Robot brain builders are pushing out of their GPT-2 era
Robotics companies developing "AI brains" for machines are racing to catch up with the rapid progress made in large language models, but experts warn that the sector remains in an early, immature phase. This was underscored this week when Unitree, China's leading robot maker, lost nearly half its market value shortly after a blockbuster IPO valued the firm at $66 billion, as analysts pointed to robots' continued inability to perform genuinely useful, value-creating work despite improving physical capabilities.
At last week's Actuate conference, a gathering for developers building robot AI systems, attendance had tripled since 2023 to around 1,500 people, reflecting soaring interest, but also exposed a persistent "robotics data crisis" — a shortage of high-quality training data. One founder, Harry Mellsop of simulation-tools startup Antioch, likened the field's current state to "GPT-2 era" language models, saying more data, compute and better simulation hardware are needed to progress. Autonomous vehicle firms are seen as furthest ahead, thanks to abundant real-world driving data, and companies including Tesla, Wayve and Uber are now applying that expertise to humanoid robotics projects, though some, like Genesis AI's CEO, argue it remains too early to settle on a single hardware approach.
- Robot AI development lags behind LLM progress despite huge investment
- Unitree's valuation halved days after its blockbuster $66bn IPO
- Firms lack quality training data; autonomous vehicle expertise seen as a shortcut
New here? Start with this
Robot “brains” are AI systems designed to help machines understand their surroundings and decide how to act. They are being developed by robotics companies, while large language models such as GPT have advanced quickly in recent years.
A major obstacle is the shortage of useful training data: robots need examples of how to handle varied real-world situations. Developers are trying to fill the gap with computer simulations, while companies such as Tesla, Wayve and Uber are bringing experience from self-driving vehicles, which gather large amounts of driving data.
The field is still at an early stage, and a robot that can move well may not yet be able to do work that is useful or commercially valuable. That makes progress important to companies and investors weighing how quickly robotics could become practical.
Both sides, in good faith
The strongest fair case each way — we don't pick a winner.
The case for
The robotics AI sector is following the proven path of large language models, which progressed from their "GPT-2 era" to remarkable capability in just two years through more data and compute. The tripling of conference attendance, massive investment, autonomous vehicle breakthroughs proving the approach works, and new simulation tools addressing the data shortage all suggest robotics is approaching its inflection point. Historical precedent shows AI progress has consistently exceeded sceptics' timelines.
The case against
Despite years of investment and visible technical progress, robots still cannot perform economically valuable work—the true test of whether AI brains have solved robotics. The post-IPO collapse of Unitree from a $66 billion valuation reflects realistic market scepticism about near-term viability. The structural reality of the "robotics data crisis" is harder to overcome than optimists suggest; unlike autonomous vehicles with billions of real-world miles, robots need diverse interaction data that simulation cannot adequately substitute. Field leaders acknowledging they remain in a "GPT-2 era" suggests breakthrough remains years away, not imminent.