General Intuition, a startup building "embodied AI", argues that robotics is on the verge of a shift comparable to the arrival of ChatGPT and foundation models. Its chief executive, Pim de Witte, contends that instead of building specialised robot models trained on vast amounts of task-specific real-world data, the industry should develop general-purpose foundation models that transfer an intuition about movement and interaction across many robots and environments. The claim matters because, if correct, it could sharply reduce the data and effort needed to build capable robots.

The company trained its foundation model on millions of hours of video game data, including which controller buttons a human pressed and when, which de Witte and lead investor Vinod Khosla see as key to human-like spatial and temporal reasoning. General Intuition raised $320 million last month at a $2.3 billion valuation. It says the same model can play a video game for hours and power a four-legged robot after fine-tuning on just eight minutes of real-world data, navigating an office using only a front camera. Rather than making robots itself, the firm aims to be the base model that other robotics companies build upon.