電子遊戲訓練模型

General Intuition introduced a foundation model for embodied AI trained on millions of hours of video game data, enabling spatial-temporal reasoning that can transfer from virtual environments to physical robots. The startup developed the model to learn movement and action patterns from controller inputs, positioning it as a general-purpose platform for robotics rather than a system tailored to a single machine or environment.

The company demonstrated that the model could play video games for extended periods and, after being fine-tuned with just eight minutes of real-world robotics data, power a quadrupedal robot using only a front-facing camera. Backed by a recent US$320 million funding round, General Intuition aims to provide a foundation model that robotics companies can adapt for their own applications, reducing reliance on extensive real-world training datasets.

For robotics developers, the platform could significantly shorten development cycles by making general-purpose physical AI easier to deploy across different machines. The launch reflects a broader shift toward foundation models that prioritize adaptability and transfer learning over building specialized AI systems for individual robotic applications.