Inductive encoders have been pushed to a height they haven't reached yet

Let me say something. For the past six months, whenever anyone talks about humanoid robot joints, inductive encoders get brought up. Brokerage research reports directly label them as the next mainstream solution, and the estimate of Tesla Optimus Gen 2 having 40 actuators and 54 encoders has been passed around everywhere.

I don't really agree with that judgment.

Last week I went through the spec sheet for this one from Gongwang Sensing and asked a few people about it. The thing itself is fine. Its principle belongs to the same category as a resolver, except it uses PCB coils, avoiding the winding and cost of the transformer approach. The structure can be made very thin, and it's genuinely less troublesome than optical encoders when it comes to oil and dust resistance. Single-turn, multi-turn, high and low precision are all covered—the product line is quite complete.

But complete and usable are two different things.

What I asked about was long-term drift and batch consistency. What they sent back was a curve from lab room temperature. I can understand that—not many domestic players are working this line, and Gongwang counts as one of the ones running ahead. But humanoid robot joints will have to rely on tens of thousands of machines running in workshops, outdoors, and in environments with temperature swings of tens of degrees for a year, and the 5,000th unit still has to match the first. There's no public data on that layer yet.

In the encoder business, the hard part is calibration, consistency, and the yield rate you tune unit by unit on the production line. Optical, magnetic, inductive—each has its own way of surviving at the principle level, and all of them can work. Optical encoders took decades to get costs down. Inductive ones have only just started. Nobody can skip that stretch in between.

So my attitude is clear: if you're positioning or betting on the "next mainstream" narrative right now, I'd suggest waiting a bit. Teams already validating servos or robot joints can put Gongwang on the candidate list, but they must test it side by side with magnetic encoders—don't deploy it alone.

Research reports can calculate market space, but they can't calculate a yield curve.