TL;DR
DOGlove is a low-cost ($<600), open-source haptic glove that provides 21-DoF motion capture and 5-DoF force feedback for dexterous robotic teleoperation. By combining cable-driven force feedback with fingertip vibrations, it allows human operators to perform complex, contact-rich tasks—like rotating a carton in-hand or squeezing a bottle—even without visual feedback.
What Problem Does This Paper Address?
Achieving human-level dexterity in robots requires high-quality demonstration data, typically collected via teleoperation. However, researchers face a frustrating trade-off: vision-based systems (like using cameras to track hands) are cheap but lack the tactile "feel" necessary for contact-rich tasks, while professional haptic gloves (like the SenseGlove or Manus VR) often cost upwards of $5,000, making them inaccessible for many labs.
Without haptic feedback, a human operator cannot perceive an object's weight, friction, or softness, leading to failed grasps or crushed objects. DOGlove aims to bridge this gap by providing a high-performance, multi-modal feedback device that is affordable, easy to manufacture, and fully open-source.
Methods
The authors developed a sophisticated hardware-software stack designed for anthropomorphic precision:
- Kinematic Design: The glove features a 21-DoF structure that closely mirrors the human hand. It uses a customized joint structure with compact Alps rotary sensors and a 24-bit ADC, achieving an angular error of less than ±1° after calibration.
- Multi-Modal Feedback:
- Force Feedback (5-DoF): A cable-driven mechanism powered by Dynamixel servos provides active resistance to each finger, allowing the user to "feel" the solidity or softness of a virtual or remote object.
- Haptic Feedback (5-DoF): Linear Resonant Actuators (LRAs) at each fingertip provide high-frequency vibrations to signal contact events.
- Retargeting Framework: To map human motion to various robotic hands (like the LEAP or Shadow Hand), the system uses Forward Kinematics (FK) to determine human fingertip positions and Inverse Kinematics (IK) via the Mink library to calculate the corresponding robot joint angles.
- Hybrid Feedback Strategy: The authors implemented a specific thresholding strategy for force and haptics. For example, haptic vibrations trigger at low forces (10g) to signal initial contact, while active force feedback kicks in at higher loads (50g) to simulate object resistance.
Datasets
While the paper does not introduce a new static dataset, it uses DOGlove to collect demonstration data for Imitation Learning (IL). Specifically, they used the 3D Diffusion Policy (DP3) algorithm. The system was evaluated across several challenging benchmarks:
- Object Perception: Distinguishing between pairs of objects (e.g., a soft toy cabbage vs. a real cabbage) based solely on touch.
- Bottle-Slipping: Controlling a bottle to slip through fingers to a specific distance without dropping it.
- In-Hand Rotation: Rotating a carton 45 degrees and placing it in a bucket.
- IL Tasks: "Press and Move Box" and "Pick and Place Teddy Bear."
Key Insights and Ideas
- The "Misleading Sensation" Threshold: The authors discovered that continuous haptic vibration during high-force tasks can actually interfere with a user's perception. They implemented a "third threshold" (100g) where haptic vibration stops, leaving only the force feedback active, which improved task performance.
- Force vs. Haptics: The user study revealed that force feedback is superior for perceiving softness, while haptic vibration is more effective for perceiving shape and contact timing.
- Accessibility as a Feature: By using 3D-printed PETG and off-the-shelf electronics, the glove can be assembled in just 6 hours. This lowers the barrier for entry for labs wanting to explore dexterous manipulation.
Key Results
DOGlove significantly outperformed vision-based baselines and "no-feedback" setups:
- In-Hand Rotation: Achieved a 100% success rate with combined haptic and force feedback, compared to only 40% with no feedback and a dismal 10% using the vision-based AnyTeleop baseline.
- Blind Operation: In the bottle-slipping task without visual feedback, force feedback improved the success rate from 50% to 90%.
- Imitation Learning: Policies trained on just 40 DOGlove demonstrations achieved an 85% success rate on the "Press and Move Box" task and 70% on the "Pick and Place Teddy Bear" task.
- Latency: The system maintains a responsive 30 Hz frequency for haptic feedback and up to 120 Hz for motion capture, ensuring a smooth teleoperation experience.
Limitations and Future Work
Despite its success, the authors note a few areas for improvement:
- Weight: At 550g, the glove is relatively heavy due to the five commercial servos, which can lead to user fatigue.
- Agility: The maximum speed and torque of the servos limit performance in very high-speed, agile tasks.
- Ergonomics: While modular, the current design may not be perfectly comfortable for all hand sizes.
Future work will focus on designing custom, lighter reduction mechanisms to balance speed and torque, and providing CAD files for multiple linkage sizes to improve the fit for a wider range of users.
Summary updated