In 1988, a robot named HelpMate started making deliveries in a hospital.
Not on a track. Not following a wire buried in the floor. Not in a roped-off corridor after hours. It drove itself through ordinary hospital hallways, in the daytime, around nurses and gurneys and visitors and wet-floor signs, using wheel odometry corrected by sonar, infrared, and vision. It carried meal trays, sterile supplies, medications, patient records, lab specimens, and mail. It ran up to twenty-four hours a day, seven days a week.
More than fifty of them went into real hospitals. This was not a demo video. It was a fleet.
The company that built it was Transitions Research Corporation. In 1997 it went public and renamed itself after its one product, HelpMate Robotics. On October 15, 1999, it sold its assets to Pyxis, the automation arm of Cardinal Health, for $13 million.
Thirteen million dollars. Eleven years after the first working robot. That is less than a seed round today. By 2003, under new ownership, the installed base had grown from about fifty units to about seventy.
Read the failure carefully, because it is not the one you expect
Here is what did not go wrong.
The autonomy worked. The navigation worked. The robot did not need a rebuilt building or a special floor or a human babysitter. It solved a genuinely hard perception and planning problem in an environment full of unpredictable people, and it did so with 1988 compute, which is to say with less processing power than the thermostat I installed last month.
What went wrong was everything downstream of the engineering. A hospital is a slow, committee-driven capital buyer. Every unit needed a sales cycle, a site survey, an install, training, and ongoing service, and each of those costs money that does not scale the way software costs scale. Fifty units in eleven years is not a manufacturing problem. It is a go-to-market problem. The machine was ready long before anyone figured out how to sell and service it profitably.
Why I bring up a thirty-eight-year-old robot
Because I get asked a version of the same question on most calls, and it is always framed as a technology question.
"Shouldn't I wait until this stuff gets good?"
It got good in 1988. What has never reliably worked is the part where somebody sells you a machine at a price you will actually pay, installs it this month instead of next year, and is still around in three years to answer the phone when it acts up. Thirty-eight years of robotics history says the engineering arrives first and the economics arrive late, if at all.
So when I designed what I sell, I designed the business model first and let the technology be as boring as it needed to be.
The boring version, on purpose
One off-the-shelf sensor goes on one piece of equipment that matters: a pump, a compressor, a boiler, a rooftop unit. It feeds a small local box running a small model. For the first couple of weeks that box does nothing but learn what your specific machine's normal is, which is not the same as any other machine's normal. The vibration signature it always has. The current it pulls on a ninety-degree day versus a thirty-degree one. The way it starts, the way it settles.
After that, its entire job, forever, is to notice when today stops looking like yesterday, and to tell a human about it three to six weeks before that drift becomes a failure.
It never drives anywhere. It has no arms, no wheels, no map, no personality. It is bolted to one machine and it pays attention. That is the whole product.
The pricing is deliberately unromantic: $99 to $199 per month, hardware under $3,000, installed in an afternoon. Nothing leaves the building, so there is no cloud bill underneath and no per-query API charge that grows as your equipment gets chattier.
I have a sump pump in Watertown that has been running on exactly this setup for over a year, and a forty-device building in Northampton doing the same thing at larger scale. Neither of them is impressive. Both of them work, and both of them have a number next to them.
The market data agrees, if you read the boring column
Two forecasts came out this week. Edge AI hardware is a $33.3 billion market in 2026, headed for $81 billion by 2032. Edge AI software is $3.1 billion, headed for $11.9 billion.
Hardware is more than ten times the size of software. Sit with that for a second, because it means the sensors and boards and chips, the things that feel like the product, are the commodity. They are getting cheaper on a schedule. A vibration sensor that cost $600 a few years ago is $50 now, and STMicroelectronics started shipping a part this month that runs the first pass of anomaly detection inside the sensor package itself, from $25 in volume.
The model is not where the value is either. Production-grade anomaly detection runs on hardware costing under $200, with latency under ten milliseconds. That is not a moat. That is a Tuesday.
Look at how the software forecast is actually segmented and you find the answer sitting there in plain language: it splits the market into solutions and services, meaning consulting, deployment, and support. That second category is the smallest thing on the chart and it is the only part that does not get commoditized, because it is not a component. It is knowing which of your forty devices is worth watching. Where on the housing to mount the accelerometer so you are reading the bearing and not the floor. How long to wait before you trust the threshold. Who to call at two in the morning, and whether they pick up.
What this means if you own a building instead of a robot company
- "Wait until the tech is ready" is answering the wrong question. The tech has been ready for narrow jobs for decades. Ask instead whether anyone will sell you a working version at a price you can afford, on a timeline you can live with, with support that outlasts the install.
- Scope is the entire safety margin. HelpMate had to handle every hallway in every hospital. My monitor has to handle one pump. That difference is why one of them cost eleven years and $13 million to not-quite-work, and the other one installs in an afternoon.
- Cheap parts are good news for you, not for vendors. When the sensor drops to $25, the honest vendor stops charging you for the sensor and starts earning the fee on judgment and response. Ask any monitoring vendor what, specifically, your monthly payment buys. If the answer is mostly hardware, the price is going to look silly in two years.
- Fifty units in eleven years is a warning about your vendor's survival, too. Whatever you install in a mechanical room, somebody has to still be there when it needs attention. Boring, small, and cash-flow positive is a feature when you are the one depending on it.
Last week I wrote about three hundred robots debuting in Shanghai without a single price tag. This week's story is what happens on the other side of that gap, to a company that did have a price and a product and real customers, and still could not make the arithmetic close. Those two posts are the same argument from opposite ends. The frontier is worth every dollar going into it. But the thing that stops your basement from flooding next spring does not need to come from the frontier, and it should not cost like it did.
HelpMate deserved better than a $13 million exit. It was a magnificent piece of work. It is also the clearest evidence I know that in physical AI, shipping something small that pays for itself beats shipping something remarkable that does not.
Your equipment does not need a robot. It needs something paying attention.
Each critical asset, your compressor, your boiler, your pump, gets one off-the-shelf sensor and a small model on a local box. It learns that machine's normal over the first couple of weeks, then watches for the day it drifts, catches trouble three to six weeks before a failure, and hands it to a person who knows the equipment. Nothing leaves the building. $99 to $199 per month, hardware under $3,000, installed this week.
See how it worksSources: HelpMate case study via Weekly Robotics #369, July 20 2026. HelpMate technical record (trackless navigation via odometry corrected by sonar, infrared and vision; meals, sterile supplies, medications, records, specimens and mail; 50+ units operating in uncontrolled hospital environments up to 24/7) via IEEE Xplore conference publications on HelpMate, the Springer “Autonomous Mobile Robots” perspective chapter, and NASA Spinoff 2003. Transitions Research Corporation founded 1984; renamed HelpMate Robotics on its 1997 public listing; assets acquired by Pyxis Corporation (Cardinal Health) October 15 1999 for $13 million; approximately 70 units installed by 2003, via Emerald “Industrial Robot”, Mergr deal summary, and NASA Spinoff. Edge AI hardware market $33.30B (2026) to $81.12B (2032) at 15.87% CAGR, report released July 17 2026; edge AI software market $3.12B (2026) to $11.86B (2032) at 24.63% CAGR, report released July 20 2026, offering segmentation separating services (consulting, deployment, support) from solutions, both via GlobeNewswire. STMicroelectronics in-sensor-AI vibration sensor (ISPU-based, on-package FFT, filtering, envelope, velocity severity and anomaly detection), from $25 in 1,000-piece orders, availability July 2026, via Edge AI and Vision Alliance. Edge anomaly detection reference economics (production deployment on hardware under $200, latency under 10ms) via 2026 edge AI implementation surveys. Field deployments at The Intersecto Watertown sump-pump site and Northampton 40-device building. Companion brief: /Users/tdeshane/lobster/research/physical-ai-brief-2026-07-21.md.