HIRO Hand: Wearable 3D-Printed Robotic Manipulator

• HIRO Hand is a wearable, 3D-printed tendon-driven robotic manipulator with 15 DOF enabling natural one-to-one human-to-robot motion mapping.
• It combines low-cost mechanical design, passive tendon actuation, and both PID and neural network controllers for effective demonstration replay and visual imitation learning.
• Experimental tests show high precision in grasp reproduction and reliable performance across diverse dexterous manipulation tasks.

The HIRO Hand is a wearable, fully 3D-printed, tendon-driven robotic manipulator designed to function dually as an anthropomorphic, dexterous end-effector and a high-fidelity human demonstration capture device. With 15 degrees of freedom (DOF) realized by five fingers, each with three active joints, it enables users to collect expert demonstration data for imitation learning while leveraging their own native tactile feedback for force modulation and dexterous manipulation tasks. HIRO Hand is uniquely characterized by direct physical coupling between wearer and robot, allowing for one-to-one joint mapping and natural demonstration trajectories that can be replayed or used for end-to-end vision-based imitation learning. The system integrates low-cost mechanical design, a passive tendon actuation scheme, and both classical PID and neural network-based controllers, aiming to overcome traditional data acquisition limitations in dexterous robot learning.

1. Mechanical Design and Sensing Architecture

The HIRO Hand is comprised of a five-fingered, tendon-driven mechanism fabricated entirely through 3D-printing processes. Each finger (excluding the thumb) comprises four phalanges (metacarpal, proximal, intermediate, distal), connected via three sequential joints: a double-axis metacarpophalangeal (MCP) joint (40° pitch, ±90° roll), a proximal interphalangeal (PIP), and a mechanically coupled distal interphalangeal (DIP) joint. The DIP and PIP are linked with a dedicated coupling wire ensuring a defined articulation constraint. Representative link dimensions for the middle finger are: a₁ = 13 mm, a₂ = 18 mm, a₃ = 17.5 mm, a₄ = 18 mm, totaling a finger length of 56.5 mm, chosen for ergonomic compatibility.

Actuation is achieved through 15 custom-selected planetary-geared DC motors (0.784 N·m stall torque, 55 g per motor), each driving a 0.4 mm stainless-steel cable routed through custom axles. Three groups of passive springs ensure return forces: (1) DIP–PIP retraction, (2) PIP–MCP return, (3) MCP neutral restoration via torsion.

The HIRO Hand implements a total of 15 miniature potentiometers (one per joint) to continuously sense absolute joint angles $q_m[i]$, enabling real-time feedback for demonstration recording and closed-loop control. Notably, no force or tactile sensors are present within the hand; the operator relies on their own sensory feedback during demonstration.

Wearability is achieved through adjustable (0–25.5 mm diameter), 10 mm-wide nylon hook-and-loop rings affixed to each phalanx, coupling the robotic and human fingers. Cables connecting hand to forearm-mount motors are bundled for user convenience. Total load on the hand is maintained below 500 g.

2. Human-to-Robot Motion and Force Mapping

The physical architecture enables direct, one-to-one mapping between the human user’s joint angles and the robotic joints, facilitated by intimate mechanical coupling. Formally, given the human configuration vector $q_h \in \mathbb{R}^{15}$, the robot’s commanded pose is

$q_d = I_{15 \times 15} q_h,$

where $I_{15 \times 15}$ is the identity matrix. Motions are visualized and planned using Denavit–Hartenberg kinematic parameters, with fingertip spatial pose determined by concatenated frame transformations:

$$T_0^3 = T_{01}(\theta_1) \cdot T_{12}(\theta_2) \cdot T_{23}(\theta_3, \theta_4)$$

and a closed-form relationship constraining the DIP joint ($\theta_4$) to the PIP angle ($\theta_3$):

$$\theta_4 = \gamma - \arccos\left( \frac{c_1 + 2l_1\sqrt{2r^2(1 - \cos\theta_3)} + 2r^2\cos\theta_3}{2l_3 s} \right),$$

with geometry constants $(c_1, l_1, l_3, r, s, \gamma)$ as defined in the canonical figure.

Force and tactile feedback are not electronically measured. Instead, users sense object contacts via the coupling rings, with no explicit measurement or estimation of tendon cable tension in the current implementation. Future extensions may employ inferred cable tensions via motor current feedback.

3. Control Strategies: Trajectory Replay and Visual Imitation

3.1 Trajectory Replay via PID Control

Demonstration trajectories $q_d[i](t)$ are captured as time-indexed joint sequences. Each DOF is tracked during replay using independent classical PID controllers:

$$\begin{align*} e[i](t) &= q_m[i](t) - q_d[i](t) \ u[i](t) &= k_p e[i](t) + k_i \int_0^t e[i](\tau)d\tau + k_d \frac{de[i]}{dt} \end{align*}$$

with representative parameters $k_p = 0.5$, $k_i = 0.1$, $k_d$ tuned for critical damping of the motor-tendon dynamics. This PID scheme is robust across diverse grasp and manipulation trajectories directly provided by the user via the wearable interface.

3.2 End-to-End Visual Behavior Cloning

To mitigate manual demonstration workload, HIRO Hand supports a visual imitation-learning paradigm. Here, a palm-mounted 200 W USB camera captures RGB sequences ($I_t \in \mathbb{R}^{160 \times 320 \times 3}$, 15 Hz) during interaction, paired with per-joint binary activation labels $u_t \in \{0,1\}^{15}$. Data acquisition employs a 1-DOF linear vertical approach, collecting approximately 600 labeled instances per task with brightness augmentation.

The controller utilizes a convolutional neural network: three convolutional layers (downsampling to $\sim10 \times 20$), followed by three fully connected layers, culminating in 15 sigmoid outputs. The network is trained for 45 epochs (batch size 75) using the Adam optimizer, minimizing the binary cross-entropy loss:

$$\mathcal{L}(\theta) = - \sum_{i=1}^{15} \left[ u_i \log \hat{y}_i + (1-u_i) \log (1-\hat{y}_i) \right].$$

At inference, binary joint commands are thresholded and relayed to the underlying PID loops.

4. Experimental Evaluation and Performance

4.1 Single-Finger Repeatability

A 100-cycle, 30 mm-lift test under 600–1500 g payloads yielded RMS repeat errors $\leq 0.14$ mm (max $\leq 0.32$ mm), corresponding to fingertip forces up to 15 N and demonstrating reliability in pose reproduction.

4.2 Grasp Taxonomy Coverage

HIRO Hand achieves 21 out of 28 Cini et al. (2019) canonical human grasp types ($\approx$80% coverage) without accessories, spanning a wide grasp taxonomy relevant for manipulation research.

4.3 Task Completion

In PID-controlled demonstration replay, the hand executed cup, glue bottle, toy ball, tongs, and key pickup; tennis ball rotation; and typing "GREAT" on a standard keyboard. All static grasps achieved 100% success, in-hand manipulations exceeded 90% completion.

4.4 Vision-Based Policy Results

Behavior cloning was evaluated across five tasks: "egg-crate" foam grasp (60% success), cup grasp (67%), wire ball grasp (75%), foam rotation (100%), and chained grasp/unscrew (75%). Mean success rate across tasks was 0.722 (range 0.6–1.0). The protocol tested both distant and proximal reorientation scenarios.

Task	Success Rate
Grasp “egg-crate” foam	60%
Grasp cup	67%
Grasp wire ball	75%
Rotate foam	100%
Grasp + unscrew faucet	75%

The results reflect reliable grasping and simple sequential dexterous actions under controlled, visually-mediated operation.

5. Comparative Analysis, Limitations, and Future Work

The HIRO Hand’s direct physical mapping confers advantages over data glove or VR-captured demonstrations, notably in leveraging the operator’s tactile and proprioceptive faculties for force and contact feedback. This results in demonstration data that is more consistent with real-world manipulation requirements.

Principal limitations include the absence of on-board force or tactile sensors, reliance on human feedback during teaching, and user fatigue during prolonged sessions. Scaling the system for more complex environments would require (1) integration of force estimation, potentially through motor current-based tendon tension inference, (2) extended sensing modalities such as fingertip pressure, (3) transfer of actuators offhand via Bowden cables for improved ergonomics, and (4) enhanced vision-based policies for manipulation in less structured, higher-DOF settings. The current low-cost (\$400), compact, and easily repeatable (0.14 mm RMS) design establishes HIRO Hand as a versatile platform for both demonstration collection and closed-loop dexterous control research.

A plausible implication is that augmenting HIRO Hand with distributed force sensing and offhand actuation could extend its utility to richer, multi-modal imitation learning and wider deployment in unstructured scenarios.