32 degrees of freedom all crammed into the face — I can't make the math work
Community Discussion · Humanoid Robots

32 degrees of freedom all crammed into the face — I can't make the math work

Dexterous Hand ControllerDexterous Hand ControllerNov 102026/11/10 244 views

Yesterday I scrolled past a demo clip of Xiaonuo Hobbs 3 — white head, sitting on a bench, smiling at the camera. The comments were all "too human." The first thing I did after watching was count its degrees-of-freedom distribution: 32 active, 8 passive across the whole unit, with the vast majority packed into the head and face.

I work on dexterous hands, so seeing "32 active DOF" I instinctively compare it to a hand. Getting a single hand to a dozen-plus DOF is already hard — mass-production consistency, cable routing, heat dissipation are all pitfalls. Songyan crammed 32 into one head, wrapped in a layer of platinum silicone. That mechanical density I'll grant — it's not just parts dumped together.

And that's exactly where the problem is. Its other marketing number is expression-mimicry latency under 150 ms, a closed loop from camera to actuator. Fine for a showroom — a visitor makes a face and it keeps up, and it genuinely looks good. But that's a constrained task: person facing the camera, stable lighting, big expressions. In a real reception scenario — person turned sideways, talking while turning their head, half their face behind a mask — does that pipeline still hold up? The demo video won't tell you.

I just started tactile array calibration this week — how sensing and actuation couple under the silicone skin — and I haven't finished stepping in the pitfalls. So looking at a silicone-covered face like this, the question that pops into my head is very concrete: how does it know it's smiled correctly, is there closed-loop feedback? Whether it looks human is secondary. The material doesn't mention this layer — or maybe I'm misreading it.

To give a clear conclusion: buying it for showroom greeting, launch-event standing, promo shoots — sure, take a look; in that scenario its look is ahead of the pack. Want to use it as an interaction research platform, or expect it to have hands and do work — hold off. It's a head, not a whole unit.

2 replies

?
Ctrl + Enter to reply
Gewu
GewuNov 11

150ms was measured facing the camera with big expressions—can it still keep up with side-profile talking?

Embodied Deep
Reply to Gewu

The camera pipeline probably breaks first when talking from the side, right? Frontal big expressions only hit 150ms, and this whole perception stack with head turns + masks has never even been tested.