Degrees of freedom (DoF) in humanoid robots

What the number counts

Each simple joint contributes one degree of freedom: one independent axis of motion. A human arm is about seven, ignoring the hand. A shoulder alone is three.

For a whole humanoid the totals vary enormously with what the manufacturer decides to include, which is the first reason to distrust comparisons. The Unitree G1, probably the most widely sold humanoid in the world, is listed at 23 degrees of freedom. Kinetix AI debuted KAI at a record-claiming 115 DoF. Those are not machines in the same category of ambition — but nor is the second five times the robot the first is.

The clearest illustration is a single product line. Unitree's R1 ships in configurations from the 20 to 26 DoF of the standard versions up to 40, and the R1 dual-arm platform spans 15 to 31 DoF depending on torso and base. Same robot, same name, nearly double the number. It is a configuration choice, not a fact about the model.

The gap is almost entirely hands. A pair of dexterous hands can carry 40 or more degrees of freedom between them: 1X's NEO hands are 25 DoF each, and AGILINK's OmniHand 3 Ultra-M is a 20-DoF hand on its own. Quote a robot with hands and the headline number doubles without the body changing at all.

Why more is not free

Every added degree of freedom is an actuator, a gearbox, an encoder, wiring, heat and mass — usually mass at the end of a limb, where it costs the most. It is also a new failure point, and humanoids are already machines with a lot of joints and a modest MTBF.

It compounds in software too. More freedom means a larger configuration space, more inverse kinematics solutions to choose between, and a harder whole-body control problem. Redundancy is genuinely useful — it is what lets a robot reach around an obstacle, or keep its balance while reaching — but it has to be paid for in control.

Which is why price and DoF do not track. Nori's A3 is a sub-1,700-dollar wheeled bimanual manipulator that removes freedom deliberately: wheels instead of legs, because an appliance does not need to climb stairs.

Where the joints actually go

The interesting DoF are the ones most robots skip.

The waist. A rigid torso is cheaper and simpler, and it stops the robot bending to a low shelf. Xpeng's IRON drew attention for a 5-DoF waist and a 2-DoF scapula — an analyst's verdict on the spine was "crazy," which in hardware terms means expensive and hard to control.

The hips. Not just how many, but their order. The field has converged on flexion-abduction-rotation rather than the older ordering, because the sequence changes what the leg can do at the limits of its range even when the count is identical.

The shoulder. Where human range of motion is hardest to reproduce, and where retargeting human motion capture most often has to approximate.

How to read the spec

Ask what is included — hands, fingers, head, waist, and whether passive joints were counted. Ask how many are independently driven, since coupled joints inflate a count without adding control. Ask what the range is at each, because a joint with 15 degrees of travel counts the same on paper as one with 180.

Then ask the question the number cannot answer: what does the robot do with them. A 115-DoF machine running a policy that only commands twenty is a 20-DoF robot carrying ninety-five joints' worth of cost.