Here is the question I get more than any other, usually about thirty seconds after someone decides they like the idea of a monitor on their equipment.
"The pump is brand new. It has never broken. So what is your monitor even going to learn from? It has nothing to compare against."
It is a good question, and for a long time my answer was a little hand-wavy. This week a robotics paper answered it for me, cleanly, from a completely different corner of the field.
A robot lab ran into the exact same wall
Researchers at Stanford and UC Berkeley published something called RoboReward. The job of the model is to watch a video of a robot doing a task, picking up a block, placing it in a bin, and score how well it is going, one to five. That score becomes the reward signal the robot trains against. A good judge means a robot that learns fast. A bad judge means a robot that learns garbage.
They found something that maps directly onto my problem. A judge trained only on successful attempts is a bad judge. It has watched a thousand videos of the task going right, so it knows what "right" looks like, but it has never once seen the task go wrong. Ask it to spot trouble and it has no idea where the line is, because it has only ever stood on one side of it.
That is the same model I would build if I only trained on a healthy pump. Feed a model a thousand hours of a pump running fine and it learns, precisely, what "running fine" looks like. It does not learn what "about to fail" looks like, because you never showed it. Those are two different skills, and only the second one is worth paying for.
Their fix is my fix: manufacture the failure
Here is the clever part. RoboReward did not wait around for robots to fail on video. They manufactured failure, two ways.
First, they took successful videos and trimmed them, so the model saw the task half-done, a third-done, stalled partway. Real footage of not-yet-finished. Second, they used a language model to write task descriptions that deliberately did not match a video, a "validated failure," so the judge learned what a mismatch between intent and outcome looks like. Trained on success plus this engineered failure, their 8-billion-parameter judge beat GPT-5 mini and Google's Gemini robotics model at scoring real robot attempts, and a robot trained on its feedback went from a 10 percent success rate to 50 percent.
They did not need to break real robots to teach the judge what broken looks like. They constructed the negatives. And that is exactly how you solve the "my pump has never failed" problem, because a pump hands you the raw material almost for free.
A machine writes its own failure curriculum
A pump does not go from perfect to seized in an instant. It degrades. A bearing broadcasts rising vibration, then a little heat, then a change in current draw, for weeks before it actually gives out. Every one of those in-between states is the physical version of RoboReward's trimmed video: a real, recorded example of not-quite-normal, generated by the machine itself, that you can label and learn from long before anything catastrophic happens.
So the monitor is not sitting there waiting for the pump to die so it can finally learn something. It is reading the slow slide, the daily drift away from the baseline it established in its first couple of weeks. The failure is not a single event at the end. It is a gradient the machine is climbing the whole time, and the gradient is the curriculum.
You do not wait for the flood to teach the model what a flood looks like. You teach it on the drift that comes weeks before the flood, and on every other pump that has ever failed.
And you seed day one with a million other pumps
The other half of RoboReward's trick was synthesizing negatives with a language model, giving the judge examples it had never actually witnessed. There is a clean physical equivalent.
Your specific pump has never failed. But pumps, as a class, have failed a million times, and those failures have well-documented signatures, the vibration shape of imbalance, of cavitation, of a bearing losing its grease. You can seed a detector with those known signatures on the day you install it, so it walks in already knowing what a failing pump sounds like even though this pump has been nothing but healthy. The machine in front of me is a stranger. The failure modes are not.
| RoboReward (robots) | A building monitor (pumps, motors, compressors) |
|---|---|
| Trained only on success = bad judge | Trained only on a healthy machine = blind to failure |
| Trimmed videos to show partial progress | The machine degrades in readable steps for weeks |
| Language model synthesizes "validated failures" | Known fault signatures for the machine's class, seeded on install |
| Needs a benchmark of 2,831 hand-verified examples | One machine, one task, graded by reality every morning |
And notice the last row, because it is where the small-building job is actually easier than the robotics one. RoboReward is trying to build one judge that generalizes across many robots and many tasks, which is genuinely hard, which is why it needs a big hand-verified benchmark. My monitor has the opposite problem, the friendly version: one machine, one job, and a basement that is either dry or not every single day. The cold-start window is real, but it is small, and reality closes it fast.
Why this is the whole game for a small building
In a factory there are engineers who can eyeball a gauge and say "that does not sound right." A small building, a church basement, a small shop, a rental with a boiler in the corner, has nobody down there. The monitor is the only thing standing between a slow degradation and a 2 a.m. disaster, and it can only play that role if it genuinely knows the far side of the line.
That is why I do not install a monitor that has only ever met a healthy machine. Each detector shows up already knowing the failure signatures for its class of equipment, learns your specific machine's normal over the first couple of weeks, and then reads the drift, the same slow slide RoboReward learned to catch in its trimmed videos, and raises its hand three to six weeks before the failure. It never needed your pump to break first. It just needed to have been shown, on purpose, what breaking looks like.
It doesn't wait for your equipment to fail first.
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. It clamps on in an afternoon, arrives already knowing the failure signatures for that class of machine, learns your specific machine's normal, and watches vibration, temperature, and current for the drift that shows up weeks before a breakdown. Then it tells a person who knows the equipment. Nothing leaves the building. $99 to $199 per month, hardware under $3,000.
See how it worksSources: The Batch, Issue 360, "New Ways to Train Robots" (RoboReward vision-language reward model from Stanford and UC Berkeley, arXiv 2601.00675; 4B and 8B reward models scoring robot task progress one to five; "validated failures" created via a language model generating mismatched command descriptions and via trimming successful videos to partial progress; RoboRewardBench of 2,831 hand-verified examples; RoboReward 8B at 0.665 mean absolute error versus GPT-5 mini at 0.691 and Gemini Robotics-ER 1.5 at 0.906; real-robot object placement 50 percent success with RoboReward rewards versus 10 percent with the competing model and 75 percent with human rewards), July 3 2026. Weekly Robotics #367 (Luxonis $14M Series A for OAK on-device cameras; Copper-rs v1.0.0 stable Rust robotics middleware), July 6 2026. Field deployments at The Intersecto Watertown sump-pump site and Northampton 40-device building. Companion brief: /Users/tdeshane/lobster/research/physical-ai-brief-2026-07-06.md.