Robot 'brain' builders admit physical AI is still in its early days
At the Actuate conference, developers building AI systems for robots acknowledged that physical AI remains far from mature, while Chinese robot maker Unitree lost nearly half its market value this week. Experts point to a lack of quality data as the key barrier to robots reliably performing useful work.
Investment in physical AI has surged in recent months, but industry insiders acknowledge the technology remains far from mature. That was underscored this week when Unitree, China's leading robot maker, saw its valuation—$66 billion after its IPO on a Chinese stock exchange—drop by nearly half. Analysts say that while robots' physical capabilities keep improving, they still lack the ability to reliably perform work that creates real value.
At the Actuate conference, which gathers developers building AI "brains" for robots, attendance has tripled since 2023, reaching 1,500 people this year, according to organizer Foxglove. Alongside the enthusiasm, there was clear awareness of the industry's central risk: a shortage of high-quality training data, sometimes called the "robotics data crisis."
The 'GPT-2 era'
Harry Mellsop, founder of simulation startup Antioch, said physical AI is currently in its "GPT-2 era," referencing the OpenAI model that preceded ChatGPT. Progress will require more data and computing power, particularly graphics processors suited to building high-fidelity simulations.
Autonomous vehicle developers are furthest along, partly because they can gather useful data from human-driven cars, and partly because avoiding collisions is simpler than manipulating physical objects. As a result, carmakers—including Tesla with its Optimus robot, along with Wayve and Uber—have begun launching humanoid robotics efforts of their own.
General-purpose versus specialized robots
The industry is also debating whether to build general-purpose robots or ones focused on narrow tasks. Companies with a specific focus, such as Gritt in solar farm construction or Bedrock in autonomous excavation, are already deploying robots in real settings, while general-purpose humanoids largely remain confined to labs.
Opinions differ on what a "ChatGPT moment" for physical AI would look like—some point to affordable autonomous driving, others to a robot that can reliably follow natural-language instructions for everyday manipulation tasks.
