Ruka-v2: Open-Source Dexterous Robotic Hand

• Ruka-v2 is a fully open-source, tendon-driven humanoid robotic hand offering 20 actuated DOF with enhanced wrist mobility and finger abduction/adduction for dexterous manipulation.
• It employs a tendon-driven actuation system with minimized distal inertia and a vision-based retargeting control loop, facilitating teleoperation and autonomous policy learning.
• User studies demonstrate a 51.3% reduction in task completion time and a 21.2% improvement in success rates, affirming its performance gains over previous models.

Ruka-v2 is a fully open-source, tendon-driven humanoid robotic hand designed for dexterous manipulation and robot learning research. Building on its predecessor, Ruka, Ruka-v2 introduces two previously absent degrees of freedom—wrist mobility and finger abduction/adduction—enabling human-like behaviors such as in-hand rotation, fine grasping of thin objects, and manipulation in confined spaces. The platform is engineered for accessibility, with all parts 3D-printable using consumer FDM equipment, open-source controller software, and standard hardware components, maintaining a total hardware cost of approximately \$1,500 (Xinqi et al., 27 Mar 2026).

1. Mechanical Architecture and Kinematics

Ruka-v2 features a fully tendon-driven actuation using Faulhaber DC motors positioned in the proximal forearm shell, minimizing distal inertia. Tendons, routed via Bowden cables, pass through a central wrist pivot and custom routing plate to the fingers, reducing frictional losses and sharp bends.

The hand incorporates a total of 18 actuated degrees of freedom (DOF) in the fingers and thumb: for each non-thumb finger (index, middle, ring, pinky), the DIP (1 DOF), PIP (1 DOF), MCP flexion/extension (1 DOF), and MCP abduction/adduction (1 DOF) are implemented, except the middle finger which is fixed for lateral spread as a reference. The thumb includes IP (1 DOF), MCP (1 DOF), and CMC opponency (1 DOF). This yields $4 \times 4 + 3 - 1 = 18$ DOF for the digits, plus a decoupled 2-DOF wrist, for a total of 20 actuated DOF.

Parallel Wrist and Finger Abduction

The wrist implements two axes—flexion/extension (±45° flex, ±30° extension) and radial/ulnar deviation (±35°)—driven independently via rectangular four-bar linkages intersecting at a single geometric pivot defined by a passive spherical joint. All finger tendons traverse a central through-hole in the wrist for direct routing. Each finger’s abduction/adduction axis is housed in an independent knuckle module with passive compliance provided by extension springs; a dedicated tendon enables active adduction.

Modular Components and Sensing

Structural components, pulleys, and housings are single-material PETG or PLA; soft fingertip pads use flexible TPU. Bearings, springs, and fasteners are off-the-shelf, streamlining assembly. Optional DIP/PIP coupling uses fixed-length strings to enforce $\theta_2 \approx \theta_1$, enhancing repeatability at the cost of compliance.

Joint angle measurement is facilitated by AS5600 magnetic encoders (12-bit), attachable via press-fit magnets and multiplexed via ESP32 QTPy for calibration and closed-loop experiments.

2. Control System and Human-to-Robot Mapping

The control stack employs a two-stage architecture:

1. Vision-based Retargeting: 3D keypoints, captured via stereo/RGB cameras and processed using Mediapipe or DepthAI, are aligned to the robot-base coordinate frame. The AnyTeleop-DexRetarget system solves:

$$\min_{\theta} \sum_{i=1}^N \lVert\mathbf{v}_i(\theta) - \mathbf{d}_i\rVert^2$$

where $\mathbf{v}_i(\theta)$ are robot link vectors and $\mathbf{d}_i$ are the corresponding human finger vectors.

2. Joint-to-Motor Linear Mapping: Each joint maps desired angle $\theta \in [\theta_{\min}, \theta_{\max}]$ to motor position $p \in [p_{\min}, p_{\max}]$ by:

$$p = p_{\min} + c\, \frac{\theta - \theta_{\min}}{\theta_{\max} - \theta_{\min}} (p_{\max} - p_{\min})$$

where $c$ is a per-joint gain for friction/tendon stretch compensation.

The control loop consists of: (1) video keypoint extraction, (2) retargeting optimizer for joint angles $\theta$, (3) mapping to motor positions $\theta_2 \approx \theta_1$0, (4) dispatch via CAN bus, and (5) optional encoder-based closed-loop correction. All calibration limits are determined by automated routines moving joints to hard-stops.

3. Empirical Performance and Evaluation

User Study: Ruka vs. Ruka-v2

A study with $\theta_2 \approx \theta_1$1 users evaluated teleoperation (Oculus + OpenTeach, 7-DOF Franka + Ruka hand) on three tasks—bread pick & place, pen grasp, and book opening. Results indicated a 51.3% reduction in mean completion time and a 21.2% increase in success rate for Ruka-v2 relative to the original Ruka.

Metric	Ruka-v1	Ruka-v2	Relative Change
Mean Completion Time	—	—	$\theta_2 \approx \theta_1$2
Overall Success Rate	—	—	$\theta_2 \approx \theta_1$3

Thermal and Mechanical Endurance

Continuous 5-hour tests (finger and wrist sweeps) showed peak/steady temperature changes: fingers (Δ0.81°C), thumb (Δ1.70°C), wrist (Δ7.25°C), with no observed thermal throttling.

Static payload tests showed:

• DIP–PIP (non-thumb): 1200 g (15 s)
• MCP (non-thumb): 780 g (15 s)
• Adduction: 150 g (15 s)
• Thumb curl: 835 g (20 s)
• Wrist supination/pronation: 1215 g (20 s)
• Wrist radial/ulnar-side up: 835 g (20 s)

Kinematic and Control Fidelity

Joint tracking averaged 8.26° absolute error (10.68% of ROM). Coupled DIP/PIP fingers achieved ±2–3° repeatability, compared to >10° hysteresis and variability for uncoupled fingers.

4. Applications in Robot Learning

Teleoperation Benchmarks

Ruka-v2 supports extensive teleoperation suites: 10 single-arm and 3 bimanual tasks (including bread pick & place, pen grasp, book opening, soup ladle scooping, calligraphy, magnetic assembly, music-box opening, and cloth manipulation), with data collection via OpenTeach and Oculus interfaces.

Autonomous Policy Learning

Autonomous learning leverages the BAKU framework (transformer with action-chunking). Observations comprise a 23-dimensional vector (7 Franka + 16 Ruka-v2 joints) concatenated with RGB visual features via ResNet-18 and MLP. Policies are trained with ≈100 demonstrations/task and Gaussian noise ($\theta_2 \approx \theta_1$4 rad, $\theta_2 \approx \theta_1$5 rad). Evaluated tasks include bread pick & place (success 8.2 ± 1.4), music-box opening (7.5 ± 2.0), and pen grasp with abduction (9.1 ± 0.8) out of 10. Loss combines standard behavior cloning MSE and transformer cross-entropy for chunked actions.

5. Limitations and Future Directions

The joint-to-motor mapping is a linear simplification. Planned work includes replacing this with data-driven tendon modeling using dense magnetic encoder feedback. Integration of tactile sensors (TPU-based e-flesh) may be affected by magnet interference, requiring further study. Additional avenues include reducing friction/slack using advanced liners, and torque-based control exploiting the tendon Jacobian:

$\theta_2 \approx \theta_1$6

to facilitate force-sensitive manipulation.

6. Open-Source Accessibility and Reproducibility

All 3D print files, CAD models, build instructions, controller code, and demonstration videos are publicly available at https://ruka-hand-v2.github.io/. The platform is designed for broad accessibility and straightforward repair, lowering barriers for hardware teams and researchers seeking to reproduce or extend dexterous robot hand research (Xinqi et al., 27 Mar 2026).