This week 1X unveiled the production hand for its NEO home robot, and the demo was genuinely stunning: 25 degrees of freedom, driven by tendons that work like ropes, wrapped in a sensor "skin" that feels pressure across the fingertips. It pours tea. It zips a jacket. It sorts grapes. The editor of Weekly Robotics, who watches this field for a living, wrote that a week earlier he'd have guessed we were two years away from hands that good.
He also gave it the best description I've read all year. He called the hand "an API to the physical world."
That phrase is worth stopping on, because it's exactly right, and it quietly explains why I run a monitoring business and 1X is trying to build a home robot with a person hidden inside it.
Every API has two halves
If you've ever written software against an API, you know it has a read side and a write side. Reading asks the world a question: what's the temperature, what's the status, what happened. Writing tells the world to change: move this, open that, turn it on. Reading is safe and cheap. Writing is where everything gets hard, because now you can break something.
A robot hand is the write side of the physical world's API. Its entire purpose is to act: to grip, pour, zip, sort. And here is the tell that most of the coverage skipped past. NEO, the most advanced consumer humanoid shipping this year, at around $20,000, still can't do the write side alone. For any task it hasn't mastered, 1X ships a fallback: a remote human "Expert" puts on a VR headset, looks out through the robot's cameras, and takes over the hands.
When you cannot fully automate an action even with a tactile robot hand and a person in the loop, that action is nowhere near solved for the messy open world. Writing to a home you've never seen, reaching for an object that's never in the same place twice, is genuinely at the frontier. It should be. It's the hard half.
My system only reads
Here is what I build. A small detector clamps onto one machine, a sump pump, a boiler, a compressor, and watches three signals: vibration, temperature, and current draw. A small model runs on a local device right there in the basement. It never grips anything. It never moves anything. It has no hand at all. When two of its three signals agree that something is wrong, it sends one plain-language text to one person.
That is the read side of the exact same API. And the read side, scoped to a single machine, is solved, cheap, and fully autonomous. No teleoperator. No human in a headset. No stranger watching a video feed. It just runs, 24 hours a day, and stays silent until it has something real to say.
This isn't because I'm cleverer than the team at 1X. It's because I picked the easy half on purpose, and I scoped it down to something closed. NEO has to re-find the world every single time it reaches, because the world moves. My sensor is bolted to one pump and never moves, so the hardest question in robotics, where is the thing and what does normal look like, gets answered once and then stays answered forever.
| Write side (act on the world) | Read side (observe the world) |
|---|---|
| A 25-DOF hand, tendons, tactile skin | Three cheap sensors on one machine |
| Must re-find the object every time | Bolted on; never moves |
| Still needs a human in the loop to do the work | Needs a human only to make the call, rarely |
| $20,000 and a payroll of remote Experts | $99 to $199 a month, no one watching |
| Frontier, unsolved for open spaces | Solved, shipping, running tonight |
Read-only is a feature, not a shortfall
For a long time I treated "it only monitors, it doesn't fix anything" as the apologetic part of my pitch. I had it backwards. Read-only is one of the strongest things about the system, and this week's news is why.
Think about what a building owner is actually afraid of. Not that a monitor will fail to notice a problem, that's the thing they're buying. What makes them hesitate is the opposite fear: that some automated system will do something, at 2 a.m., that it shouldn't. Trip a pump. Cut a compressor. Get hacked and turned into an action. Every one of those fears lives on the write side. My system doesn't have a write side. It physically cannot move, open, or actuate anything in your building, because it has no hand to do it with. There is nothing there to break loose or be hijacked into acting.
The machine reads. The human decides and acts. And the human only gets pulled in when the read is real. That division of labor is why the whole thing fits inside a $99 subscription instead of a VR headset and a staff of operators.
The privacy story is the same shape. The NEO model has, by design, a remote stranger who can see through the cameras in your living room. My monitor never streams video anywhere. Nothing leaves the building except a text message when something is genuinely wrong. "It can only watch its own machine, and it can never touch anything, and nothing it sees ever leaves your basement" turns out to be exactly what a nervous owner wants to hear.
Pick the half that ships
None of this is a knock on NEO. The write side of the physical world is the grand prize, and somebody has to keep pushing on it, tendons and tactile skin and remote Experts and all. I'm rooting for them.
But there's a whole business sitting on the read side, and it's open right now, today, with off-the-shelf parts and small models, while the write side is still figuring out how to reliably pour tea without a person in a headset. The frontier is teaching robots to act. The field, the part you can actually deploy this month, is teaching cheap sensors to read one machine well. I know which half is solved, because mine has been running in a basement for two years and it has never once needed a human to grab the controls.
Read-only by design. It can watch everything and touch nothing.
Each critical asset, your pump, your boiler, your compressor, gets its own small detector built from off-the-shelf sensors, running a small model on a local device. I clamp it on, teach it what normal looks like for that exact machine, and tune it to text a specific person, in plain language, only when something is really wrong. It can't move, open, or actuate anything, and nothing it sees ever leaves your building. $99 to $199 per month, hardware under $3,000.
See how it worksSources: 1X NEO production hand (25 degrees of freedom, tendon-driven / quasi-direct-drive, tactile-sensor skin; demos pouring tea, zipping jackets, sorting grapes) and "an API to the physical world" framing — Weekly Robotics #368, July 13, 2026, and 1X / trade coverage. NEO ~$20,000 early-access pricing, US home deliveries in 2026, 10,000 pre-orders in five days, and remote-"Expert" VR teleoperation fallback — 1X and trade reporting, July 2026. Edge predictive-maintenance economics (12–18% maintenance-cost reduction; 65% of teams planning AI adoption by end of 2026) — 2026 industry guides. Sensor-side commoditization ($600 in 2019 → under $50 in 2026) and the read-only monitoring approach — The Intersecto field deployments at the Watertown sump-pump site and Northampton 40-device building. Companion brief: /Users/tdeshane/lobster/research/physical-ai-brief-2026-07-14.md.