| CARVIEW |
Hardware Design
To ensure precise MoCap capabilities and a comfortable wearing experience, DOGlove is designed to closely resemble the anthropomorphic structure of the human hand. It features 21-DoF motion capture and 5-DoF haptic force feedback.
Finger Assembly
DOGlove is composed of the five finger assemblies, along with the palm base structure. The design of each finger assembly follows a modular approach, ensuring consistent structural elements across all fingers. The exploded view of a single finger is illustrated in the figure below. The highlighted area indicates the basic components of a rotary joint.
DOGlove leverages a cable-driven mechanism to deliver force feedback to each finger while maintaining a compact and cost-effective design. Additionally, each fingertip is also equipped with a linear resonant actuator (LRA) to provide realistic haptic feedback. This integration of force and haptic feedback creates an immersive and responsive interface for dexterous manipulation.
Action Retargeting
We apply the 5-DoF haptic force feedback to the human operators’ fingertips and adopt a retargeting method focused on fingertip positions. Our approach combines Forward Kinematics (FK) to compute human fingertip positions and Inverse Kinematics (IK) to calculate the corresponding robotic hand positions.
We deploy the system on the LEAP hand in real-world scenarios and test it with various robotic hands in the MuJoCo simulator.
Teleoperation Demo
Rotate and place a milk carton.
Pick and place a teddy bear.
How Does Haptic Force Feedback Enhance Teleoperation?
We conduct a series of quantitative experiments and a user study.
In the above experiment, the human operator must perform a bottle-slipping action using only feedback from DOGlove.
A trial is denoted successful if the bottle slips without falling.
Without visual cues, force feedback significantly improves the success rate.
Adding haptic feedback further enhances overall performance.
To further evaluate control accuracy, when visual input is allowed, the operator must slip the bottle to a precise target distance of 9 cm.
We then measure the deviation between the actual slipping distance and this target.
When visual input is available, force feedback helps operators reduce slipping deviation more effectively.
Haptic force feedback enables operators to achieve a higher success rate and a faster average completion time, as haptic feedback provides contact information, while force feedback indicates the proper timing for in-hand rotation.
User Study
Five untrained human operators participate in this user study. They must distinguish between five pairs of objects using only haptic force feedback from DOGlove, without any visual or auditory input.
Basic Pair: Different shape.
The ball and the box have distinctly different shapes.
Basic Pair: Similar softness and shape, different size.
The two toys have similar softness and shapes but vary in size.
Basic Pair: Similar shape, different size.
The peanut bottle and the coffee paper cup share a similar cylindrical shape, but their diameters differ slightly.
Challenging Pair: Similar size and shape, different softness.
Two identical bottles, one filled with pure water (soft) and the other filled with shaken carbonated cola (hard).
Challenging Pair: Similar shape, different size and softness.
A toy cabbage (softer, larger) and a real cabbage (harder, smaller, and more compact).
Imitation Learning with DOGlove
DOGlove enables the collection of high-quality demonstrations.
3D Diffusion Policy (DP3) is selected as our imitation learning algorithm, and we use Realsense L515 to acquire the point cloud inputs, which are then downsampled to 1024 points using farthest point sampling. The data collected by DOGlove is used to train policies for various downstream tasks. We evaluate imitation learning performance on 2 contact-rich tasks.
Press and Move Box: During data collection, the box is randomly placed within a 30×20 cm area,
and DOGlove collects 40 demonstrations to train the policy.
In evaluation, the box is also randomly placed in the same area.
Across 20 trials, the success rate is 85% (17/20).
Pick and Place Teddy Bear: During data collection, the teddy bear’s initial position is randomized within a 30×20 cm area,
and DOGlove collects 40 demonstrations to train the policy.
In evaluation, the bear is again randomly placed in the same area.
Across 20 trials, the success rate is 70% (14/20).
BibTeX
@article{zhang2025doglove,
title={DOGlove: Dexterous Manipulation with a Low-Cost Open-Source Haptic Force Feedback Glove},
author={Zhang, Han and Hu, Songbo and Yuan, Zhecheng and Xu, Huazhe},
journal={arXiv preprint arXiv:2502.07730},
year={2025}
}Contact
If you have any questions, please feel free to contact Han Zhang.
Acknowledgement
We would like to thank Zhengrong Xue, Gu Zhang, Changyi Lin, Mengda Xu, and Yifan Hou for their invaluable advice and fruitful discussions on hardware design and learning policies. We also appreciate Wenhao Ding and Laixi Shi for their insightful discussions and feedback. Additionally, we thank Yichuan Gao, Xiaoyan Yang, Xinyao Qin, and Botian Xu for their assistance with the user study. Special thanks to Skyentific, Gennady Plyushchev, for their innovative contributions to the unconventional cable-driven joint design. We are also grateful to Tiansheng Sun and Guanghan Pan for their open-source repository of the HTC Vive Tracker Python API. We extend our appreciation to Yitong Wang for her help in creating elegant graphic renderings of the hardware design. Tsinghua University Dushi Program supports this project.