MEKAWIRE
IntermediatearXiv

A fully articulated humanoid dexterous hand with human-level joint structure enabled by linkage-based actuation

Jihao Li, Jianguo Wang, Haotian Guo, Yanzhe Wang, Guodong Lu, Huixu DongZhejiang University (Grasp Lab, School of Mechanical Engineering; Robotics Institute, Ningbo)
In 30 seconds

A Zhejiang University preprint presents Torch-Hand, a 21-degree-of-freedom humanoid hand driven entirely by screw linkages from motors in the palm: single fingertips push 52.7 to 64.7 N, the whole hand grips with more than 100 N and lifts a 40 kg dumbbell, and fingertips return to the same point within 0.031 mm. It reproduces all 33 standard human grasp types, so strength and dexterity no longer have to be traded against each other in one hand.

Research question

Can a humanoid hand copy the human joint layout, including a five-axis thumb, and still be strong, precise and simple to control, without tendons or soft actuators?

Problem

Humanoid hands usually give up one property to get another. Motor-in-joint and gear hands are precise but bulky and weak at sustained loads; tendon-driven and pneumatic hands are compact and compliant but stretch, couple joints unintentionally and deliver little force. Hands that look human often cannot hold a heavy tool, and hands that are strong rarely have a fully opposable thumb.

Previous approach

The authors' comparison table shows the spread. LEAP Hand, a popular research hand with motors in the joints, carries about 2 kg with 19.5 N fingertip force; the linkage-based ILDA Hand reaches 18 kg and 34 N with 15 DOF; most tendon and pneumatic hands stay under 6 kg and 30 N.

New approach

Torch-Hand mirrors the human joint hierarchy: a thumb with five DOF, including a three-axis base joint for wide opposition, and four DOF in each finger. All actuators sit in the palm and drive the fingers through rigid screw-driven sliders and linkages arranged like a serial robot arm, so each motor maps to one joint motion. Joints are mechanically decoupled, yet software can drive several together for whole-hand postures, and the screw linkage self-locks to hold a load once a grasp is formed.

Results

Repeat positioning error averaged 0.0111 mm for the fingers (within plus or minus 0.026 mm) and 0.0115 mm for the thumb (within plus or minus 0.031 mm). With two flexion joints driven together, fingertip forces were 60.72 N (index), 52.65 N (middle), 61.51 N (ring) and 64.69 N (little). Whole-hand grip exceeded 100 N on a grip ball, the hand crushed an aluminium can and lifted a 40 kg dumbbell. It reached every Kapandji thumb-opposition position and reproduced all 33 grasp types of the standard taxonomy on both printed and real objects. In a nine-task matrix it threaded 0.35 mm nylon line through a 0.5 mm needle eye, ran a pipette dilution, twisted a lipstick and pulled a nail with a claw hammer.

Limitations

This is a preprint and the hand is a hardware platform: the tasks were demonstrated, not run as repeated trials with success rates, and no learned policy or autonomy is reported. The hand has no tactile or force sensing yet. The authors say long-term durability, fatigue and reliability under dynamic tool use still need testing, and the paper gives no mass, finger speed or cost, which decide whether a humanoid can carry two of these hands.

Industry impact

Hands are the part of humanoids where Chinese makers are now competing hardest: [Xiaomi](/companies/xiaomi) disclosed its own 21-DoF bionic hand with about 1,500 tactile points on Oct 27, validated on 1,827 car-plant tasks. Torch-Hand shows the same joint count can carry a 40 kg payload with all-linkage drive, an approach suited to factory tool use where tendons creep. For hand suppliers and humanoid makers it is a reference design to watch; commercial impact depends on mass, cost and durability data that the paper does not yet give, so think two to three years.