Got a Harmonic Joint? Get It Spinning on a Single Axis First

A friend who builds complete machines hyped this up in the group chat for half a month, saying that Yiyou's RHU humanoid-specific harmonic joint is flat enough to squeeze into shoulders and elbows to save space. I grabbed one last week, wanting to see how long it takes to get started from zero. This post is for people who've never touched an integrated joint — from wiring to making it spin, all the pits I fell into are written in here.

First, figure out what this lump of metal in your hand actually is

An integrated joint, in plain terms, is stuffing four things — motor, reducer, encoder, driver — into one shell. You give it power and commands, it gives you an angle, and you don't have to match anything up yourself in between.

Harmonic is a type of reducer. Ordinary gears rely on two gears meshing; harmonic relies on a thin metal cup that can deform, pushed open by an elliptical thing, to bite into the outer rigid gear. The advantage is almost no backlash, small return error — used on humanoid robot elbows and wrists, the motion only looks non-jittery this way.

The RHU uses dual absolute encoders. Absolute means that after power loss it still knows what angle it was stopped at, no need to home every time you power on. Dual, generally one at the motor end and one at the output end — the output end one is the joint angle you actually care about.

At equivalent torque output strength, volume reduced by 30% compared to traditional designs

This is its main selling point. I measured the shell with calipers, and it's indeed a notch shorter than the old joint of the same torque I have on hand. The wheeled chassis over at our warehouse only handles movement, and the upper body joints have always been used as a makeshift; it was seeing this size that made me think about swapping.

Wiring — short that safety wire first

Power connects to 24 to 48V; I used 48V, because in that wheeled complete-machine solution in the materials all joints are uniformly 48V, and I didn't want to stock another kind of power supply. Communication first goes over CANFD, two wires, CAN_H and CAN_L, don't connect them backwards. Short the two STO wires — this step is the easiest to skip.

STO is Safe Torque Off, and the factory default is the cut-off state. If you don't give it an "I'm safe on my end" signal, the driver is dead — light on, scannable, but just won't move. I burned twenty minutes on this the first day, only finding out after flipping through the manual.

After wiring and powering on, the normal state is the driver indicator light staying on, no red flashing. If it flashes rhythmically, first check the power supply voltage and the CAN wires.

I ran both solutions

Solution A, a USB-CANFD adapter box plus the manufacturer's host software. Plug into the computer, pick the baud rate, scan nodes, change ID, write target angle, hit send. From wiring to the joint moving, on my end about forty minutes.

Solution B, EtherCAT. You have to install a master station on the computer, configure the slave description file, then run through it again. The upside is that later when you hang a dozen-plus joints for synchronization, the jitter is smaller than CANFD, and the whole machine doing walking motions will be noticeably different. The downside is slow to get started — I messed around for an entire afternoon before getting the first joint to move.

Newbies go with A first, get the joint turning properly, then decide which bus the whole machine uses. The RHU supports both protocols, switching doesn't require changing hardware.

Places that are easy to mess up

The first pit is the enable sequence. Give 48V main power first, then the enable command; the reverse reports undervoltage, and when cutting power do it in reverse. The second pit is power supply wattage. I initially used a low-wattage 48V supply, it spun fine with no load, but restarted as soon as I loaded it — the current at the instant a harmonic joint stalls is bigger than you'd think, leave enough margin on the supply. The third pit is the zero position. Although the absolute encoder doesn't forget position after power loss, the first time you mount it onto a structural part, you still have to confirm once that the mechanical zero and the encoder reading match; if they don't, all subsequent angles are off. On my end the first mount was off a bit, and the motion looked awkward.

When writing commands, there are three loops — position, speed, current. Newbies first move only the position loop, give speed a conservative value, and keep the current limit loose at first. Loosen it up after it runs smoothly.

If a single axis can turn, that means the path of wiring, communication, enable, command is through. Next you can try two things: string two joints on the same CANFD bus, write a simple linked motion, and see whether they can reach the target angle simultaneously; or add the optional torque sensor, push the joint with your hand, and see whether it can follow your force. The latter is where humanoid robots are truly hard.