Physical AI has become one of the hottest sectors in venture investing, with companies raising billions of dollars to apply the technologies behind Large Language Models to robotics. This surge of interest helped fuel a massive IPO for Unitree, China's leading robot maker, which saw its valuation hit $66 billion upon listing on the country's equivalent of the NASDAQ. However, the excitement quickly soured as the company lost nearly half its value within weeks. Analysts point to a fundamental issue: while robots' physical capabilities have improved dramatically, they still lack the intelligence to perform value-creating tasks reliably.
The Data Crisis in Robotics
At last week's Actuate conference—a gathering of developers building AI brains for robots—the mood was optimistic yet cautious. The event has tripled in size since its 2023 debut, drawing 1,500 attendees, according to organizer Foxglove, a company that helps physical AI model builders manage and visualize their data. But the underlying risks were on display: a booth for Avala, another physical AI infrastructure player, promised to solve 'the robotics data crisis.'
That crisis is the shortage of high-quality training data for AI models. Attempts to build generalized robots capable of any task remain far off, and end-to-end learning for specific tasks hasn't yet delivered reliable, commercially viable performance. For developers, the path forward mirrors the strategies of frontier AI labs: curate more diverse datasets, experiment with different training regimes, and design better reinforcement learning scenarios.
The 'GPT-2 Era' of Physical AI
Harry Mellsop, co-founder of Antioch—a startup building simulation tools for model builders—describes physical AI as being in its 'GPT-2 era,' referring to the OpenAI model that predated ChatGPT. To overcome this plateau, he argues, the field needs more data and compute, particularly GPUs optimized for ray tracing to create high-fidelity simulations. The analogy is apt: just as GPT-2 laid the groundwork for ChatGPT, today's rudimentary robot brains may precede a breakthrough that remains out of reach without further investment in infrastructure.
Autonomous Vehicles Lead the Way
The most advanced area of physical AI is autonomous driving, partly because companies can collect vast amounts of real-world data from human-driven cars, and partly because the primary objective—avoiding collisions—is simpler than manipulating the physical environment. Much of the tooling for robot model-building actually comes from AV companies; Foxglove, for instance, was founded by former employees of Cruise, General Motors' erstwhile self-driving subsidiary.
Now, these automotive players are increasingly betting that their investments in machine learning infrastructure will let them compete with dedicated humanoid robot makers. Tesla is already pursuing this with its Optimus robot, while AV-focused firms like Wayve and ride-hailing giant Uber have launched robotics labs dedicated to humanoid form factors as R&D efforts.
A Shared Foundation, Different Embodiments
'I think you need to start in vehicles... manipulation robotics is like self-driving five years ago,' Alex Kendall, CEO of Wayve, told TechCrunch. 'The data infrastructure, the simulation, ML ops infrastructure—will probably be shared, but the specific world model for the simulator will require different post-training. There's going to be a lot more commonality than not, but there will also be differences for different embodiments.'
Kendall believes it's premature to commit to any single hardware platform. Advances in sensors and other components are arriving quickly, and a truly general model should be more agnostic. As 2026 unfolds, the industry is bracing for a period of consolidation, where data infrastructure and simulation tools become commoditized, and the differentiators will be the ability to collect high-quality data and train models that can adapt across diverse robotic forms.
The road ahead is steep, but the parallels to the early days of LLMs offer both caution and hope. If physical AI can replicate the trajectory from GPT-2 to GPT-4, the current 'data crisis' may eventually be seen as the necessary growing pain of a transformative technology.
via TechCrunch