Robot "brain" builders say physical AI is stuck in its GPT-2 era

Physical AI is one of the hottest sectors in venture investing, with companies raising billions to apply the tools that built large language models to robotics. That excitement helped deliver a big IPO for Unitree, China’s leading robot maker, valuing the company at $66 billion after its debut on China’s Nasdaq equivalent — before the stock lost nearly half its value this week, as analysts pointed to a familiar problem: while robots’ physical capabilities are improving, they still lack the know-how to do value-creating work reliably.

At last week’s Actuate conference, a gathering of developers building AI “brains” for robots that has tripled in size since 2023 to 1,500 attendees, according to organizer Foxglove, the excitement was clear — as was the underlying risk, captured by a booth for physical-AI infrastructure player Avala promising to solve “the robotics data crisis.” That crisis is the lack of high-quality training data for AI models; attempts to build generalized robots that can do any task remain far off, and end-to-end learning still hasn’t delivered commercially reliable products for specific tasks. Antioch founder Harry Mellsop described physical AI as being in its “GPT-2 era” — the OpenAI model that predated ChatGPT — needing more data, compute, and better reinforcement-learning training regimes, particularly GPUs optimized for the ray tracing used in high-fidelity simulations.

Autonomous vehicles are furthest along, in part because relevant data can be collected from human-driven cars and because avoiding contact is an easier goal than manipulating the physical environment; much of the model-building tooling in the space, including Foxglove itself, comes out of autonomous-vehicle companies. Wayve CEO Alex Kendall argued it’s too early to commit to one hardware platform, since sensors and other components are advancing quickly and a truly general model should stay hardware-agnostic; both Wayve and Uber have launched robotics labs betting their AV-derived tooling can extend to humanoid form factors. Genesis AI president and co-founder Théophile Gervet, whose company raised a $105 million seed round this year, disagreed, arguing it’s too early in the wave for a “brain-only” strategy and that there is still room to co-design hardware and AI together. Robotics companies targeting specific tasks — Gritt building solar farms, Agility deploying industrial robots, and Bedrock operating excavators autonomously — are getting into the field, while general-purpose humanoids remain largely lab-bound. “No customer cares about the general-purpose robot that works at 80% success rate,” Gervet said. “We see a lot of other players go general, but there is no value provided because there’s no vertical focus.”

Foxglove announced a new product this week, built on Nvidia’s Cosmos open-weight world model, letting engineers search dense visual and lidar data with natural-language queries to speed up evaluation and simulation. On when physical AI might have its own “ChatGPT moment,” Kendall pointed to eyes-off autonomy for under $1,000 of hardware in a car as his benchmark, while Gervet described it as manipulation that “just works out of the box” at roughly 80%-plus reliability. Foxglove CEO Adrian Macneil was more skeptical of the framing altogether, arguing ChatGPT’s moment came from instant, mass distribution that physical products can’t replicate, and that he’d rather see robotics reach an “Apple II moment” — the point where people can simply buy a useful home robot.

Why it matters

If physical AI really is in its GPT-2 era, the capital pouring into humanoid and robotics-foundation-model startups is pricing in a leap — from lab demos to reliable, general-purpose deployment — that, by developers’ own account, several more years of data and compute are needed to reach.