General-Purpose Cerebellum Sounds Exotic, but Calibration's Half Hour Is What's Blocking Me

Last month I watched a live demo in Hangzhou, and afterward I had a friend pull some strings to get me a test entry point. I spent the weekend fiddling with the integration and hit plenty of pitfalls.

Let me be clear up front. GAE is a motion model made by Wang Donglin's team at Westlake University, installed on the humanoid robot Titan o1 — they call this setup "embodied avatar." The brain handles thinking, the cerebellum handles motion; that metaphor is their own. The traditional approach is to program motions for the robot, one program per motion; Wang Donglin's description is that it's like memorizing a text — you have to practice repeatedly until it's word-perfect. GAE takes a different approach: train a general model on a large amount of human motion data first, and it can follow along even with motions it was never taught. The official line is 20,000 hours of human motion data, zero-shot generalization. I happen to be trying imitation learning these days, so I hooked this up too while I was at it.

I don't have a full robot, just a remote-access setup — I turn on the camera on my end, and it drives the robot on the other end. The pitfalls are basically all in the first half.

Changing motions really doesn't require retraining. I deliberately made a motion it definitely had never seen: bend down, go around to the side, move the bottle on the table half a turn. It followed along, with the amplitude and rhythm intact; end-effector precision was so-so, but the direction was right. This is a different world from traditional programming. Before, when a customer changed a process step, you'd spend two days labeling data; now at least you don't have to start from zero.

One brain, multiple forms isn't pure marketing either. At a customer site I saw the same model switched onto a wheeled base with dual arms, and the motion style was consistent. For teams building multi-form products, what you save is the cost of reinventing the wheel.

The bad parts are obvious too. Calibration — the docs make it sound easy, but in practice I fiddled with it for almost an entire afternoon. Skeleton alignment, coordinate origin, motion amplitude normalization — if any one of these steps is off, the robot's followed motion comes out crooked. I only have an ordinary camera, no motion-capture suit, so jitter gets passed straight through.

Latency is also a problem. I make a motion, it follows, and the gap in between is visible to the naked eye. Fine for demos, not okay for close human-robot collaboration on a process step. If the network drops, it just stops there, and I couldn't find a clear statement on the emergency-stop strategy in the docs.

Wang Donglin says this algorithm is at least six months ahead of similar international teams. I'm skeptical. Around the same time, I saw public materials from another company at CVPR also working on a general cerebellum, and their figure was "2 billion frames" — the two don't even use the same unit for data volume, so there's no direct comparison.

And acceptance metrics — this is what I care about most. The metric they give is generalization success rate; what my customer wants is cycle time, repeat positioning accuracy, and running eight hours straight without a hitch. Two different languages.

Depends on the situation. For showrooms, research platforms, motion data collection, early validation of multi-form products — this thing is usable right now; what you save is retraining and calibration time, and the integration barrier isn't that high. For production lines, assembly, unattended operation — don't rush; this demo is still far from mass production.