A dancer in a spandex suit, and why the humanoid form matters

At Tesla’s AI Day on Thursday, after a long and genuinely interesting technical presentation about their self-driving stack and the Dojo training computer, Elon Musk announced a humanoid robot. It’ll be about 5’8” and 125 pounds, move at up to 5 miles per hour, and do “dangerous, repetitive, boring” tasks. A prototype is supposedly coming next year. Then a person in a white bodysuit came on stage and did a robot dance.

The internet had a good time with it, and fair enough. There’s no robot. There are some slides and a costume.

But I think the underlying idea is right, and I’d take it more seriously than the demo deserves.

The argument for humanoids is not that the human form is the most efficient design for any single task. It isn’t. For moving boxes on a flat floor, wheels win. For welding car bodies, a fixed industrial arm wins. The argument is that the world is built for human bodies. Stairs, door handles, counters, shelves, tools, car seats, keyboards, clothes. A robot that has to work in human spaces without rebuilding them needs to fit in those spaces and use those tools, and the most direct way to do that is to be roughly human-shaped, with hands.

There’s a second argument that interests me more. If you want a machine to learn how the physical world works the way people do, by acting in it, then the body determines what it can learn. A human-shaped body with realistic joint ranges, torque limits and hands can learn from the same kind of interaction people have, and can learn from watching people, because its body maps onto theirs. A wheeled base with a gripper will learn something too, but it won’t learn to use a screwdriver, and it can’t learn by watching someone else use one.

Tesla does have some real advantages. They have a lot of experience building actuators, batteries and motors at scale, a vision system trained on huge amounts of real-world video, and a manufacturing culture. They also have a habit of announcing timelines that don’t hold, and a humanoid robot is an extremely hard problem. Boston Dynamics has worked on Atlas for years and it’s still a research platform.

My guess is that the first Tesla Bot we see will be clumsy, and that people will laugh at it too. That’s fine. The bodies are going to keep improving whether anyone laughs or not. The harder question is where the intelligence comes from, and I think the answer is a lot of learning in simulation and from video, with bodies realistic enough that what’s learned transfers. The hands will be the hardest part, and the most important.