Adcock maps four stages of humanoids, Goldberg asks if data is dead
Two of the loudest voices in the field put opposite bets on the table in the same week.
Brett Adcock laid out a four-stage model of humanoid maturity, and each stage is gated by the one before it. First, build great humanoid hardware. Second, make that hardware work with a whole body, AI-first architecture. Third, scale intelligence. Fourth, scale manufacturing and integrate humanoids into the economy at massive scale.
These first two chapters can’t be brute-forced with capital
The split that matters in his post is where money helps and where it does not. The first two stages cannot be brute-forced with capital, he says, because they need deep engineering across hardware, controls, AI, and systems integration. The last two can be bought. Doing them correctly, in his framing, will require tens and eventually hundreds of billions of dollars.
His claim about stage three is the one worth arguing over. Once you are past the hardware and the control stack, Adcock says the primary bottlenecks become data and compute, and that humanoid robotics will ultimately require far more of both than large language models did. That is a big number to assert. Nobody has shown it, and Adcock does not show it either. It is a bet about how much of the physical world a policy has to see before it generalizes.
Chris Paxton, who works in the field and posts regularly about what does and does not transfer to real robots, backed it in one line. He called it a great take on humanoids.
Goldberg pushes the other way
Days later Ken Goldberg posted from IROS 2026 with the opposite question. He described many conversations at the conference about the future of robotics and automation, then asked whether agentic robotics built on pretrained LLMs could replace what he called the quixotic quest for demonstration data. Or, paraphrasing Nietzsche, is data dead.
The two positions are not a fight anybody settled. Adcock's version says the wall ahead is data and compute, so pour capital into collecting and training. Goldberg's version says the wall might be an illusion, because a model that already knows a lot about the world and can plan with tools may not need mountains of teleoperated demonstrations to act in it.
What neither post contains is evidence. Adcock is describing an outlook, not results. Goldberg is asking a question at a conference, not announcing a finding. Read them as two bets on where the next few billion dollars of robotics spending should go, from two people with very different incentives and vantage points.
Could agentic robotics with pretrained LLMs replace the quixotic quest for demonstration data?
I think this is a great take on humanoids