This Startup Thinks Robotics Is About to Have Its ChatGPT Moment

General Intuition, a new startup, believes robotics is on the verge of its own 'ChatGPT moment.' Instead of training robots with expensive, time-consuming real-world data, the company is using millions of hours of video game footage to build foundation models for physical AI. These models learn common-sense physics and object interactions from virtual environments, which can then be adapted to control real robots with much less real-world data.

The core idea is that video games already simulate realistic physics — gravity, friction, object permanence — so AI trained on game data can develop a general understanding of how the physical world works. This approach could dramatically lower the cost and time required to train robots for tasks like navigation, grasping, and manipulation, potentially unlocking smarter home assistants, warehouse robots, and autonomous vehicles.

While the startup hasn't released a public demo yet, its research suggests that virtual training data could be the key to making general-purpose robots more practical. If successful, this method might accelerate the robotics industry much like large language models accelerated text-based AI.

For those interested in following the progress, General Intuition is expected to publish more details and potentially open-source components of their work in the coming months.