The iRobot Founder Bet Against Humanoids. Your Building Should Take the Same Bet.

Two stories sat next to each other on the IEEE Spectrum robotics front page this week. The first was a profile of Hello Robot's Stretch 4, a wheeled arm-on-a-pedestal that the editors framed as "practical home automation." The second was the return of Colin Angle, the founder of iRobot, with a new venture called Familiar Machines & Magic — companion robots built to support daily wellbeing in the home.

Neither robot has legs.

That is the news. The only consumer-robotics founder who has ever shipped a product at scale, on his second swing at the home market, chose not to build a humanoid. The most-watched practical home robot of 2026 rolls on wheels. Both of them were on the front page in the same week that Figure and 1X announced their humanoid manufacturing ramps.

The funding is going to both ends of the spectrum. But the credibility this week went to the practical end.

Why the wheeled bet keeps winning at the bench

Bipedal locomotion is a research problem with a defensible thesis: if you want a robot that can use staircases, kitchens, and door handles designed for humans, you eventually need legs. The thesis is correct. The cost of being correct, today, is that you spend most of your compute keeping the robot upright instead of solving the task.

The Hello Robot answer is that homes have floors, and a wheeled platform with a vertical extension covers 95% of the surfaces a domestic robot needs to reach. The Colin Angle answer is similar — companion robots do not need to climb stairs to be useful at daily-wellbeing tasks. Both companies are routing their engineering budget into perception, manipulation, and the long messy middle of human-robot interaction, not into staying balanced on two contact points.

The bet is: solve the boring problem with cheap hardware and proven control loops. Let somebody else fund the bipedal research arc.

The same bet shows up in building monitoring

This is the bet that has structured every building-monitoring deployment I have shipped. There is a "humanoid" path for instrumenting a 30,000 square foot commercial building. It looks like a $50,000 building automation system, a vendor-specific dashboard, a service contract that requires a vendor-certified integrator for any change, and a five-year refresh cycle. It exists. It works in some buildings.

The "wheeled" path is what shows up in our case studies. The agentic sump pump watching a basement with an edge node and a microphone runs 97 recovery cycles in a single rainy night, with no cloud round trip and no vendor lock-in. The community-center smart-building deployment cut 42% off energy costs in six weeks using off-the-shelf sensors and an MQTT bridge.

Neither deployment is a research demo. Neither one would survive a humanoid budget. Both of them work because the same architectural shift Hello Robot and Familiar Machines just bet on — practical hardware, proven control loops, perception layered on top — is now cheap enough to ship into a single building at a price the owner will sign.

The architectural floor keeps dropping under the same bet

Two supplier-side data points from the past seven days that make the wheeled bet cheaper to take this quarter than it was last quarter:

• Espressif's ESP32-S3 with ESP-DL shipped a reference design for human-activity recognition running on accelerometer data, on-device, no cloud. The same MCU costs about $4 in single units. The implication for building monitoring is that the floor for "tiny ML on a streaming sensor" is now well below the cost of running the wire to the sensor.
• Skydio raised $110M Series F at a $4.4 billion valuation, explicitly to expand US drone manufacturing. The capital is for drones, not buildings, but every dollar that flows into US-manufactured edge inference hardware shortens the supply chain for everyone else who deploys at the edge. The cost of integrating a sensor module into a small-building deployment was 60 days of engineering in 2022. It is closer to 14 days in 2026.

The week's third data point is more direct. Nexus Labs' May 6 Owner Signal briefing surfaced that Microsoft runs four parallel ML models — linear regression for occupancy timing, random forest for ramp times, and two more — across roughly 50 buildings on the Redmond HQ campus, retrained every day. Microsoft is the largest sophisticated commercial-real-estate owner that publishes anything about its own architecture. The fact that the loop is daily, not weekly or monthly, is the news. The fact that it is four small models running in parallel, not one giant model, is the architecture.

That is the wheeled bet, exactly. Practical, cheap-per-model, retrain often, route the engineering budget into the part that actually moves money.

What this means for a service contract you write this quarter

The marketing surface of the physical-AI industry will keep being humanoids and demos through 2026. That is fine. It generates the funding that funds the supplier-side stack that the wheeled deployments are built on. None of it is bad.

What changes this quarter is what a small-building service contract should explicitly promise. Three concrete pieces of language to add, all backed by something that landed in the news this week:

1. "We run small, practical models on this building's data, retrained often." Backed by Microsoft's published 4-model daily-retrain loop. The proposal-defense conversation that used to take 45 minutes now collapses to a hyperlink.
2. "We bet on practical edge hardware, not vendor lock-in." Backed by Hello Robot Stretch 4 and Familiar Machines & Magic — both shipping practical solutions on cheap hardware in the same week. The owner who has been watching humanoid videos on YouTube now has a counter-example to anchor the conversation against.
3. "We expect the floor under sensor costs to keep dropping, and we will not lock you into a five-year refresh." Backed by Espressif's $4 MCU running tiny ML and Skydio's US manufacturing build-out. The deflation curve is real and the customer is buying into it, not against it.

What I'm watching next

Three follow-ons over the next month that would convert this week's signal into structural change:

1. Whether Hello Robot or Familiar Machines publish a price. The Stretch 4 has been research-grade for years. If either company prices a unit at the cost of a mid-tier laptop, the wheeled-vs-bipedal argument settles for the consumer market and the same argument propagates back into commercial real estate.
2. Whether Microsoft publishes a longer-form description of the four-model HVAC loop. Right now it is a paragraph in a Nexus Labs briefing. If Microsoft puts a blog post or a paper on it, every practitioner deploying ML for HVAC scheduling has a citable reference architecture and the proposal cycle gets shorter again.
3. Whether the Embedded Vision Summit (May 11–13 in Santa Clara) produces a new module announcement that further compresses the integration cost on small-building deployments. The summit is running as this post publishes. The thing to watch is whether any of the 90 sessions surface a born-aligned sensor module that lands at a sub-$100 price point.

The bottom line

Demos are how a field gets funded. Practical hardware is how a field gets deployed. The headline robotics stories of the past seven days happen to point at the practical end. The headline building-monitoring story of the past seven days — Microsoft's daily retrain loop — points there too. Espressif and Skydio just confirmed the floor underneath both bets keeps dropping.

The right thing to do this quarter is write the proposal that names the practical bet explicitly. The right thing to do this year is keep shipping the wheeled version of physical AI into the buildings that need it, while the rest of the industry sorts out whether the bipedal version eventually scales.

Read the case studies and related posts: How edge AI prevented a basement flood | 42% off energy costs on a community center | Closing the construction loophole on a small building | Physical AI is not a robot