compared to teleoperation method, making DEXOP a powerful tool for advancing robot dexterity.

Submission Number: 20