Robotics labs are spending billions to generate real-world data for training physical AI, overcoming a critical training material deficit. Unlike web-scraped text models, robotics requires diverse, complex datasets. Companies like Generalist, Physical Intelligence, Neura, and Skild are building "robot gyms" with bimanual manipulators to emulate physical tasks. Start-ups are deploying recording equipment globally to capture first-person video footage, collecting millions of hours of "egocentric" data. "There is no internet for us to download from," said Joe Fox Jr of Scale AI. "It is an arms race to get the best data."
Robotics companies raised nearly $48 billion year-to-date according to PitchBook, making it a hot AI venture sector. Start-up Mecka raised $60 million from Nvidia and Sequoia. Physical Intelligence, valued at $5.6 billion after raising $600 million, tested models in San Francisco Airbnbs. Figure, a humanoid robotics firm, outlined a scheme paying users to record home tasks, amassing roughly 44,000 active users paid about 94 cents per task. The Silicon Valley company plans over $1 billion in computing and services spending. Encord, Scale AI, and other data firms servicing OpenAI and Anthropic are opening robot gyms worldwide. Robotics demands more diverse data than large language models, requiring depth-perceiving hardware and precise force calculation. Simulation and "world models" aim to bridge the data gap.
Source: Financial Times Companies · Summarized by HeadlinesBriefing