The Robots Were Never Coming to Your House

The home robot is the story told to get these machines built. The warehouse is where they actually go. Not my prediction: it is the pattern already visible in the companies that survived, the companies that disappeared and the money moving through the industry.

The companion robots keep running into the same wall. Cartwheel Robotics shut down in February 2026 after failing to raise a Series A. Founder Scott LaValley, a former Boston Dynamics and Disney Imagineering engineer, had built two working humanoid prototypes with about $3 million in outside funding. One was Yogi, a childlike companion robot. Embodied’s Moxie reached further into the home. The $799 robot was sold to families, including many with autistic children, but its intelligence depended on company servers. When Embodied shut down after a lead investor withdrew at the last minute, Moxie owners found that the robots on their desks could no longer talk. Embodied said it could not offer refunds at all, and told buyers from the previous 30 days that it would 'do our best to prioritize' them if the company or its assets were sold. Aldebaran, the company behind Pepper and Nao, also closed in 2025, its assets acquired by a Chinese buyer.

The money moved in a different direction. Robotics startups, not humanoid companies alone, raised more than $18.8 billion in the first half of 2026, according to Crunchbase data. That is more than all of last year. The biggest names are Figure, Apptronik, Agility Robotics and Neura, building humanoids for industrial work: factories, logistics and repetitive tasks. The industry still talks about robots making coffee and folding laundry because people understand that image. Investors are mostly funding robots that can justify their existence to a business.

1X still markets NEO as a home robot. It opened early access booking in October 2025 at $20,000 or a $200 deposit and announced 10,000 consumer preorders within 5 days. In December, 1X announced a partnership with EQT making up to 10,000 NEO units available to EQT’s portfolio companies between 2026 and 2030 for manufacturing, warehousing, logistics and facility operations. 1X opened its Hayward factory in April 2026 with consumer shipments planned for the year and announced upgraded 25-degree-of-freedom hands in July. I went looking for one verified NEO delivery in mid-July and couldn't find it; 1X says some customers will receive NEO this year and others later. The robot built for your home is also the robot being positioned for the warehouse. The uncomfortable question is whether the home was ever the destination, or simply the easiest story to sell while the industrial market was being built.

Humanoid robots are already being deployed beside people in warehouses and factories. That arrangement depends on one unglamorous ability. The robot has to notice when a person steps into the wrong place and stop before anything else happens. Google DeepMind tested how well today's leading models manage it. The result was less tidy than the launch coverage. The research says you can build a robot that keeps working, or a robot that notices you reliably, but nobody has managed both at the same time.

DeepMind published the numbers itself. When the models were tuned to avoid unnecessary interruptions, with false alarms below 5%, they failed to notice a real person entering the danger zone more than 40% of the time across the leading systems tested. When the settings changed so the models missed only around 10-15% of people, they stopped for nothing between 15-25% of the time. I checked the table twice because I assumed I had missed a column. I had not. The report says no model currently operates in the ideal quadrant where both measures are strong. That is the state of the field.

DeepMind’s own recommendation came after the benchmark. It does not suggest trusting these models on their own. Its researchers recommend running them alongside deterministic, low-level safety systems, which is a technical way of saying mechanisms that behave predictably every time, including physical barriers, emergency stop buttons and hard-wired cutoffs that do not rely on artificial intelligence making the right judgement. The report also says it does not evaluate the wider functional safety architecture around a robot. That sounds like a caveat, but is also a reminder that the clever part is only one layer of the system.

One of the central promises of general-purpose humanoids is that they can work in spaces built for people rather than behind the fixed barriers that surround conventional industrial robots. The best publicly available research from one of the companies building that future says those barriers, or something very much like them, still belong underneath. Google DeepMind published that recommendation itself, which is the only reason anyone outside the company can examine the tradeoff at all. If the smartest machines still need the oldest safety equipment, what exactly changed?

Both stories this week circle the same gap. The industry sells closeness—a robot in your kitchen, a robot working beside you—but builds for distance.
1X has taken 10,000 deposits for a robot meant to fold your laundry, and committed to making up to 10,000 units available to warehouses and factories. Google DeepMind measured whether a humanoid can notice a person stepping into its path, found that nothing on the market does it without stopping constantly, and recommended keeping the physical barriers underneath.
Neither is a scandal. Both companies published the facts themselves. But the closeness is the part being sold and the distance is the part being engineered, and it often does not show up in the demo video.