Robot AI developers are trying to break out of their “GPT-2 era”

Physical AI has become one of the hottest areas in venture investing, with companies raising huge sums to bring language-model-style breakthroughs to robotics. That enthusiasm was on display when Chinese robot maker Unitree went public and was valued at $66 billion. Within a week, however, its stock lost nearly half its value, with analysts pointing to a core problem: while robots' physical abilities keep improving, they still can't reliably perform work that creates real value.

At the Actuate conference, which has tripled in attendance since 2023 to around 1,500 people this year, excitement was matched by concern over what one exhibitor called a “robotics data crisis” — a shortage of high-quality training data. Building a truly general-purpose robot remains far off, and current end-to-end training methods haven't produced commercially reliable products.

Harry Mellsop, founder of simulation startup Antioch, argued that physical AI is currently in its “GPT-2 era” — the stage before the leap that produced ChatGPT. Getting past that will require more data and computing power, especially GPUs suited to high-fidelity simulation.

Autonomous vehicle developers are seen as furthest along, partly because they can gather driving data more easily and because their core task — avoiding collisions — is simpler than manipulating objects. That's why carmakers, including Tesla with its Optimus robot, along with Wayve and Uber, have launched humanoid robotics labs of their own.

The industry remains split over whether to pursue general-purpose models or narrow, task-specific ones. Companies with a narrow focus — such as solar-installation firm Gritt, industrial robotics company Agility, and Bedrock, which runs autonomous excavators — are already deploying robots in real settings, while general-purpose humanoids remain largely confined to labs.

Opinions differ on when robotics will have its own “ChatGPT moment.” Wayve CEO Alex Kendall points to affordable, reliable autonomous driving as the trigger, while Genesis AI CEO Théophile Gervet says it will come when a robot can reliably perform simple physical tasks from natural-language commands. Foxglove CEO Adrian Macneil is skeptical there will be a single breakthrough moment at all, predicting instead a gradual “Apple II”-style arrival of useful, affordable home robots.