The N2D Haptic Glove: A Multi-Finger Glove for 2D Directional Force Feedback for Contact-Rich Manipulation

Authors Redacted

Anonymous Submission

Under Review

The N2D Haptic Glove worn on a hand — **The N2D Haptic Glove** — the first multi-fingered glove design that offers 2D force feedback along the axial direction of multiple fingers, enabling probing, gentle grasping, and slip detection.

Abstract

Humans rely on directional fingertip forces to probe and regulate contact during manipulation, yet most wearable haptic gloves render only vibration or single-axis force, leaving force direction ambiguous. Without directional cues, users must infer contact force from vision alone, often leading to over-pressing, inconsistent control, and reduced precision in robotic teleoperation. We present the N2D Haptic Glove, a multi-finger wearable device that renders planar flexion–extension fingertip forces using capstan-drive transmissions for high-transparency force feedback. Through benchtop validations and a user study involving haptic teleoperation of a robotic arm and hand, we demonstrate that compared to visual-only and single-axis baselines, planar fingertip feedback significantly reduces contact force error during precise manipulation, improves trial-to-trial consistency, and enhances overall user experience in axial probing tasks.

Contributions

The first multi-fingered haptic glove capable of delivering planar force feedback to the fingertip.
A mechanical design that achieves high transparency through low-friction, zero-backlash transmissions for clean, accurate haptic rendering.
Validation of 2D planar directional feedback ability with benchtop testing.
A teleoperation user study demonstrating the benefits of N2D’s directional haptics over no-haptic and 1D-haptic baselines.

Hardware Design

The N2D Haptic Glove is designed to preserve the natural biomechanical DOFs of the hand. The MCP joint is treated as a saddle joint allowing flexion–extension and adduction–abduction; PIP and DIP joints are modeled as revolute joints constrained to flexion–extension. The thumb additionally performs opposition at the TMC joint. The thumb, index, and middle fingers — central to most grasping and probing tasks — are actuated.

The device consists of three modular finger subsystems mounted along a structural backbone. Each subsystem uses a crossed four-bar linkage with an additional link to permit full flexion–extension while enabling 2D planar force feedback. Planar fingertip forces are generated by actuating the proximal links via capstan cable drives. Each linkage is mounted on a passive rotary joint for finger abduction–adduction, with the thumb having an additional passive joint for TMC opposition.

All active DOFs are driven by GB2208 gimbal motors through capstan transmissions, avoiding geared transmissions that introduce backlash, friction, and damping. Actuators are mechanically grounded at the base of the linkage, preventing motor masses from rotating about the finger joints. The total glove weighs 562 g (battery off-hand).

Deconstructed view of the N2D Haptic Glove finger linkage — Deconstructed view of the N2D Haptic Glove showing finger linkage joint layout and its correspondence to hand anatomy.

Comparison to Existing Fingertip-Feedback Systems

System	Multi-DOF Forces	Kinesthetic	Multi-Digit
SenseGlove R1	✗	✓	✓
HaptGlove	✗	✓	✓
DOGlove	✗	✓	✓
NURing	✓	✓	✗
Fluid Reality	✗	✗	✓
AirPush	✓	✗	✗
N2D Glove (ours)	✓	✓	✓

System Architecture

The Glove PC computes finger poses from encoder data and sends joint commands to the Simulation/Robot PC via ROS. Each motor–encoder pair interfaces with a Teensy 4.1 microcontroller running field-oriented control in torque mode at 1 kHz, while higher-level torque commands stream from the Glove PC at 10 Hz over a serial connection. An external VIVE Tracker 3.0 provides absolute wrist pose, fused with fingertip pose estimates. Sensed interaction forces are returned through ROS and mapped to motor commands, closing the bilateral control loop.

Kinematic Modeling

Single finger linkage with three DOFs and labeled coordinate frames — Kinematic analysis of a single finger linkage with three DOFs defined by $q_0, q_1, q_2$ ( $q_3$ being a driven joint).

Each finger is modeled as a rigid linkage with three revolute joints $q_{f,0}, q_{f,1}, q_{f,2}$ and link lengths $a_{f,0}, a_{f,1}, a_{f,2}, a_{f,3}$ . The driven angle $q_{f,3}$ is solved via the cosine law:

CD = \ell_4 = \sqrt{a_2^2 + \ell_5^2 + 2 a_2 \ell_5 \cos q_{f,2}}

Directional haptics are then rendered by mapping Cartesian forces in the world frame, $\mathbf{f}^w_f \in \mathbb{R}^3$ , to joint torques $\boldsymbol{\tau}_f \in \mathbb{R}^2$ via the Jacobian transpose:

\boldsymbol{\tau}_f = \mathbf{J}^w_f(\mathbf{q}_f)^\top \mathbf{f}^w_f

Because each finger has only two active DOF, the achievable Cartesian forces lie in a configuration-dependent 2D subspace of $\mathbb{R}^3$ . The Jacobian-transpose mapping naturally projects any commanded force onto this subspace; the un-actuated direction (finger abduction/adduction) is not reflected.

Validation Experiments

A voltage–torque mapping is identified per joint via a quasi-static calibration. Circular planar force trajectories were commanded and measured with an ATI Axia80 3-DOF sensor.

RMSE_x = 0.032 N
RMSE_z = 0.014 N
RMSE_|e| = 0.034 N, RMSE_θ = 2.81°

The fingertip Jacobian is well-conditioned with $\kappa(\mathbf{J}^w_f) \approx 2.93$ , and the manipulability $w(\mathbf{q}_f) \geq 0.82$ over 58% of sampled configurations — confirming that directional forces can be rendered cleanly across most reachable hand postures.

Force calibration setup with finger linkage mounted on an ATI force sensor — Force calibration setup showing the world coordinate frame ( $x_{\text{world}}$ , $z_{\text{world}}$ ) and the tip coordinate frame ( $x_{\text{tip}}$ , $y_{\text{tip}}$ ).

Force tracking results: commanded vs. measured along x and z axes, and in the x z plane — Individual commanded and measured force components along the x and z axes (left), and measured forces in the x–z plane compared to the commanded circular trajectory (right). RMSE: $0.032$ N (x), $0.014$ N (z), $0.034$ N (magnitude), $2.81°$ (angle).

Workspace range-of-motion scatter for small and large hands — Workspace range of motion for a small and large hand. The overlaid manipulability indicates configuration-dependent directional force capability, showing well-conditioned force generation throughout typical operating regions.

Teleoperation User Study

Sixteen participants teleoperated a Franka Emika Panda arm with an Inspire RH56DFTP dexterous hand to press a digital scale toward target weights of 50 g and 100 g, under three feedback modes: no haptics, 1D haptics, and full 2D haptics. Two contact orientations were tested: axial probing (fingertip contact) and transverse probing (finger pad contact).

Franka arm and Inspire hand scale-pressing teleoperation setup — Teleoperation setup: a Franka arm with an Inspire dexterous hand presses a digital scale toward a target weight under three feedback conditions.

Results

In axial probing, median absolute error decreased progressively from no-haptics → 1D → 2D feedback. At the 100 g target, 2D feedback significantly outperformed both 1D ( $p = 0.032$ ) and no-feedback ( $p = 1.43 \times 10^{-5}$ ) conditions, demonstrating that 2D directional feedback enhances force regulation where single-DOF resistive feedback is insufficient.

Condition	Mode	Median Abs. Err. (g)	$p$ vs. Vis	$p$ vs. 1D
Axial 50 g	Visual	37.68	—	$2.53\!\times\!10^{-4}$
Axial 50 g	1D	30.80	$2.53\!\times\!10^{-4}$	—
Axial 50 g	2D	24.99	$2.56\!\times\!10^{-6}$	0.276
Axial 100 g	Visual	50.25	—	0.032
Axial 100 g	1D	34.73	0.032	—
Axial 100 g	2D	21.20	$1.43\!\times\!10^{-5}$	0.032
Transverse 50 g	2D	27.85	$5.07\!\times\!10^{-4}$	0.876
Transverse 100 g	2D	46.52	0.0315	0.985

Boxplots of absolute error for 50 g and 100 g target weights across feedback modalities — Boxplots of absolute error (g) for target weights of 50 g and 100 g. In axial probing, median error decreases progressively from visual-only to 1D to 2D feedback.

NASA-TLX surveys show that 2D haptic feedback significantly reduces cognitive load and increases perceived feedback quality for both axial and transverse probing.

NASA-TLX survey results across feedback modalities — Individual NASA-TLX survey results for axial and transverse probing, demonstrating that multi-directional haptic feedback significantly reduces cognitive load while improving overall performance for participants.

Conclusion

Directional fingertip haptics are essential for realistic physical interaction in teleoperation. The N2D Haptic Glove demonstrates that multi-directional 2D feedback is both feasible and beneficial for contact-rich tasks, particularly those involving axial probing — pushing small buttons, palpation, peg-in-hole insertion, and fingertip-based search in confined environments. The glove provides a foundation for next-generation wearable haptic interfaces supporting more natural and effective human–robot interaction across teleoperation, immersive VR simulation, and imitation learning.

BibTeX

@misc{n2dglove2026,
  author = "{Authors Redacted}",
  title  = "The N2D Haptic Glove: A Multi-Finger Glove for 2D Directional
            Force Feedback for Contact-Rich Manipulation",
  year   = "2026",
}