Tutor Intelligence opened the doors to DF1 last month. It is, by their own claim, the largest robotic data factory in the United States. 100 bimanual semi-humanoids in a converted mill in Watertown, Massachusetts, picking sponges and bags of snacks. 45 to 50 remote teleoperators in Mexico and the Philippines, scoring the picks. A vision-language-action model called Ti0, training on what the system collects.
The MIT-CSAIL spinout raised $34 million in Series A funding last December to build it. They have now built it. Their CEO, Josh Gruenstein, framed the bet plainly to The Robot Report: there is no Wikipedia for robots, you cannot simulate your way around that fact, and the only path to scale is collecting real-world data with humans in the loop. He compared DF1 to the Large Hadron Collider — an instrument of discovery for physical AI.
I have a sump pump in a basement that has been doing this same job, for free, for two years.
What DF1 Is Actually Producing
DF1's output is not a model. It is data. Specifically, it is real-world manipulation traces — what the robot saw, what the human teleoperator commanded, what the robot did, whether the result matched the goal. Every one of those traces is a labeled training example for Ti0.
Tutor Intelligence claims that running 100 robots in parallel against the same policy lets them detect and correct errant behaviors 100 times faster than running a single robot. An edge case that would take eight hours of robot operation to surface in a one-off setup shows up in five minutes of DF1 operation. That is the LHC analogy made operational. They are not training a smarter robot; they are running a particle accelerator that throws bags of chips at a hundred grippers until something interesting falls out.
The interesting thing that falls out is structured: a stream of input-output pairs grounded in actual physical reality, labeled by humans who watched what happened, ready to be fed into a learning algorithm.
What the Sump Pump Has Been Producing
Two years ago I put a wifi-connected float sensor and a current sensor on a sump pump in a residential basement. It cost less than four dollars in parts. The pump cycled every time it needed to. Most cycles, the float swung up cleanly and the pump engaged. Some cycles, the float hesitated for a second or two before engaging. Once, last spring, it ran for forty seconds without the water level dropping — a sign the impeller was packed with debris.
Every one of those cycles is a labeled training example. The label was generated by physics. The water either left the pit or it didn't. The pump either drew the expected current or it didn't. The basement either flooded or it didn't. The sensor stream is the input, the outcome is the label, and the labels arrived for free, every time.
Two years of that. 730 days. Tens of thousands of cycles, each one a real-world manipulation trace from an embodied agent operating in actual physical conditions.
This is, structurally, the same asset class as one shift of DF1 output. The robotics industry is paying enormous sums to manufacture it. The building practitioner inherits it.
The Mistake Most Building Owners Make
When I describe this to a building owner, the response is almost always the same: "So I should wait until the AI gets smarter, and then start collecting data."
It is the exact wrong sequencing. The model can change next month. Open-weight VLAs from NVIDIA, Hugging Face, Genesis AI, and now Tutor Intelligence are dropping every few weeks. What you cannot do next month, with any amount of money, is retroactively create the last two years of pump cycles. The basement either has the data or it doesn't.
This is the same logic that drives Tutor Intelligence's whole business. They could have spent the $34 million on bigger models. They didn't. They spent it on the data collection rig, because they understand that without the data, the model is nothing. The model is a multiplier on the data. Zero data times any multiplier is still zero.
The 730-day-old pump times a smarter reasoner is a real number. The pump that started monitoring last week times the same reasoner is also a real number, but it is much smaller, and it cannot be made bigger by waiting.
What the $14/hour Price Tells You
Tutor Intelligence is also selling a single-arm version of their robot, called Cassie, at $14 to $18 per hour, usage-based, no contract. Productiv, a kitting company in Virginia, told The Robot Report the robot was profitable from Day 1. The traditional alternative was a $150,000 capex purchase from an integrator.
This is the same pricing shape as building monitoring. No capex, no integrator markup, no multi-year contract, walk away anytime. The difference is that Cassie has a hard upper bound on hours per month — there are only so many shifts in a week. A sensor pack on a sump pump runs 24 hours a day, every day, and it costs about $0.14 per hour to do so. Two orders of magnitude cheaper than Cassie, for an asset that never sleeps.
Tutor Intelligence has just printed the public reference price for usage-based, no-contract physical AI. It will become harder, not easier, for traditional integrators to defend $150K capex deals against this. Building monitoring sits on the same trend line, two orders of magnitude further down the cost curve.
What to Do With This Week
If you are operating a small commercial building or a one-off industrial asset, the move this week is not to wait for the smarter reasoner. It is to start the recording. The sensor pack costs less than a service call. The data it produces compounds every day it stays online. When the next Ti0-class model lands and runs locally on a Jetson — and it will, probably this year — the building that has been recording for two years has something to fine-tune against. The building that started yesterday has a sample size of one.
Tutor Intelligence is paying $34 million and 50 teleoperators to manufacture this asset class. The building practitioner gets it for the price of a four-dollar sensor and the discipline to keep the recorder running.
Start your building's data factory this week.
We deploy edge AI monitoring on small commercial buildings using off-the-shelf sensors. Pump cycles, panel readings, HVAC performance, vibration patterns — all captured locally, all yours, ready for the next generation of reasoning models. Under $500 to start, no contract, walk away anytime.
See What We BuildRead the case studies: Making a sump pump agentic | Buildings don't need a world model — they have yesterday's data | The sim-to-real gap doesn't exist in your building | The four-dollar sensor lies about your basement