CYJ Hand-0: Biomimetic 21-DOF Robotic Hand

• The paper introduces CYJ Hand-0, a 21-DOF humanoid hand that mimics human anatomy at 1:1 scale using a hybrid tendon-driven system.
• It integrates motor-driven flexion with SMA-based extension, validated through Kapandji tests, gesture actions, and diverse grasp experiments.
• The design emphasizes lightweight, modular fabrication via a 3D-printed metal frame, though challenges in control and thermal management remain.

CYJ Hand-0 is a 21-degree-of-freedom humanoid dexterous hand that combines a fully articulated, human-scale skeletal architecture with a hybrid tendon-driven actuation system based on DC motors and shape memory alloys (SMAs). The platform is explicitly configured to mimic the anatomy, kinematics, and functional dexterity of the human hand while maintaining lightweight construction, modular fabrication, and compact actuator integration. Its reported implementation uses a fully 3D-printed AlSi10Mg metal frame, high-strength fishing line as artificial tendons, a custom rear support structure carrying hybrid actuation modules, and an Arduino Mega 2560-based control system. Mechanical and kinematic experiments were reported to validate biomimetic dexterity, including all 10 positions of the Kapandji thumb test, 32 gesture actions, and more than 30 grasping experiments (Chai et al., 19 Jul 2025).

1. System concept and principal specifications

CYJ Hand-0 was designed around five concurrent objectives: high anatomical and kinematic biomimicry at 1:1 human scale, high dexterity with many DOFs and thumb opposability, robust load capacity with lightweight construction, modular low-cost fabrication via metal 3D printing, and a compact integrated hybrid actuation system using both DC motors and SMA wires. The hand excludes the wrist in its DOF count and reproduces the commonly cited 21-DOF robotic-hand target through an explicit thumb carpometacarpal implementation (Chai et al., 19 Jul 2025).

Specification	Value
Degrees of freedom	21
Digits	Five
Scale	1:1 human scale
Hand mass	380 g
Frame material	Fully 3D printed AlSi10Mg metal
Tendon material	12-braid high-strength fishing line, 0.55 mm diameter, nominal 40 kg pull
Flexion actuation	DC screw–nut linear motor modules
Extension / lateral actuation	SMA-based linear modules
Control system	Arduino Mega 2560
Reported dexterity tests	Kapandji 10/10, 32 gesture actions, more than 30 grasping experiments

The mass figure is particularly notable because the paper places it against a typical human hand mass of approximately 413 g, or about $0.575\%$ of a 75 kg body mass. The intended balance is therefore not merely anthropomorphic geometry, but anthropomorphic scale and mass distribution. A plausible implication is that the design aims to preserve human-like inertial characteristics while relocating the majority of the actuation hardware away from the hand itself.

2. Anatomical architecture and joint organization

The mechanical architecture follows human-hand nomenclature directly, including carpals, metacarpals, phalanges, and wrist. The reported implementation comprises 18 distinct component types and 80 printed metal parts. The palm uses an integrated carpal block and separate, non-parallel metacarpals, producing a naturally concave, cupped palm that the paper associates with improved power grasps (Chai et al., 19 Jul 2025).

The three universal fingers—index, middle, and ring—share an identical 4-joint “side-swinging type” mechanism. Joint 1 provides MCP lateral swing for abduction/adduction; Joint 2 provides MCP flexion/extension; Joint 3 provides PIP flexion/extension; and Joint 4 provides DIP flexion/extension. Joint 1 is orthogonal to Joints 2–4, while Joints 2–4 are parallel. The little finger follows the same structure with a shorter proximal phalanx. The thumb has three anatomical joints—CMC, MCP, and IP—but the CMC is decomposed into two 1-DOF coupled segments, one for adduction/abduction and one for flexion/extension. The paper states that this replaces a full anatomical saddle-joint model with a functionally appropriate but simpler actuation and modularity scheme.

The paper explicitly contrasts a simplified human-hand accounting with the robotic design. For the human hand, after omitting subtle DOFs such as finger axial rotation, palm deformation, and detailed TMC complexity, it reports

$$\text{Total DOFs} = 10\times 0 + 5\times 0 + 10\times 1 + 5\times 2 = 20.$$

For CYJ Hand-0, it reports 21 DOFs, including an explicit 2-DOF thumb carpometacarpal design. This distinction is important because it clarifies a common misconception: the reported 21 DOFs do not imply a full reproduction of every subtle human-hand articulation, but rather a specific functional allocation that prioritizes opposability and dexterous posture generation.

Joint range	Human hand	CYJ Hand-0
MCP	$0\text{–}90^\circ$	$0\text{–}90^\circ$
PIP	$0\text{–}110^\circ$	$0\text{–}110^\circ$
DIP	$0\text{–}90^\circ$	$0\text{–}90^\circ$
Thumb Ab–Ad	$0\text{–}70^\circ$	$0\text{–}53^\circ$
Thumb Ex–Fl	$$\text{Total DOFs} = 10\times 0 + 5\times 0 + 10\times 1 + 5\times 2 = 20.$$0	$$\text{Total DOFs} = 10\times 0 + 5\times 0 + 10\times 1 + 5\times 2 = 20.$$1

These ranges show that the non-thumb fingers are intended to be essentially human-like, whereas the thumb trades some abduction/adduction range for enlarged flexion/extension excursion. This suggests a design emphasis on functional opposition and reach across the palm rather than exhaustive anatomical equivalence.

3. Kinematic formulation and tendon mechanics

For the non-thumb fingers, the paper models anatomical flexion coupling between the PIP and DIP joints as

$$\text{Total DOFs} = 10\times 0 + 5\times 0 + 10\times 1 + 5\times 2 = 20.$$2

This coupling is used to represent the tendency of DIP flexion to accompany PIP flexion. Each universal finger is then modeled as a 4-DOF serial manipulator using the Denavit–Hartenberg convention, with the base frame located at the intersection of the first two joint axes and link lengths $$\text{Total DOFs} = 10\times 0 + 5\times 0 + 10\times 1 + 5\times 2 = 20.$$3 mm, $$\text{Total DOFs} = 10\times 0 + 5\times 0 + 10\times 1 + 5\times 2 = 20.$$4 mm, and $$\text{Total DOFs} = 10\times 0 + 5\times 0 + 10\times 1 + 5\times 2 = 20.$$5 mm (Chai et al., 19 Jul 2025).

The fingertip position $$\text{Total DOFs} = 10\times 0 + 5\times 0 + 10\times 1 + 5\times 2 = 20.$$6 is given as

$$\text{Total DOFs} = 10\times 0 + 5\times 0 + 10\times 1 + 5\times 2 = 20.$$7

$$\text{Total DOFs} = 10\times 0 + 5\times 0 + 10\times 1 + 5\times 2 = 20.$$8

$$\text{Total DOFs} = 10\times 0 + 5\times 0 + 10\times 1 + 5\times 2 = 20.$$9

Inverse kinematics begins with

$0\text{–}90^\circ$0

after which $0\text{–}90^\circ$1 and $0\text{–}90^\circ$2 are solved numerically, for example by Newton iteration, and $0\text{–}90^\circ$3 is obtained by back-substitution. The paper notes that this mapping from desired fingertip pose to joint angles is central to grasp posture and motion planning.

The tendon system is also modeled geometrically. Flexion is produced by three palmar flexor tendons, one per phalanx; extension by two dorsal extensor tendons, located proximally and at the middle phalanx; distal return is assisted by a spring rather than a dedicated distal extensor tendon; and lateral abduction/adduction is produced by two side tendons per finger. Tendon routing passes through holes and tubes in the AlSi10Mg frame, with grease used at contact points to reduce friction.

For the proximal joint, the paper reports that when the joint flexes from $0\text{–}90^\circ$4 to $0\text{–}90^\circ$5, the tendon length shortens by 16.86 mm, leading to a required linear-actuator travel of at least 20 mm. It further reports that the moment arm increases with flexion, thereby reducing the required tendon tension at high flexion and allowing larger fingertip force for the same actuator force. This tendon analysis is used to size the actuator stroke and estimate torque and force capability.

4. Hybrid SMA–motor actuation and integrated actuator packaging

The defining feature of CYJ Hand-0 is its hybrid actuation architecture. Finger flexion is assigned to DC screw–nut linear motor modules, while finger extension and lateral abduction are assigned to SMA-based linear modules. The rationale reported in the paper is explicit: DC motors provide high controllability, high displacement accuracy, self-locking through the screw mechanism, and stable output force, whereas SMA wires provide high specific force, low weight, inherent compliance, and compact integration, albeit with thermal lag, nonlinearity, and hysteresis (Chai et al., 19 Jul 2025).

The motor-driven flexion unit is built around a GA12-N20-298 motor. The reported parameters are a rated voltage of 12 V, rated current of 0.18 A, rated speed of 100 rpm, rated torque of 1.5 kg·cm, and screw pitch of 0.7 mm. The module itself has a 20 mm stroke, 40 g weight, rated output force of 400 N, and a speed of 1.17 mm/s. Mechanically, the motor drives a lead screw, and the moving nut pulls the flexor tendon downward to flex the finger.

The SMA extension and abduction modules use NiTi wires from DYNALLOY. The paper lists density $0\text{–}90^\circ$6, phase transition temperature $0\text{–}90^\circ$7, specific heat $0\text{–}90^\circ$8, thermal conductivity $0\text{–}90^\circ$9, Poisson’s ratio 0.33, and thermal expansion coefficients of $0\text{–}90^\circ$0 in martensite and $0\text{–}90^\circ$1 in austenite. The module incorporates multiple SMA wires of different diameters, a permanent magnet, an electromagnetic coil, a compression spring, a carbon-fiber tube, and fixing plates at both ends. By combining SMA wires with electromagnetic assistance, the module is reported to achieve a maximum stroke of 18 mm and a maximum force of 784 N.

At the system level, CYJ Hand-0 uses a separate arm-like structure to house actuators and route tendons to the wrist. The total actuator count is 32: 15 motor-driven modules for flexion and 17 SMA modules for extension and lateral motion. The 15 motor modules are arranged in three axial layers of five modules each, stacked cylindrically; 12 SMA modules are inserted between them. The support structure is 3D printed in white resin, weighs 780 g, is hollow for tendon routing, attaches on its upper side to the wrist, and on its lower side to a UR10 robot arm. The paper characterizes this as a compact power “backpack” that preserves low hand mass while sustaining substantial force output.

5. Control architecture and empirical evaluation

The control stack is based on an Arduino Mega 2560 microcontroller. Flexion control uses 15 Tonkeys TKS-M8-P0 motor controllers, with the Arduino receiving high-level commands from a PC via the Arduino IDE and issuing commands that enable forward/reverse rotation and start/stop control. The paper does not specify closed-loop position sensing and describes the control as essentially low-level position or displacement control of the linear modules via motor rotation. For the SMA subsystem, the Arduino outputs PWM signals that drive relays controlling power to the SMA modules, thereby modulating heating intensity through duty cycle (Chai et al., 19 Jul 2025).

No dedicated encoders, force sensors, or temperature sensors are described in the control architecture. Instead, motion characterization was conducted offline using video processing with Tracker software, and finger load testing was performed by hanging known masses at the fingertip. The absence of an integrated sensor suite is therefore not incidental but a defining feature of the reported prototype stage.

The experimental results span finger kinematics, load capacity, gesture replication, thumb opposition, and grasp taxonomies. In finger flexion experiments, time-position data extracted from video were converted into joint-angle trajectories, and the reported joint ranges matched the design targets for MCP, PIP, and DIP motion. In load tests on the universal finger, distal-phalanx loading from 200 g to 1200 g was completed successfully. The paper states that the structural load-bearing capacity of a single finger exceeded 4 kg, but that when loads greater than 1.5 kg were applied, motors overloaded and burned out. The practical safe rating was therefore reported as 1.2 kg per finger, yielding an overall hand load of approximately 8 kg.

Dexterity evaluation emphasizes both standardized and demonstrative tests. CYJ Hand-0 reportedly passes all 10 positions of the Kapandji thumb opposition test. For gesture reproduction, it replicates 11 common index-finger gestures, including 7 that are easily performed by a human and 4 that are usually difficult without external assistance, and it demonstrates 32 distinct whole-hand gestures. A skin or flesh-like covering is also shown for a more anthropomorphic external appearance.

In grasping experiments, the hand achieved five of Schlesinger’s six basic grasp types: cylindrical grasp, spherical grasp, three-finger pinch, side pinch, and two-finger pinch. Hook grasp was not tested. Under Cutkosky’s 16-grasp taxonomy, the paper reports successful demonstration of 9 types, including large diameter, small diameter, medium wrap, adducted thumb, light tool, disk, sphere, tripod grasp, and side pinch. Additional demonstrations included precision grasping of a pen and grasping a full 550 ml water bottle.

6. Materials, limitations, and broader research significance

The manufacturing strategy centers on fully 3D-printed AlSi10Mg metal for the skeletal structure and key parts. The reported advantages are the ability to realize complex anatomical shapes, integrate internal passages for tendon routing, obtain a high stiffness-to-weight ratio, and reduce assembly and manufacturing costs. Many joint components are modular and shared across fingers, facilitating replacement and fabrication. The tendons use 12-braid fishing line, mark 9.0, with 0.55 mm diameter and nominal maximum pull of 40 kg. Tendon endpoints are secured using figure-eight loops and double fisherman’s knots. The paper notes friction at holes and tubes, mitigated with grease, and mentions potential long-term creep or wear as not extensively analyzed (Chai et al., 19 Jul 2025).

The main limitations reported are equally specific. Motor burnout occurred at fingertip loads above 1.5 kg, making motor thermal capacity rather than structural strength the immediate bottleneck under heavy loading. SMA actuation is controlled with relatively simple open-loop PWM and relay logic, without explicit temperature or strain feedback. No integrated sensory system is present for joint angle, force, or temperature, and the overall control architecture remains Arduino-based and open-loop, without advanced motion planning or impedance control. SMA speed and thermal management are also identified as potential constraints for dynamic manipulation.

Future work is described in terms of structural optimization, improvement of the sensor system through added joint-angle, force, and temperature sensing, and development of a closed-loop control system for more precise position control, better SMA temperature and strain regulation, and improved robustness and repeatability. The paper also points toward increased application value in prosthetics and robotic manipulators integrated with platforms such as UR10.

The broader significance claimed for CYJ Hand-0 is its combination of a fully metal 3D-printed, anatomically faithful skeleton, 21 DOFs at human scale, and a hybrid SMA–motor tendon-driven actuation architecture that separates motor-driven flexion from SMA-driven extension and abduction. This places the platform at an intersection of biomimetic morphology, additive manufacturing, and unconventional actuation assignment. A plausible implication is that the hand is intended less as a finalized industrial manipulator than as a research platform for studying how anatomical fidelity, tendon routing, and mixed actuator modalities interact under dexterous-manipulation constraints.

A further plausible implication, supported by adjacent hand-robotics research, is that CYJ Hand-0 could benefit from external perception and teleoperation pipelines built around real-time hand pose and mesh estimation. For example, ReJSHand reports real-time monocular 3D hand pose estimation and mesh reconstruction at 72 FPS with a 1.91M-parameter model, and explicitly frames its outputs as suitable for robotics, teleoperation, and dexterous manipulation interfaces (An et al., 8 Mar 2025). Such integration is not part of the CYJ Hand-0 paper, but it suggests one concrete direction for coupling biomimetic hardware with lightweight perception modules.