Manifesto
LLMs learned to talk and think by reading everything we’ve ever written, but we’ve never recorded how we move.
Solving robotics is fundamentally different from solving language, and the humanoid problem makes that especially clear. There’s no internet’s worth of data waiting to be scraped, and the physical world is notoriously unforgiving. Video comes closest, but it shows what our bodies do, not the forces behind it. So every humanoid company has to collect, or manufacture, its own data.
Even the largest players fall back on training in simulation, synthetic data or paid data-collection operations. None of these scale.
A high-quality, full-body embodied dataset the size of the internet does not exist, and no amount of money spent on today’s methods will create one.
The physical world needs its own internet: an enormous, passively generated stream of high-quality full-body data. There is only one way to build it, and that is to make a physical product people choose to use en masse.
Solution: Powered Exo
I stress powered. Wireless mocap suits already exist and capture much of the same motion, but you’ll never get millions of people to wear them. Mocap asks people to work for the data; an exo works for them.
You need a device people want to wear: one that makes them perform better than they otherwise could, and collects high-quality data along the way. It can’t be a hindrance in any way, as today’s exosuits are. If it’s even mildly annoying, no one will want one.
And everyone must want one. When a product clears that bar, embodied humanoid data will be solved.
That’s what we’re building.
