Why we’re building HALO Dynamics.
We didn’t set out to build a robotics data company. We set out to build a sport.
The Origin Hunt was supposed to be a natively-VR sport — not a game you play sitting down with a controller, but something you play with your whole body, moving through a world that pushes back. To make that real, we had to build hardware nobody sells: a haptic suit that could put force on a hundred points across your body, shoes that could feel the ground and predict your next step before you took it, gloves that could sense what your hands were doing, and an omnidirectional treadmill so you could run forever without leaving the room.
We spent months heads-down on that stack. And somewhere in the middle of wiring the hundredth motor and debugging a flex sensor at 2 in the morning, it clicked: the thing we’d built to play a sport was, almost by accident, one of the most valuable scientific instruments in robotics.
Here’s why.
Every serious lab trying to build physical AI — humanoids, manipulation, general-purpose robots — is stuck on the same problem, and it isn’t compute. It’s data. Large language models had the entire internet to learn from. Robotics has no such thing. There is no internet of robotic data. A robot has to learn how the physical world responds when you touch it, push it, grip it, step on it — and that knowledge doesn’t exist in any dataset at scale, because the most important part of it is invisible to cameras.
You can watch a video of someone opening a bottle a thousand times and never learn the one thing that actually matters: how hard they gripped it right before the cap broke its seal, and how they eased off the instant it did. Force and contact are invisible to video. They’re invisible to almost everything the labs are currently training on. And they are exactly what our hardware measures directly.
That’s the whole thesis. The bottleneck to embodied intelligence has moved from compute to real-world, action-labeled data — and we own the instrument that captures the rarest, most decision-relevant slice of it: whole-body, contact-rich, force-labeled human movement. It’s self-labeling, too. We don’t need an annotation pipeline; the human body provides its own ground truth. Every grasp, every step, every impact comes with the force attached.
So we made a decision. HALO Dynamics isn’t a VR sports company that happens to have cool hardware. It’s the embodied-intelligence data foundry, and the sport is what proves the instrument works.
Longer term, the data is only half of it. We’re building a self-supervised world model that learns from this data the way a body learns — by predicting what it will feel next, and improving wherever it’s wrong. No labels, no reward engineering; the world supervises itself. This isn’t a wild bet. The approach has been shown to work — Meta’s V-JEPA 2-AC controlled real robots with no task-specific reward, trained mostly on raw video. What it has never had is the data we capture. That’s the part we think changes things. To be clear about where we are: we’re building this now, not claiming it’s done. The hardware is in hand. The first thing we have to prove — the thing everything else rests on — is that the prediction loop actually works on the data we capture. That experiment is what we’re running. We’ll show the curve when it moves, not before.
I’ll be honest about the stage we’re at, because I think founders who pretend to be further along than they are end up fooling mostly themselves. We’re early. The suit, the shoes, the gloves, the treadmill exist and work. The data pipeline is coming together. The model is a hypothesis we’re in the middle of testing. There’s a real chance the hard parts stay hard. But the instrument is real, the data it captures is genuinely scarce, and the timing — with the whole industry waking up to the data bottleneck at once — is not an accident we want to waste.
If you’re a builder who likes watching hard things get made in real time, we’re going to be writing here as we go — the wins and the 2am debugging both. And if you’re a lab that needs the kind of data video can’t give you, that’s exactly what we’re building. Come talk to us.
We built an instrument for a game. It turned out to be the data layer for physical intelligence. Now we’re going to find out how far that goes.
— HALO Dynamics