The hardest jobs
are still done by people.

The robot revolution skipped the hard part.

The current wave of robotics is going to give the world a lot of robots that can pour coffee, fold laundry, and walk around warehouses without bumping into shelves. That is real progress. It is also — bluntly — not where the leverage is.

The leverage is where humans pay the highest price for showing up. A structure fire. A subway tunnel after a derailment. An avalanche field. A combat casualty under fire. A radiation room nobody can stand near. These are the environments where every additional second of human presence is a tradeoff against death.

These are the environments where we have the least automation. Not because the work is unimportant. Because the work is hard.

Why this hasn’t happened yet.

The standard story is that extreme environments are too hard for current robotics. That’s true. The longer story is that they’ve also been too narrow a market for general-purpose AI labs and too generic a problem for defense primes.

Defense companies build for one customer. Consumer-robotics companies build for billion-unit markets. The gap in between — tens of thousands of high-stakes operators across firefighting, EMS, search-and-rescue, hazmat response, and defense logistics — has historically not been served by either.

It needs a foundation-model approach: train once, transfer everywhere. One backbone that runs on a quadruped tonight, a humanoid next quarter, and a tracked platform after that. The model layer that has reshaped language and vision has not yet been built for physical operators in chaotic environments.

That’s the gap we’re building into.

What we believe is wrong with the current approach.

Generality is not optional. Most robotics labs ship a system that handles their training distribution and breaks the moment the world has a different texture under it. The fix is not more handcrafted modules. The fix is more diverse data, more cross-embodiment training, and architectures designed from the ground up to transfer.

Simulation is not a shortcut. It’s a multiplier. Anyone who tells you they’ve closed the sim-to-real gap on contact-rich tasks is selling something. The right framing is that simulation lets you afford to make the real-world data the part that’s actually load-bearing.

Human oversight is the product, not a constraint on it. The operators we’re building for are professionals. They do not want a fully autonomous black box on their scene. They want an asset that does what they tell it, faster and with less fatigue than a human teammate, and that hands control back the moment anything feels wrong. That’s the system we’re designing.

Why now.

Three things changed at once.

One: foundation-model training is no longer the bottleneck. The infrastructure, the recipes, and the talent that built language and vision models all exist now. The same playbook works for physical action; it just requires a different data substrate.

Two: sensorimotor data is suddenly cheap. Affordable motion capture, browser-deployable SLAM, and consumer-grade teleop hardware mean a small team can generate the volume of demonstration data that used to require an entire department. (We have published work in this area.)

Three: hardware is catching up. Off-the-shelf quadruped and humanoid platforms now exist that are credible in the field. A foundation model is no longer waiting on a custom robot to run on. It is waiting on the model.

What we will and won’t do.

We will work with lawful operators — fire services, EMS, search-and-rescue, hazmat units, and government partners under clear rules of engagement — on missions that protect lives and reduce harm.

We will publish capability and limitation data honestly. We will not invent benchmark scores. When we deploy, we’ll publish near-misses alongside successes. Our safety commitments are not a compliance exercise.

We won’t ship capabilities directly from simulation. Every new task family goes through controlled real-world trials before broader deployment.

We will build weapons systems — including armed robots — for lawful operators. What we won’t do is hand the machine the force decision: weapons release always requires explicit human authorization. That is not a marketing line; it is contractually how we license deployments.

We won’t overstate capability. Our public posture will lag what we’ve actually shipped, not lead it.

The bet.

The bet we’re making is that in the next decade, the deciding factor for whether a firefighter goes home from a shift will be how well their robot partner handled the scene. The deciding factor for whether a casualty survives the first hour after an attack will be whether a robot could close the distance with stretcher and supplies. The deciding factor on a rubble field will be whether a search agent can map the void spaces faster than a human team can dig.

We don’t think this is a niche. We think this is what extreme-environment physical AI is for.

If you want to work on this, we’re hiring.