Physical AI’s GPT-2 Era: Why Robots Still Can’t Do the Work

Physical AI startups are raising billions to bring large language model techniques to robotics, but the sector is hitting a wall: robots lack the data and intelligence to perform value-creating tasks reliably. This reality was underscored this week when Unitree, China’s leading robot maker, lost nearly half of its market value after a $66 billion IPO on China’s STAR Market, as analysts pointed to the gap between physical capabilities and practical know-how.

Why the Data Crisis Is Holding Back Physical AI

At last week’s Actuate conference in San Francisco, developers building AI brains for robots gathered to confront what Avala, a physical AI infrastructure company, called “the robotics data crisis.” The event, organized by Foxglove, tripled in size since 2023 to 1,500 attendees, reflecting both excitement and urgency. The core problem: there isn’t enough high-quality training data to teach robots generalized skills, and end-to-end learning for specific tasks hasn’t yet produced reliable commercial products.

Harry Mellsop, founder of Antioch, a startup building simulation tools, likened physical AI to the “GPT-2 era” of OpenAI—before ChatGPT proved the power of scale. He believes more data and compute, especially GPUs optimized for ray tracing to create high-fidelity simulations, are needed to push the field forward. Autonomous vehicles are furthest ahead because they can collect real-world driving data and the primary task is collision avoidance, not manipulation.

AV Companies Pivot to Humanoids

The tooling developed for autonomous driving is now being repurposed for humanoid robots. Foxglove was founded by former Cruise employees, and both Wayve, an AV startup, and Uber have launched robotics labs focused on humanoid form factors. Alex Kendall, Wayve’s CEO, told Bitcoin World, “You need to start in vehicles… manipulation robotics is like self-driving five years ago.” He argues that data infrastructure and simulation will be shared, but world models will need different post-training for different embodiments.

However, not everyone agrees that AV expertise translates directly. Théophile Gervet, CEO of Genesis AI, which raised a $105 million seed round this year, countered, “We’re too early in this wave for a brain strategy to work; there’s lots of opportunities to co-design hardware and AI.” His company is vertically integrated, building both hardware and software, to address the co-design challenge.

Vertical vs. General: The Strategy Debate

Robotics companies are split between those targeting specific tasks and those pursuing general-purpose humanoids. Gervet noted, “No customer cares about the general purpose robot that works at 80% success rate,” but added that if you build for a narrow vertical on top of a weak model, “you’re going to get crushed by the company building on GPT-4.” Vertical players like Gritt (solar farms), Agility (industrial settings), and Bedrock (autonomous excavators) are already deploying robots, gaining real-world data but risking obsolescence as general models improve.

Bedrock CTO Kevin Peterson explained that starting with excavation helps understand “manipulation in the wild,” but the company plans to build an intelligence layer across construction machines. Managing the dense visual and lidar data is a challenge, which is why Foxglove announced a new product built on Nvidia’s Cosmos world model, allowing engineers to search data with natural language queries for faster debugging.

When Will Physical AI Have Its ‘ChatGPT Moment’?

Sam Altman recently predicted a ChatGPT moment for physical AI in a few years, but experts have differing views. Kendall points to consumer vacuum bots as the largest robot deployment, and says a true breakthrough would be “eyes-off autonomy for less than $1000 [worth of hardware] in a car”—a market Wayve is targeting by licensing models to automakers. Gervet defines the moment as “manipulation that just works out of the box,” where a robot can understand natural language and perform basic tasks with 80%+ reliability.

Foxglove CEO Adrian Macneil disagrees with the concept entirely. “There will not be a ChatGPT moment for robotics,” he told Bitcoin World, because distribution in the physical world is far harder than digital. He hopes for an “Apple II moment” where consumers can buy a home robot that does useful things. The path forward remains uncertain, but the industry is betting that solving the data crisis will unlock the next wave of value.

Conclusion

Physical AI is at a critical inflection point, with billions in investment but limited real-world utility. The data crisis, the vertical vs. general debate, and the challenge of distribution are the key hurdles. As companies like Wayve, Genesis, and Foxglove push forward, the winners will likely be those who can bridge the gap between impressive hardware and reliable, valuable software.

FAQs

Q1: Why did Unitree’s stock drop after its IPO?
Unitree’s valuation fell by nearly half as analysts noted that while its robots have advanced physical capabilities, they lack the AI software to perform value-creating tasks reliably, undermining investor confidence.

Q2: What is the “robotics data crisis”?
It refers to the shortage of high-quality training data needed to teach AI models to control robots effectively. Without diverse and reliable data, models cannot generalize to new tasks or environments.

Q3: Will autonomous vehicle companies lead in humanoid robotics?
Some experts like Wayve’s Alex Kendall believe AV infrastructure will be shared, but others like Genesis AI’s Théophile Gervet argue that hardware-AI co-design is necessary, making it too early to say which approach will dominate.