Robot brain builders are pushing out of their GPT-2 era

Physical AI has become one of venture capital’s favorite bets, with billions flowing toward efforts to bring the tricks behind large language models into robotics. The pitch is simple enough. If models can learn language from vast internet-scale data, maybe robots can learn to move through the world the same way. The reality, as the industry keeps discovering, is messier.

Unitree’s big China listing showed how much money is chasing the idea. The robot maker briefly hit a $66 billion valuation after its IPO on China’s NASDAQ equivalent, then lost nearly half that value this week. Analysts say the problem is not whether robots can move better. It’s whether they can do work that matters, consistently, in the real world.

That gap was all over last week’s Actuate conference, where developers building AI brains for robots gathered in force. Foxglove, the organizer, said attendance reached 1,500 and that the event has tripled in size since it started in 2023. A booth from Avala advertised a fix for “the robotics data crisis,” a blunt reminder that high-quality training data is still the bottleneck. General-purpose robots remain far off, and even task-specific end-to-end systems haven’t yet produced dependable commercial results.

So the field is trying to borrow from frontier AI playbooks: collect or synthesize more diverse datasets, vary training setups, and improve reinforcement learning. Harry Mellsop, founder of Antioch, called physical AI its “GPT 2 era,” suggesting the industry still needs much more data and compute before it clears the next hurdle. He pointed to GPUs optimized for ray tracing as especially important because they help produce high-fidelity simulations.

Autonomous vehicles are the furthest along, partly because cars already generate useful real-world data and partly because driving is mostly about avoiding contact rather than manipulating objects. That’s why so much of the tooling comes from the AV world. Foxglove itself came out of former Cruise employees. And now the crossover is getting more explicit: Tesla is pushing Optimus, while Wayve and Uber have both launched robotics labs aimed at humanoid form factors.

The argument inside the sector is no longer just about whether robots will work. It’s about where to start. Wayve CEO Alex Kendall says vehicles are the right place, because manipulation robotics is where self-driving was five years ago. Genesis AI CEO Théophile Gervet disagrees, arguing it’s too early for a pure brain strategy and that hardware and AI should be co-designed. His company raised a $105 million seed round this year.

Meanwhile, the businesses actually shipping robots are usually narrow ones. Gritt is building solar farms, Agility is deploying in industrial settings, and Bedrock is operating excavators autonomously. Those systems may not be general, but they are out in the field, collecting the kind of data that matters. Foxglove’s new product, built on Nvidia’s Cosmos open weight world model, is aimed at helping engineers search dense visual and lidar data with natural language queries so they can debug faster and run better evaluations and simulations.

The big open question is what a real robotics breakthrough will look like. Kendall says consumers, not investors, will decide that, and he points to cheap eyes-off autonomy in cars as one possible moment. Gervet wants out-of-the-box manipulation that works at roughly 80% reliability. Foxglove CEO Adrian Macneil thinks robotics may never get a single ChatGPT-style spike at all, because real-world distribution is much harder. He’d settle for something more old-school: an Apple II moment, or an IBM PC moment, where ordinary people can buy a home robot that does useful things.