UTRF-RoboHand: Teleoperation & Actuation

• UTRF-RoboHand is a dual-component platform featuring a computational framework for teleoperation retargeting and an innovative under-actuated robotic hand for dexterous manipulation.
• It employs synchronous tendon routing to coordinate three-joint finger actuation using a single motor per finger, reducing mechanical complexity while ensuring adaptive compliance.
• The teleoperation system integrates computer vision-based pose estimation with a learned MLP retargeter to map human hand motions to high-DOF robotic commands at 24 Hz efficiently.

UTRF-RoboHand is a universal teleoperation and under-actuated manipulation platform that refers to two research lines: (1) a computational framework for retargeting uncalibrated human hand-arm motions to high-DOF robot hands (“Universal Teleoperation Retargeting Framework”); and (2) a physical robotic hand comprising five under-actuated tendon-driven fingers using synchronous tendon routing for compact, high-stiffness, and adaptive compliance operations. Both components have been established in recent literature as complementary advances in dexterous robot hand teleoperation and actuation, targeting scalability, robustness, and practical deployment for manipulation tasks (Yuan et al., 11 Dec 2025, Sivakumar et al., 2022).

1. System Architecture and Motivation

UTRF-RoboHand addresses two canonical bottlenecks in dexterous manipulation: morpohology-agnostic control and efficient actuation. The teleoperation framework allows any operator, with minimal setup, to control multi-DOF robotic hands and arms using only a monocular RGB camera and image-based pose estimation. The actuation mechanism leverages synchronous tendon routing to realize each three-DOF finger with a single actuator, optimizing weight, mechanical complexity, and compliance.

Key design objectives are:

• Universality: Glove-free, marker-free usage, agnostic to operator and robot morphology.
• Efficiency: Minimal sensor, computational, and actuation overhead.
• Dexterity and Compliance: Large DOF workspace with predictable, tunable stiffness.

2. Synchronous Tendon-Driven Finger Design

Each Under-actuated Tendon-driven Robotic Finger (UTRF) in the UTRF-RoboHand is a three-link, three-joint structure actuated by a single motor. Two antagonistic tendon groups (flexion and extension, each with three tendons) are routed synchronously, with coupling tendons wrapped over cylindrical guides of radii $R_1$, $R_2$, $R_3$. Pulling either tendon group yields synchronous rotation of all three joints at fixed angular velocity ratios:

$$\frac{\Delta \theta_1}{\Delta \theta_2} = \frac{R_2}{R_1}, \quad \frac{\Delta \theta_2}{\Delta \theta_3} = \frac{R_3}{R_2}$$

This topology achieves full finger flexion/extension with just one actuator per finger, ensuring mechanical coupling and predictable movement while enabling adaptive compliance through tendon elasticity. Table 1 summarizes core prototype properties.

Parameter	Value	Notes
Link lengths	$L_1=35$ mm, $L_2=30$ mm, $L_3=25$ mm	3D-printed resin
Pulley radii	$R_1=10$ mm, $R_2=8$ mm, $R_3=6$ mm	Brass inserts
Tendons	1 mm stainless steel wire, $E=200$ GPa	6 per finger (3 flex/3 ext)
Actuator	Linear DC motor, 24 W	In palm or forearm

Predictable joint couplings enable analytically derived kinematics, workspace, and compliance models (Yuan et al., 11 Dec 2025).

3. Kinematic, Static, and Compliance Modeling

The UTRF kinematic model exploits the synchronous routing:

$\theta_i = \frac{q}{R_i},\quad i=1,2,3$

$$x(q) = \sum_{i=1}^3 L_i\cos\left(\sum_{j=1}^i \frac{q}{R_j}\right), \quad y(q) = \sum_{i=1}^3 L_i\sin\left(\sum_{j=1}^i \frac{q}{R_j}\right)$$

Tendon elasticity is modeled as: $$T_i = k_i (L_{T_i}' - L_{T_i}), \quad k_i = \frac{E_iA_i}{L_{T_i}}$$

Where $T_i$ is tendon tension, $E_i$ Young's modulus, $A_i$ cross-section, $L_{T_i}$ rest length. Finger compliance is obtained by iteratively solving for equilibrium and correcting wrap angles:

$$K_{\mathrm{tip}} = \frac{\partial F_{\mathrm{tip}}}{\partial \delta y}$$

A prototype exhibits a measured tip stiffness of $1.2\times10^3$ N/m under 3 kg loading, with deflection prediction error $<0.322\%$ of finger length (Yuan et al., 11 Dec 2025).

4. Teleoperation Retargeting Framework

The universal teleoperation component employs real-time computer vision (OpenPose, FrankMocap) and a parallel processing pipeline for 2D-to-3D pose estimation. It maps detected human hand and body parameters $$(\beta_h, \theta_h, \varphi_h; \beta_b, \theta_b, \varphi_b)$$ into robot joint commands via learned MLP retargeters.

The retargeting loss minimizes key-vector deviations: $$E_\text{key}((\beta_h, \theta_h), q_a) = \sum_{i=1}^{10} \|\mathbf{v}^h_i - c_i \mathbf{v}^a_i\|_2^2$$ with auxiliary penalties for self-collision and temporal smoothness: $$L_\text{total} = E_\text{key} + \lambda_c D(f_\phi(x)) + L_\text{tempo}$$

The hand retargeter (input: 55-dim human features) outputs 16 Allegro joint angles at 24 Hz, outperforming online IK (0.17 rad/joint RMSE vs. 0.25 rad/joint for baseline) (Sivakumar et al., 2022).

5. Deployment and Benchmarking

Five UTRF fingers are mounted on a rigid palm for full-hand manipulation. Actuation uses either ten linear motors (2 per finger) or five rotary capstans (1 per finger). Joint angles are controlled via PID on tendon displacement $q(t)$ such that $\theta_i(t)=q(t)/R_i$.

Experimental protocols include:

• Grasping tasks with objects: ping-pong balls, bottles, smartphones, wood blocks, and delicate materials.
• Stability under external disturbances up to 5 N.
• No object slip or delicate item damage observed; grasp stability maintained under hammer impacts (Yuan et al., 11 Dec 2025).

Teleoperation studies found novice and trained users could perform dexterous tasks (e.g., dice pickup, pouring) with mean times $<30$ s and high success rates. Responsiveness and intuitiveness were repeatedly noted, despite single-camera depth ambiguities and occlusions (Sivakumar et al., 2022).

6. Comparative Evaluation and Limitations

The actuator-per-finger design reduces mechanical complexity and weight compared to traditional three-actuator fingers. The retargeter’s run-time (24 Hz) supports interactive teleoperation, and synchronization mechanisms reliably prevent self-collision and local minima that affect traditional IK baselines.

Identified limitations include:

• Fixed compliance—stiffness only tunable by pre-tension, no active modulation.
• Tendon friction, backlash, and hysteresis are not fully mitigated.
• Absence of tactile sensing; grasp detection remains open-loop.
• Single-camera vision introduces depth-scale ambiguities; hand shapes outside training set degrade accuracy (Yuan et al., 11 Dec 2025, Sivakumar et al., 2022).

7. Prospective Extensions

Suggested improvements for UTRF-RoboHand are:

• Variable-stiffness integration (e.g., antagonistic springs, low-melting-point alloys) for dynamic compliance control.
• Embedding miniaturized force or tactile arrays for closed-loop manipulation.
• Extension to out-of-plane actuation via universal joints and advanced routing.
• Retargeting adaptation to diverse robot morphologies by retraining with updated keypoints and losses.
• Object-aware retargeting and multi-view camera fusion for enhanced robustness in teleoperation (Yuan et al., 11 Dec 2025, Sivakumar et al., 2022).

UTRF-RoboHand demonstrates a scalable, robust approach to dexterous robotic hand manipulation, bridging efficient mechanical realization and universal, human-intuitive teleoperation.