PaperlayerView exact revision on arXiv

Source-linked AI summary

DOGlove: Dexterous Manipulation with a Low-Cost Open-Source Haptic Force Feedback Glove

Han Zhang, Songbo Hu, Zhecheng Yuan, Huazhe Xu

arXiv:2502.07730v1cs.RO

TL;DR

Dexterous-hand teleoperation needs affordable, precise motion capture and richer sensory feedback than vision alone provides. DOGlove addresses this gap with an open-source haptic glove and retargeting framework, achieving successful contact-rich manipulation and imitation-learning data collection. Its main scope limits are weight, speed and torque constraints, and possible discomfort for some users.

  • Problem

    Dexterous-hand teleoperation and imitation learning require precise, high-quality demonstrations, while existing systems can be expensive and visual feedback provides limited interaction information.

  • Method

    DOGlove combines 21-DoF motion capture, 5-DoF force feedback, fingertip haptics, action and haptic-force retargeting, and teleoperated demonstration collection.

  • Results

    Across 10 trials, the imitation-learning policy for rotating and placing a carton achieved a 90% success rate (9/10), while user studies showed object-property perception without visual feedback.

  • Takeaways & Limitations

    DOGlove provides precise haptic perception for contact-rich teleoperation and integrates with existing robot-learning methods for collecting demonstrations.

  • Takeaways & Limitations

    DOGlove weighs 550g, its agile teleoperation is constrained by servo speed and torque, and it may be uncomfortable for some users.

Abstract

from arXiv · show

Dexterous hand teleoperation plays a pivotal role in enabling robots to achieve human-level manipulation dexterity. However, current teleoperation systems often rely on expensive equipment and lack multi-modal sensory feedback, restricting human operators' ability to perceive object properties and perform complex manipulation tasks. To address these limitations, we present DOGlove, a low-cost, precise, and haptic force feedback glove system for teleoperation and manipulation. DoGlove can be assembled in hours at a cost under 600 USD. It features a customized joint structure for 21-DoF motion capture, a compact cable-driven torque transmission mechanism for 5-DoF multidirectional force feedback, and a linear resonate actuator for 5-DoF fingertip haptic feedback. Leveraging action and haptic force retargeting, DOGlove enables precise and immersive teleoperation of dexterous robotic hands, achieving high success rates in complex, contact-rich tasks. We further evaluate DOGlove in scenarios without visual feedback, demonstrating the critical role of haptic force feedback in task performance. In addition, we utilize the collected demonstrations to train imitation learning policies, highlighting the potential and effectiveness of DOGlove. DOGlove's hardware and software system will be fully open-sourced at https://do-glove.github.io/.

1 Tsinghua University, 2 Shanghai Qi Zhi Institute, 3 Shanghai AI Lab https://do-glove.github.io/

DOGlove addresses dexterous-hand teleoperation challenges by combining precise motion capture with multimodal haptic feedback in a low-cost, open-source glove. Its retargeting framework supports contact-rich manipulation, evaluation without visual feedback, and imitation-learning data collection.

  • The system supports complex contact-rich demonstrations, evaluates haptic feedback without visual input, and collects demonstrations for imitation-learning policies.
  • Dexterous-hand teleoperation requires intuitive, responsive, and precise motion capture because high DoFs and hand complexity increase operational demands.
  • Vision-based tracking can be inaccurate or obstructed during hand-object interactions, while visual feedback alone makes intuitive control challenging.
  • DOGlove combines 21-DoF anthropomorphic motion capture, 5-DoF cable-driven force feedback, and fingertip linear resonant actuators.
  • Action retargeting maps glove fingertip positions to robotic hands, while haptic force retargeting conveys contact information during teleoperation.

II. RELATED WORK

Prior work uses videos, trajectories, tracking systems, and teleoperation hardware to collect dexterous-manipulation data, but visual gaps and limited sensory richness remain challenges.

  • Imitation learning requires substantial task-specific, high-quality demonstrations, making efficient data collection a persistent challenge.
  • Video and trajectory demonstrations are accessible but can transfer poorly because recorded human visuals differ substantially from robot perception.
  • Existing dedicated grippers mainly target parallel grippers, while MoCap systems use camera, glove-based, marker-based, or commercial tracking.
  • Teleoperating dexterous hands remains difficult despite leader-follower, VR, and commercial motion-capture approaches.
  • Haptic feedback enriches sensory information beyond vision and can improve immersion and perception of robot status and movement.

III. GLOVE DESIGN OBJECTIVES

DOGlove is designed around accessibility, manufacturability, and responsive operation, targeting a low-cost alternative to expensive commercial force-feedback gloves.

  • DOGlove is designed to capture hand poses and provide haptic force feedback for intuitive teleoperation.
  • The design prioritizes low cost, ease of manufacturing, and high performance for accessibility by the research community.
  • Under 600 USD, DOGlove costs far less than commercial products priced above 5,000 USD.
  • The glove uses commodity 3D printing and readily sourced components, and the complete system can be assembled within 6 hours.
  • Joint encoders provide angle data within ±7.2° error, while the haptic system supports multiple waveforms and operation at a minimum frequency of 30 Hz.

IV. HARDWARE SYSTEM

DOGlove’s hardware models human-hand kinematics with modular, encoder-equipped rotary joints and adjustable linkages designed for comfortable, accurate motion capture.

  • Kinematic Design: The glove’s kinematic design uses constraints and an anthropomorphic structure to emulate natural human-hand motion while supporting comfort and precise motion capture.
  • Kinematic Design: Finger DIP and PIP joints are 1-DoF hinges, whereas MCP joints are 2-DoF ball joints supporting flexion-extension and adduction-abduction.
  • Kinematic Design: The thumb uses 1-DoF IP and MCP hinges, a 2-DoF TM ball joint, and an additional wrist DoF for pronation-supination.
  • Kinematic Design: DOGlove implements hinges with rotary-linked mechanisms and ball joints with two orthogonal rotary joints, each equipped with encoders.
  • Finger Design: Modular finger assemblies share a common framework, while linkage lengths are simulated and customizable to accommodate adult hand sizes and preserve motion.
  • Finger Design: Each rotary joint uses a shoulder screw, bearing, encoder, and locknut, with PETG-printed finger bodies for fabrication and durability.
  • Finger Design: The finger assembly is constrained below 26 mm in width and requires at least 0.5 N·m stall torque for effective force feedback.

1) Joint Encoders:

DOGlove uses compact, low-cost rotary encoders integrated into its finger linkages to capture joint motion. Calibration reduces encoder-related angular error from ±7.2° to within ±1°.

  • 1) Joint Encoders:: Sixteen compact rotary encoders combined with five servo motors provide 21-DoF motion capture.The encoders are integrated into 3D-printed joint structures and their voltage output is converted through an ADC and MCU pipeline.
  • 1) Joint Encoders:: The encoder voltage is mapped directly to joint angle under the assumption that 0 V corresponds to 0° and approximately 3.3 V corresponds to 360°.This mapping uses the ADC output voltage as the joint-angle signal.
  • 1) Joint Encoders:: ±2% encoder linearity error produces ±7.2° angular error before calibration.An external high-precision joint encoder is used to build a correction table for each encoder.
  • 1) Joint Encoders:: Calibration reduces the measured joint-angle error to within ±1°.The correction table maps each encoder’s voltage reading to the corresponding actual joint angle.

2) Cable-Driven Force Feedback Structure:

DOGlove uses a compact pulley and cable mechanism to transmit bidirectional torque despite misaligned servo and finger-joint axes. The design avoids the weight and cost of using two servos per finger while supporting fingertip tactile feedback.

  • 2) Cable-Driven Force Feedback Structure:: A cable-driven mechanism transfers servo torque across misaligned rotary axes in a compact design.Compared with bevel gears, cables avoid the space demands and potential shaft deformation associated with transmitting large torque through orthogonal gears.
  • 2) Cable-Driven Force Feedback Structure:: Traditional cable systems provide unidirectional tension and use springs for opposite-direction force, producing unrealistic feedback sensations.Using two servos per finger could address the issue but would increase glove weight and cost.
  • 2) Cable-Driven Force Feedback Structure:: DOGlove’s pulley system provides bidirectional torque transmission through a 0.6 mm stainless steel braided cable and a 1:1 pulley ratio.Fixed pulleys redirect the cable path to reduce transmission friction.
  • 2) Cable-Driven Force Feedback Structure:: Clockwise and counterclockwise servo rotation shift cable tension between Cable B and Cable A to drive finger-joint motion.This arrangement produces flexion movement when tension shifts to Cable A.
  • 2) Cable-Driven Force Feedback Structure:: The pulley configuration enables bidirectional torque transmission in a simple, compact, and cost-effective design.This mechanism supplies force feedback without the two-servo-per-finger arrangement described as heavier and more expensive.
  • 2) Cable-Driven Force Feedback Structure:: Each fingertip includes a linear resonant actuator that provides precise vibration stimuli for tactile feedback.The LRAs are 8 mm in diameter, 2.5 mm high, and resonate at approximately 240 Hz along the fingertip-normal Z axis.

C. Wrist Localization

DOGlove supports wrist localization and fingertip-focused action retargeting across different robotic hands. Its rigid anthropomorphic structure and scaling-based inverse kinematics enable proportional gesture transfer and precise fingertip alignment.

  • C. Wrist Localization: DOGlove accommodates external wrist-localization devices through a shell with a 1/4 inch screw connector.The experiments use an HTC Vive Tracker for real-time wrist-position tracking, while other devices can also be used.
  • C. Wrist Localization: Action retargeting computes human fingertip positions with forward kinematics and maps them to robotic-hand positions with inverse kinematics.The method focuses on fingertips because they are the primary contact area during object interactions.
  • C. Wrist Localization: A scaling factor in inverse kinematics adapts fingertip mapping to size differences between the human hand and target robotic hand.This allows the robotic hand to mirror the operator’s gestures proportionally.
  • C. Wrist Localization: Precise fingertip alignment supports grasping small objects by making corresponding robotic fingertips touch when the operator brings them together.The retargeting is evaluated on the LEAP Hand in the real world and on the Shadow, Inspire, and Allegro Hands in simulation.

B. Haptic Force Retargeting

DOGlove retargets fingertip force and haptic feedback from a robotic hand to the operator using sensor thresholds and actuator control. The combined strategy supports distinguishing object properties without vision and improves complex contact-rich manipulation.

  • B. Haptic Force Retargeting: Each LEAP Hand fingertip uses a 1-D force sensor with a 3 kg measurement range and 1 g precision.The quantitative experiments use these measurements to select a combined haptic and force-feedback strategy.
  • B. Haptic Force Retargeting: A 10 g threshold initiates haptic feedback, while a 50 g threshold is used for more realistic force feedback after operators’ sensitivity is considered.The thresholds are part of the combination strategy summarized in Table I.
  • B. Haptic Force Retargeting: Force readings are clamped to [0g, 3000g] and mapped linearly to the Dynamixel servos’ KP gain.The servos operate in current-based position-control mode, while haptic feedback uses waveform ID 56, Pulsing Sharp 1-100%.
  • B. Haptic Force Retargeting: The combined haptic force-retargeting strategy enables operators to distinguish object shape, size, and softness without visual feedback.The paper reports that it also improves performance in complex, contact-rich manipulation tasks.

VI. EXPERIMENTS

The experiments evaluate DOGlove’s haptic force perception, teleoperation efficiency, and compatibility with imitation learning. A user study tests object discrimination without visual or auditory feedback using force and haptic feedback conditions.

  • Evaluation Scope: DOGlove is evaluated across haptic force perception, teleoperation efficiency, and imitation-learning compatibility.The evaluation includes a user study, quantitative experiments, teleoperation tasks, and downstream policy training.
  • User Study: Five untrained operators distinguish five object pairs using only DOGlove feedback while wearing an eyemask and headphones.The pairs vary in shape, size, and softness, and performance is measured by success rate.
  • User Study: All participants effortlessly distinguish basic object pairs 1–3 without visual or auditory feedback.For challenging pair 4, most participants perceive softness using force feedback, while some use haptic feedback during deformation.
  • User Study: For challenge pair 5, deformation of the softer toy cabbage makes softness difficult to distinguish using force feedback alone.The deformation makes the toy resemble the size of a real cabbage.
  • User Study: Combining haptic and force feedback slightly reduces sensitivity for both challenging pairs, producing marginally lower accuracy.This result contrasts with the separate strengths of force feedback for softness perception and haptic feedback for shape perception.

B. Bottle-Slipping

Bottle-slipping experiments examine teleoperation without and with visual feedback, while carton rotation tests a longer contact-rich manipulation. Force feedback improves blind slipping and precision, whereas combined haptic force feedback benefits carton-task success and completion time.

  • Bottle-Slipping Without Visual Feedback: Force feedback significantly improves bottle-slipping success without visual feedback, while additional haptic feedback further enhances overall performance.Because the robotic fingers remain continuously in contact with the bottle, haptic feedback provides no additional information in this condition.
  • Bottle-Slipping Without Visual Feedback: Retargeting differences make small human finger-position changes produce large LEAP Hand deviations, causing AnyTeleop to struggle with slipping.The cited result links the difficulty to differences in retargeting strategies rather than to the bottle task alone.
  • Bottle-Slipping With Visual Feedback: With visual feedback, force feedback minimizes slipping deviation more effectively, although haptic feedback may interfere with task precision.DOGlove used solely for motion capture achieves the same success rate as haptic force feedback but produces a larger average deviation.
  • Rotating and Placing the Carton: The carton task requires rotating the object more than 45 degrees and placing it into a small bucket without dropping it.Operators must manage hidden finger–carton contact points caused by visual obstacles.
  • Rotating and Placing the Carton: Haptic and force feedback significantly improve carton-rotation success and reduce completion time in the long-horizon contact-rich task.Haptic force feedback achieves a higher success rate, while force feedback alone has a comparable average completion time.

D. Imitation Learning

The imitation-learning experiment uses demonstrations collected with DOGlove to train policies for dexterous downstream tasks. The reported results include 70% success for one task and 90% for carton rotation and placement, while the system remains limited by weight, actuator performance, and fit.

  • Method: DOGlove-collected demonstrations train 3D Diffusion Policies for downstream dexterous manipulation tasks.Point-cloud inputs from a Realsense L515 are downsampled to 1024 points using farthest point sampling.
  • Results: 70% success (14/20) is achieved across trials, with failures primarily caused by the teddy bear slipping from an insufficiently firm grasp.The result demonstrates policy performance on a grasping task using human-collected demonstrations.
  • Results: 90% success (9/10) is achieved for rotating and placing the carton after training with 3 human-collected demonstrations.This is evaluated on the same contact-rich setup described for the teleoperation experiment.
  • Limitations: DOGlove weighs 550g because it uses five commercial servos, and agile teleoperation is constrained by their maximum speed and torque output.The authors are investigating lighter servos and a customized reduction mechanism to balance speed and torque.
  • Limitations: Some users may find DOGlove uncomfortable despite its accommodation of most hand sizes.The authors are developing multiple linkage sizes to improve customization and wearability.
Loading 2502.07730v1…