This startup thinks robotics is about to have its ChatGPT moment

A startup is developing what it calls a foundation model for robotics, arguing that the field should shift from building specialized robot models for specific tasks toward creating general-purpose models that can transfer knowledge of movement and interaction across different environments. The approach mirrors how the AI industry moved from task-specific language models to general foundation models that can be fine-tuned for various applications. The company trained its model on millions of hours of video game data and demonstrated that a robot could perform tasks after being fine-tuned on just eight minutes of real-world data, suggesting that general models could reduce the need for collecting massive amounts of specialized robotics data. Rather than building robots itself, the startup aims to become a foundation model that other robotics companies can build upon.