
Can't Bring the Helix Model Home, So I Tested It a Different Way
A friend got me onto Helix. He works at a humanoid robot integration company, and he sent me a video on WeChat, saying Figure's robot went into a customer's home, found things on its own and picked them up, and told me to check if it was all hype.
I said sure. But let's get one thing straight first: Helix isn't an app you can download, nor is it hardware you can bring home. It's a vision-language-action model from Figure AI — in plain terms, the decision-making system installed in a humanoid robot's brain. Vision-language-action, abbreviated VLA, means it can process what it sees and what it hears at the same time, then directly output actions, with no human translating in between. It runs on the Figure 03 robot. I don't have a Figure 03, and it's not easy to get one in China.
So I took a different approach and crashed at his lab for an afternoon. I've been filming robot dog videos for a while, and I've had hands-on time with humanoid robots for about a month, so I'm not a total outsider looking at this stuff.
His company doesn't represent Figure, but they've accumulated a lot of first-hand material, plus a recording of an on-site customer operation. The operator interface I saw was simpler than I expected — just an input box. You type "put the things on the table into the fridge," and the robot breaks down the steps itself, then executes. My friend let me type two commands myself: one to grab a cup, one to gather scattered items into one place. The cup one worked. The gathering one got stuck for a moment — the robot hesitated for a few seconds over a soft package before moving.
Helix's core design is a dual-system setup. One is called S2, responsible for understanding scenes and language, thinking at roughly seven to nine times per second. The other is called S1, responsible for turning thoughts into concrete actions, running at two hundred times per second. The idea is that the slow one thinks about the big picture, and the fast one keeps the hands from shaking. In the recording, two robots carry things together — one turns, the other immediately adjusts its posture to follow. That moment was genuinely smooth.
The pleasant surprise is in generalization. In the official demo, it picked up a cactus it had never seen before — spiky, irregularly shaped — and still managed to grab it. It relies on the model recognizing things itself, without pre-programming every object. My friend said when they reproduced it internally, they also tried everyday items they hadn't seen before, and the success rate was higher than he expected.
Where it gets stuck is also very real. That S2 frequency of seven to nine times per second means its understanding has gaps — when it hits a situation requiring immediate reaction, it can only rely on S1 to tough it out. Early versions of Helix only controlled the upper body, around thirty-five degrees of freedom, with the legs managed by a separate system; the full-body version that merged walking, balance, and manipulation into one network came later. Also, when the scene changes, performance drops. Change the room layout, change the lighting, and the same task takes several more tries.
My most direct takeaway from testing it is that it connects understanding human speech and doing physical work into one chain, but that chain isn't thick enough yet. Whether it's worth your attention depends on whether you have a robot that can run it.
There are several things it does well. Natural language directly connected to actions, without training separately for each task — that's genuinely convenient. Generalized grasping is among the best in its class. The multi-robot collaboration segment isn't a edited effect; it makes sense logically. The problems are also very real: noticeable latency, still immature handling of sudden situations, an extremely high barrier to entry — basically no chance for individuals and small teams — and the performance drop when changing environments shows it's still a ways off from being stable.
So the conclusion is: it depends. If you're doing home service robots, have a full-robot team, and need generalized grasping capability, it's worth keeping an eye on, using its technical approach as a reference. If you're a small team wanting to quickly build a demo and sell products, don't benchmark against it — your data loop can't support this kind of play. As for content creators and individual users like me, in the short term it has nothing to do with you; just watching the demos is enough.
If you really want my advice: if you're working in embodied intelligence, don't rush to chase Helix's demo effects. First ask yourself one question — how many hours of your own scenario data do you have? No matter how strong the model is, without data to feed it, it's just a video.
Physix Frontier