Physical Intelligence raises $70M seed

Physical Intelligence, a robotics startup building foundation models to control robots directly, emerged from stealth with a $70 million seed round, according to reporting on the raise. Investors included Thrive Capital, OpenAI, Khosla Ventures and Sequoia Capital, valuing the company at roughly $400 million.

The company’s approach differs from most robotics work of the period, which paired narrow, task-specific controllers with separate perception and planning systems. Physical Intelligence instead builds what it calls vision-language-action models: a single network that takes in camera images and natural-language instructions and outputs the low-level motor commands that drive a robot’s joints and actuators, aiming for one model that generalises across robot bodies and tasks rather than one trained per robot.

The founding team combined robotics and machine-learning researchers from Google Brain, Stanford and UC Berkeley, including Karol Hausman, who had led robot-manipulation work at Google Brain, and Sergey Levine, a prominent figure in deep reinforcement learning for robotics.

The seed round was an early instance of foundation-model methods, and foundation-model money, moving from language into physical control — a pattern venture investors would fund heavily through 2024 and 2025 as the “robot foundation model” category grew. Physical Intelligence itself raised substantially larger rounds within the following year as the company published its first models and moved toward commercial deployment.