Robot Hand Drive Types Compared: Tendon, Direct Joint Drive and Coupled Linkage

For a humanoid to do real work, its hands eventually have to handle objects. Robot-hand spec sheets usually list only DoF, weight and force, so hands that look alike on paper can be built in completely different ways. This article compares three products with different drive approaches, the tendon-driven Shadow Dexterous Hand, the directly driven Unitree Dex3-1 and the coupled Inspire RH56DFX, using their official specifications, and works through three calculations that reveal the drive type from the published numbers.

Three ways to drive a hand

• Tendon drive: motors sit outside the hand, usually in the forearm, and cables pull the finger joints. The hand itself can be small and light, but cable tension and friction have to be managed.
• Direct joint drive: a motor sits in or next to each joint. Force and position are easy to control joint by joint, but the motors set the size and weight of each finger.
• Coupled drive: one actuator moves several joints through a linkage. Fewer actuators make the hand lighter and simpler, but coupled joints cannot move independently.

Real products often mix these approaches. The classification below follows the drive information each manufacturer publishes.

Official specifications

Sources: Shadow Dexterous Hand, Unitree Dex3-1, Inspire RH56DFX. The Shadow page says the hand is tendon driven but does not state where the motors are, and the Inspire page does not name its actuator type. Classifying the RH56DFX as coupled is an inference from its published 6 DoF driving 12 joints.

Calculation 1: the joint-to-actuator ratio shows coupling

Dividing joint count by the number of independent actuators gives a rough measure of how coupled a hand is:

Shadow moves 24 joints with 20 motors, so $c = 1.2$. Dex3-1 moves 7 joints with 7 motors, so $c = 1.0$. RH56DFX moves 12 joints with 6 DoF, so $c = 2.0$. Values near 1 mean joints can be controlled independently; values near 2 mean the segments of each finger fold together. Wrapping a hand around an object often works fine with coupling, but rolling an object between the fingertips needs more independent joints.

Calculation 2: weight per actuated DoF

Dividing hand weight by actuated DoF compares how much mass each controllable axis costs. Shadow comes to 4.3 kg ÷ 20 ≈ 215 g, Dex3-1 to 710 g ÷ 7 ≈ 101 g, and RH56DFX to 540 g ÷ 6 = 90 g.

On this measure alone the RH56DFX looks most efficient, but two things belong next to the number. First, the Shadow hand is a research hand with more than 100 sensors, so a large share of its mass is sensing and wiring. Second, a tendon hand may keep its motors outside the hand, so hand weight alone does not tell you the load at the end of the arm. When computing arm payload, add hand weight to object weight; the shoulder-torque method from how to read a humanoid spec sheet applies directly.

Calculation 3: convert joint torque into fingertip force

When a hand does not publish fingertip force, you can estimate it from joint torque and finger length. With joint torque $\tau$ and distance $L$ from the joint to the fingertip:

Taking the 0.86 N·m measured maximum torque (with rotation) from the Dex3-1 page and assuming a 6 cm finger gives about 14 N. The finger length is an illustrative assumption, and a real grasp in which several joints push together will give a different result. The same page separately lists a 500 g grip load for a 5 cm hard round object, and conditioned official values like that are safer to design with.

Where maintenance differs

The drive type also changes what has to be inspected in service. The items below follow from the structure; no product publishes replacement intervals, so ask the vendor.

1. Tendon drive: if a cable stretches or wears, the same motor rotation produces a different finger position. The 40 tendon load sensors in the Shadow hand can be read as a way to monitor cable tension continuously. Confirm the tensioning procedure and how cables are replaced before you buy.
2. Direct joint drive: with motors inside the fingers, heat and impacts reach the motors directly. Check whether fingers can be replaced individually and whether recalibration is needed afterwards.
3. Coupled drive: fewer actuators mean fewer failure points, but linkage wear affects several joints at once. A hand like the RH56DFX that holds its pose without power will not drop an object in an outage, but it may need a defined procedure to release a jammed object.

What the drive type means for learned policies

Hand structure also shapes the action space a learned policy has to output. Every independent actuator is one more dimension the policy must predict at every control step, so a coupled hand with 6 DoF presents a much smaller action space than a tendon hand with 20 actuated DoF. A smaller action space usually needs fewer demonstrations to learn a grasp, while a larger one gives the policy room for in-hand manipulation that a coupled hand simply cannot perform. Sensor count matters in the same way: the palm cameras and fingertip tactile inputs described in how Figure’s Helix evolved only help if the hand exposes those signals at a rate the policy can use. When choosing a hand for a learning project, write down the action dimension and sensor channels alongside weight and force.

Which drive to consider first, by task

The differences between actuators and reducers themselves continue in rotary versus linear humanoid actuators.