Salim Ismail has a dog that "literally looks like a teddy bear." His worry, he told the panel, is that a household robot will look at it, decide "oh, laundry," and put the dog in the washing machine. "That's going to be a bad day," he said. "Worse for the dog."
The joke carried a serious argument about when robots will move from factories and hazardous sites into ordinary homes. It came up on an episode of Moonshots with Peter Diamandis, recorded on October 16, 2026, and published the next day. Diamandis, founder of XPRIZE and Singularity University, was joined by Ismail, founder of Open ExO and a general partner at Exponential Venture Capital; Emad Mostaque, founder of Intelligent Internet; and Anousheh Ansari, CEO of XPRIZE. The panel did not settle the question. Diamandis closed it with: "We're going to find out."
A robot for every job
Diamandis opened with a picture of where manufacturing might lead. He pointed to Jeff Holden's company Atomic Machines, which he described as going "from prompt to product": you describe what you want, and a machine builds it with millimeter-level accuracy from multiple materials. If that works, he said, someone who needs a robot to clean windows, repair air-conditioning systems, lay bricks or mow the lawn could ask for the ideal design, have it built and then put it to work. "So do we end up with Star Wars with like thousands of different robot designs?" he asked.
The household-robot application is Diamandis's own extrapolation. In his October 7, 2026, post on the company's blog, Holden describes an AI-native digital manufacturing system that combines several processes and materials, and says its immediate focus is microscale machinery: tiny devices that work through mechanical, fluidic, thermal, optical, chemical and electrical effects.
Dull, dirty and dangerous first
Ismail answered that this kind of automation works best where tasks repeat often, then spreads to related work such as plumbing, which involves many small manipulations within one trade. For now, he said, robots will start with what he called DDD jobs, "dull, dirty, dangerous": toxic cleanups, underwater manufacturing and similar work. He expects it to "be a while" before they enter homes, citing liability and what he called institutional lag.
The application that excites him most is surgery. When Diamandis suggested surgery hardly counts as dangerous or dirty, Ismail said it is dangerous for the patient. Nobody wants to be among a surgeon's first ten heart operations, he said; you want to be around the 500th. Too far down the line has its own risk, because after a few thousand operations a doctor "gets lazy" and stops paying attention. A robot could smooth that experience curve out, he argued, first supporting the best surgeon and over time doing the operations itself, working at the highest possible precision.
Diamandis added that the one question to ask a surgeon is how many times they did the procedure that morning, since the surgeon who has seen the most varied cases is most likely to succeed. With federated learning across all robots, he said, the machines would have seen millions of cases and would "share all those memories." Federated learning is a way of training one shared model on data held by many separate machines without pooling the raw data in one place.
The kitchen got its own moment. The discussion turned to a robot chef that would cook the best food for a person all the time in their own kitchen. Diamandis extended the idea, imagining sensors embedded in the body that measure protein, hydration and vitamin levels, so the robot cooks exactly what a person needs at that moment, "and it tastes good."
Heavy machines, small children
Ansari said a safety problem has to be solved before robots enter homes. "These are still big, heavy machinery," she said. A dog, a cat or a small child moving around one could simply get stepped on.
The discussion noted that 1X makes its home robot "soft and squidgy" for this reason. 1X's description of its NEO robot explains that softness as hardware design: a deformable three-dimensional lattice material cushions the internal components, covered joints keep fingers away from pinch points, and tendon-driven actuation, in which cables pull the limbs much as tendons move human joints, is meant to keep movements low in energy. The company also describes household autonomy as something that develops over time. For chores the robot has not learned, owners can schedule a human expert to supervise it remotely.
Ismail's laundry scenario raised a different problem: edge cases, the rare situations a system was never trained for. There are "infinite numbers of edge cases," he said, and he pointed to driving. Autonomous driving has taken a long time to get right, he said, even though it is "a very bounded domain" that is reasonably well understood.
Diamandis pushed back. The robots' physical performance is improving, he said, but they will also be "running super intelligent models." He did not think such a model would look at "your fluffy little dog and say laundry." It would recognize the dog and "make decisions better than the majority of humans," he said, because the bodies and the models are improving at the same time.
The conversation then turned to why general-purpose models matter here. A specialized model might put the dog in the laundry, the argument ran, but a generalized one would not. When Ismail asked whether this required world models, AI systems that build an internal simulation of how the physical world behaves, the answer was no. The recent generations of leading AI models had "suddenly understood 3D," the reply went, and could create whole 3D assets for games, though "no one's quite sure how." That unexpected leap, the argument continued, is why they are now generalizing.
Within a year, or after the supply chain
Diamandis also disputed the idea that robots would be confined to dull, dangerous and dirty work. Ismail clarified that he had said they would start there. Diamandis thought they would move past that stage within the next year, if the dexterity and capability the panel had been discussing were really achieved.
Ismail's reply rested on economics. Robots will pay for themselves in those industrial roles for the first few years, he predicted, because "we have no supply chain right now for making billions of robots." Bringing them into general households would take billions of humanoid robots, he said, or at least millions in wealthy countries. Until manufacturing catches up, "the ROI won't be there," meaning the return on investment would not justify the cost for most homes. He labeled this "my prediction."
Is the humanoid the right shape?
Ansari questioned a premise behind much of the debate. "I really don't believe we ultimately, the humanoid form factor is the right form factor," she said. Some tasks do not suit human bodies at all; a robot may need to be small enough to fit through a narrow opening to reach a location. She argued for specialized bodies matched to each task, or for something like Transformers, the toy and film robots that change shape. General models would help with that, she said, because nobody would need to build a separate model for every body type. "Got someone on your side?" Ismail asked, and thanked her.
The discussion also turned to the cost of hands. One remark held that the most advanced robot hands now cost about $8,000, down from $20,000, with as much dexterity as a human hand or more. A laundry robot does not need that, the remark continued, so spare-parts companies might appear that sell add-ons such as extra arms, again in the style of Transformers.
Ismail built on Ansari's point with an image of a modular robot: a body "with a bunch of sockets in it," fitted with "a spider type thing for one thing, and a regular arm for another thing, and a rope for another thing," and buyers around the world choosing among the parts.