What the mid-body joints actually do

A humanoid’s workspace is not its arm length. It is arm length plus everything the torso can add. A waist yaw joint lets the upper body sweep a half-circle of shelf without moving the feet — on a robot that walks at a careful 1–2 m/s and re-plans every step, not stepping is a real productivity feature. Waist pitch converts a knee-punishing deep squat into a hip hinge when the target is a tote on the floor. And waist roll lets the torso lean into a lateral reach instead of shuffling sideways.

The second job is balance. The torso is the heaviest link in the chain, and a powered waist lets the controller use it the way a human uses theirs: as an inertial appendage for regulating angular momentum while the legs deal with footholds. The third job belongs to the neck: pointing the cameras. Fixed-head robots must turn the whole torso — or the whole robot — to look at something. That is exactly the compromise Tesla removed when Optimus Gen 2 arrived with a two-axis actuated neck.

From one axis to five: how makers budget the mid-body

Unitree’s line shows the whole spectrum in one catalogue. The 23-DoF G1 base model carries exactly one waist degree of freedom — a lumbar yaw joint — and no actuated neck; its DoF budget is spent on six joints per leg and five per arm. The G1-EDU can be optioned with two additional waist degrees of freedom, taking the torso to a full three axes, and tops out at 43 DoF with the dexterous-hand package. The larger H1-2 keeps the single-waist philosophy at bigger scale: 27 DoF as seven per arm, six per leg and one waist joint, with a documented waist range of −3.14 to +1.57 radians — essentially a full half-turn one way.

Fourier sits at the other end. The GR-3, pitched as a "care-bot" for interactive and service roles, spends three degrees of freedom on the waist (pitch, roll and yaw) and two more on the head, inside a total of up to 55 with the optional 12-DoF hands. That is five mid-body axes on a robot whose job description is reading rooms and facing people — gaze and posture are the product. Tesla’s trajectory points the same way: the Gen 1 Optimus body used 28 structural actuators and a fixed head, while Gen 2 added a two-axis actuated neck. Boston Dynamics went furthest of all: the electric Atlas, listed at 56 DoF by third-party spec databases, uses custom 360-degree rotational joints that let the torso and head swivel in ways no human spine allows — its party trick is turning its upper body to face backwards rather than turning around.

Why the waist is an expensive place to add a joint

Every waist axis is a hip-class actuator in the worst possible location. It has to carry the bending moment of the entire upper body — torso, two arms, head, battery and whatever the hands are holding — which puts waist joints in the same torque bracket as hips: Unitree’s G1 documentation quotes a maximum joint-motor torque of 120 N·m, the H1-2’s largest motor is rated at 360 N·m, and Fourier quotes a 440 N·m peak for the GR-3’s strongest joint. And every kilogram of motor and reducer spent there sits high above the ankles, raising the very pendulum inertia the balance controller has to fight.

That is why the single-yaw waist is the industry’s default compromise. One axis buys the biggest workspace win — sweeping the upper body across a workstation — for one actuator, and yaw is the axis where the load is mostly torsional rather than a raw gravity moment, so the joint can be lighter than a pitch axis of the same reach. Pitch and roll, the axes that let a robot hinge to the floor and lean, are exactly the ones that see the full upper-body gravity moment; they are the expensive ones, which is why they arrive as options (G1-EDU) or on robots whose mission justifies them (GR-3). It is the same budget logic we traced in the leg and arm instalments of this series: every axis must earn its mass.

• Load: a waist pitch joint carries the gravity moment of everything above it, at maximum lever arm when the robot bends flat — the load case that sizes the reducer.
• Mass placement: actuator mass at the waist and neck raises the centre of mass and the inertia the legs must stabilise, so mid-body axes are penalised twice.
• Cabling: power and data for both arms, the head and the sensors all cross the waist; every added axis is another moving harness through the most congested cross-section in the robot.
• Audibility: as reducer maker IMSystems puts it, torso and neck joints are "moderate loads, high visibility — a whining neck joint is the first thing a human notices from across the room."

The neck: small torque, high stakes

Torque-wise the neck is the easiest joint in the robot: it moves a head, not a payload. That makes it the joint where designers optimise for everything except torque — speed, quietness, smoothness and range. The pragmatic minimum is two axes, pitch and yaw, enough to point a camera anywhere in the forward hemisphere; that is what Optimus Gen 2 and the GR-3 carry. Research platforms go further. Pollen Robotics’ Orbita is a three-degree-of-freedom parallel spherical joint built specifically to give robot heads human-like motion realism, and a well-cited ICRA design study argued for a four-DoF neck to reproduce the human combination of lower-neck and upper-neck motion at human speeds.

There is also a control dimension: the neck is a stabiliser. Work published in Sensors demonstrated camera stabilisation using an actuated (soft) robotic neck with a robust fractional-order controller — the head counter-rotates against torso motion so the vision pipeline sees a steadier world, the robotic version of the vestibulo-ocular reflex. And gaze research treats neck and eyes as one redundant chain: a Cartesian 6-DoF gaze controller coordinates them so the fast, light axes do the quick corrections while the slow, heavy ones follow. On a working humanoid, those quick corrections are worth real money: steadier pixels mean better manipulation from the same camera and compute.

FAQ

Sources