Arms are not legs

A humanoid leg is a torque problem: six joints per side, sized around a knee that must catch the robot’s full weight at every step. We covered that budget in our leg joint architecture explainer. The arm is a different problem entirely. Its joints see a fraction of the load — the base Unitree G1’s knee is rated at 90 N·m while its whole arm is specified to carry about 2 kg — but they must reach around obstacles, orient a hand in any direction, and pass power and signal wiring out to fingers that are getting more numerous every generation.

That is why arm design is dominated by kinematics rather than torque. A hand position and orientation in space consumes six degrees of freedom; the seventh joint that most flagship arms carry buys redundancy. With seven joints, the arm can hold the hand perfectly still while the elbow swings through space — the same trick you use to keep a cup level while moving your elbow off a table edge. Robotics literature treats this null-space motion as the key to reaching into shelves and around fixtures, and it is why inverse-kinematics research for 7-DoF humanoid arms is its own subfield.

The seven-joint consensus — and who pays for less

Across published spec sheets, the flagship platforms cluster tightly. Tesla’s Optimus Gen 2 carries 7 DoF per arm including a 3-DoF wrist; Figure’s arms are reported at 7 DoF; 1X built its upper body around 7-DoF arms precisely for the elbow-repositioning redundancy described above. Fourier’s GR-2 spreads 53 total DoF across the body with 12-DoF hands on each arm. Apptronik’s Apollo, built around the company’s in-house linear actuators, is reported at 6 DoF per arm — one of the few flagships to give up the redundant joint.

The cost engineering shows at the other end of the market. The $13,500 base Unitree G1 ships with 5 DoF per arm — three shoulder axes, an elbow, and a single wrist roll. The two remaining wrist axes, pitch and yaw, are sold as an option on the G1 EDU tier, alongside the dexterous hand. When a manufacturer needs to cut arm cost, the wrist is the first thing to go: it is the joint furthest from the torso, so every gram there costs shoulder torque, and it is the joint whose absence a scripted demo hides best.

The wrist is the hard part

Ask arm designers where the pain is and the answer is consistent: the wrist. It has the smallest packaging envelope on the robot, sits at the end of the longest lever, and everything the hand needs — power, sensor wiring, and increasingly tendons — must pass through it while it moves. Yale’s review of prosthetic and robotic wrist design documents how rarely full three-axis wrists appear in practice, and academic groups keep proposing parallel mechanisms and cable-driven wrists, from KIT’s high-torque-density compact wrist to MIT CSAIL’s DexWrist, precisely because a serial stack of three motors at the end of an arm is heavy, slow and fragile.

The production robots show two responses. Figure 02’s wrist is engineered as a conduit: Figure says it matches a human wrist’s range of motion while encasing and routing all of the power and signal wiring for the 16-DoF hand. Tesla’s Optimus V3 patents go further. With the hand’s actuators relocated into the forearm, roughly 22 hand DoF are driven by tendons — three per finger — that must cross a 2-DoF wrist without changing length. The patents describe a transition zone in the wrist where tendon bundles exit the forearm as a lateral stack and are re-ordered into a vertical stack entering the palm, geometry chosen to minimise friction, stretch and torque crosstalk between channels as the wrist moves. The wrist, in other words, has become a cable router with bearings.

Payload: the honest numbers

Arm payload is where marketing and physics diverge, so it pays to read the fine print. The figures below are what manufacturers and trade press actually publish.

• Unitree G1: about 2 kg per arm on the base model, about 3 kg on the EDU — stated plainly on Unitree’s own spec sheet for a 35 kg robot.
• Fourier GR-2: 3 kg single-arm payload on a 63 kg robot, even though its FSA 2.0 actuator family peaks at 380 N·m — peak joint torque and sustained arm payload are different quantities, as our thermal limits explainer details.
• Apptronik Apollo: a 25 kg payload — but that is a whole-robot, two-armed figure for a 72.5 kg machine built for tote handling.
• Figure 02: The Robot Report cites hands engineered to carry up to 25 kg each; Optimus Gen 2’s arm payload is reported around 20 kg. Neither company publishes a continuous-duty arm rating.

FAQ

Sources