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

The significant market value drop for China's Unitree, following its $66 billion IPO, underscores a critical challenge for physical AI in Asia and globally. While billions are flowing into startups aiming to bring large language model techniques to robotics, the core issue remains a lack of high-quality training data and the intelligence needed for robots to reliably perform complex tasks. This situation is akin to the "GPT-2 era" for OpenAI, suggesting that a breakthrough moment for physical AI, where robots can perform manipulation tasks reliably out of the box, is still several years away. The current focus on repurposing autonomous vehicle tooling for humanoid robots, as seen with Foxglove and Wayve, reflects an industry attempting to bridge this data gap, but its direct applicability to manipulation remains debated. Companies like Genesis AI are pursuing vertically integrated approaches, co-designing hardware and AI, rather than relying solely on a "brain strategy." This divergence highlights a fundamental tension: whether to build general-purpose humanoids that currently operate at an 80% success rate, or to focus on narrow vertical applications like those by Gritt in solar farms or Bedrock in autonomous excavation, which gather real-world data but risk obsolescence as general models improve. For Asian markets, the success of these vertical players will be a key indicator of practical AI adoption, as customers prioritize reliability over general capabilities.