Robot brain builders say they are moving past their GPT-2 stage

Robotics foundation model founders are now describing their own field’s trajectory in specific, self-critical terms: past the earliest generation of robot brains, but not yet at the equivalent of a language model’s ChatGPT breakthrough. Harry Mellsop, founder of the simulation startup Antioch, told TechCrunch that physical AI is moving out of its “GPT 2 era,” referencing the OpenAI model that predated ChatGPT by roughly three years. Coming from someone raising money in the category, the comparison reads as candor about how much runway is left, not a confession of being stuck.

The stakes showed up first in public markets, not conference booths. Unitree, China’s leading robot maker, went public at a $66 billion valuation, then lost nearly half that value within the week, according to TechCrunch. Analysts cited in the report point to a specific gap: the robots keep improving physically while the software running them still cannot reliably do value-creating work.

That gap is why data, not hardware, dominated the conversation at Actuate, a developer conference for physical AI builders that TechCrunch reports has tripled in attendance since 2023, reaching 1,500 people this year. One infrastructure vendor at the event, Avala, marketed itself around solving what it called “the robotics data crisis.” The pitch names the actual bottleneck: general-purpose robots need far more diverse training data than any single company has collected.

Autonomous vehicles are the furthest along, and the reasoning TechCrunch’s sources gave is structural rather than about who has raised the most money. Cars generate training data automatically from millions of human drivers, and their core task (avoiding contact) is simpler than the manipulation problems humanoid robots must solve. Alex Kendall, CEO of the self-driving company Wayve, told TechCrunch that “manipulation robotics is like self-driving five years ago,” and argued the simulation and ML-ops tooling built for cars will carry over to humanoids even as the underlying world models diverge by embodiment.

Not everyone in the piece agrees that shared infrastructure is the right bet. Théophile Gervet, CEO of the vertically integrated humanoid company Genesis AI, disagreed. Genesis closed a $105 million seed earlier this year, and Gervet told TechCrunch it is “too early in this wave for a brain strategy to work,” arguing hardware and software still need to be co-designed rather than bolted onto a shared general model.

Gervet drew a sharper competitive line on strategy: a startup building a narrow application on a GPT-2-equivalent model will get “crushed” once a rival ships on a GPT-4 equivalent, so vertical-focused robotics firms are effectively racing the underlying model layer as much as each other. That vertical-versus-general split already shows up in whose robots are actually deployed. Bedrock has excavators operating autonomously on job sites, Gritt has moved into building solar farms, and Agility has machines running in industrial settings, while general-purpose humanoids, by contrast, remain confined to labs. “No customer cares about the general purpose robot that works at 80% success rate,” Gervet told the outlet.

The piece surfaces no agreement on what a breakout moment for the sector would even look like. Kendall pointed to consumer vacuum robots as the largest current deployment and said the real test is exciting consumers rather than investors, floating sub-$1,000 eyes-off driving autonomy as one marker. Foxglove CEO Adrian Macneil told TechCrunch he doubts a single “ChatGPT moment” exists at all for the category, comparing the more realistic outcome to the gradual Apple II or IBM PC eras of early personal computing rather than a one-week distribution spike.

For operators evaluating physical AI vendors over the next quarter, the practical read from TechCrunch’s reporting is to weight deployment data and vertical focus over general-purpose demos. The founders closest to the technology say general humanoid intelligence has moved past its earliest stage but is not close to solved, and that the near-term winners will be narrow products that ship reliability, not breadth.

Reporting by Tim Fernholz for TechCrunch, published August 26, 2026.