Humanoid platforms are moving from demo to deployment, and each one carries somewhere between six and forty precision joints. Cobots crossed this bridge a decade earlier with the same constraint set: the machine hauls its own actuators, runs from a limited power source, and works near people. If you're architecting joints for either, four budgets interact: mass, energy, force, and coordination. The mistakes compound down the kinematic chain.
The mass cascade
In a serial chain, a gram at the wrist is not a gram: it's a gram plus the shoulder torque to swing it, the bigger shoulder joint to make that torque, the structure to carry the bigger joint, and the battery to feed it. Run the arithmetic once and it stops being abstract: 100 g saved at a 600 mm reach takes roughly 0.6 N·m off the shoulder's gravity load, which can be the difference between two adjacent frame sizes at the shoulder, which saves hundreds of grams more. Mass savings upstream of the chain multiply; that's why they're worth paying for distally.
It's also why distal joints justify exotic materials. Our lightweight series puts 33 N·m peak in a 375 g magnesium-housed Ø52 joint precisely for wrist and forearm stations, while proximal joints (where mass rides near the axis) can take standard aluminium frames and spend the savings elsewhere.
The energy budget
A tethered robot's efficiency problem is a heating bill. A battery machine's efficiency problem is its runtime spec. Two consequences for joint selection:
- Most humanoid joints spend most of their time nearly stationary, holding posture: the exact regime where a strain wave gear's fixed losses dominate and light-load efficiency collapses. Size frames to run at healthy load fractions rather than luxuriously oversized, and put brakes on posture-holding axes so holding costs zero current.
- Legs are different animals: high transient torque, shock at every footfall, low precision demands. This is planetary territory, which is exactly the mixed harmonic-arms, planetary-legs architecture visible in current humanoid platforms.
The force budget
Working near people caps the forces you're allowed to produce, and the collaborative standards want those limits demonstrated, not asserted. Motor-current estimates fog out exactly in the gentle-contact regime that matters; joints with calibrated output torque sensing (our AS-SF and AS-SJ3 families run ±25 to ±500 N·m at up to 0.5% F.S. accuracy) make the safety case measurable per axis. Not every joint needs it: a sensible pattern is sensing at the shoulder and elbow (highest energy), current-based limits at the wrist (lowest inertia).
The coordination budget
Forty joints acting as one body is a bus-architecture problem before it's a control problem. Whole-body control wants cyclic synchronous setpoints on tight, synchronised cycles (EtherCAT's home turf) while hands, grippers, and peripherals are natural CAN-FD subnets hanging off the limb controllers. Deciding the bus topology per limb, early, is much cheaper than re-wiring a torso.
A per-axis worksheet
| Axis | Dominant constraint | Sensible architecture |
|---|---|---|
| Shoulder / hip | Torque, thermal, safety forces | Harmonic joint, torque sensing, brake |
| Elbow / knee | Torque-to-mass, shock (knee) | Harmonic (elbow) / planetary (knee) |
| Wrist / ankle | Mass above all | Lightweight harmonic frames, current-based limits |
| Fingers / grippers | Envelope, count, cost | Small planetary joints on a CAN-FD subnet |
| Posture-holding axes | Standby energy | Brake variants: holding should cost nothing |
Prototype honestly
One last discipline: prototype with the duty cycle, not the demo. A joint that looks perfectly sized for the scripted showcase will meet its real life (thermal soak, posture-holding hours, the occasional fall) in the field. Log temperatures and currents from the first prototype onward; the data decides your frame sizes better than any spreadsheet.
Building a platform and want the per-axis worksheet filled in against real part numbers? Send us the reach, payload, and duty cycle. Sizing a whole limb is the same working-day service as sizing one joint.
Related: Integrated joint, or build your own? · How to size a harmonic drive.