A startup called General Intuition believes it has found the shortcut to general-purpose robotics: millions of hours of video game footage paired with controller inputs.
The company, which raised $320 million at a $2.3 billion valuation in June, argues that the robotics industry is approaching a GPT-3-like inflection point where a single foundation model will make task-specific training for every robot, environment, and scenario obsolete.
The thesis
Current robot builders do “lots of specialized work focused on individual embodiments, environments, and robots,” General Intuition CEO Pim de Witte told TechCrunch. He argues that this approach is about to become redundant in the same way GPT-3 made specialist NLP models obsolete.
The insight is that video games already contain vast amounts of action data, not just visual information, but logs of what buttons were pressed, in what sequence, with what timing. This action data, General Intuition believes, is the key to building AI systems with human-like spatial-temporal intuition about the physical world.
Lead investor Vinod Khosla shares this view, backing the thesis that action data can teach AI models how the physical world works without requiring massive real-world data collection.
What they’ve demonstrated
General Intuition trained a foundation model on millions of hours of video game data. After fine-tuning on just eight minutes of real-world robotics data, the model powered a quadrupedal robot that performed zero-shot navigation using only its front-facing camera, no other sensors, no prior exposure to the test environment.
“The fact that the robot was actually able to zero-shot on just the front camera, with no other sensors, in the office with dynamic objects being introduced and people walking by was a very big surprise to us,” de Witte said. “I think it’s a sign of what’s to come.”
Business model
General Intuition does not build robots. It aims to become the foundation model provider for physical AI, the base layer that robotics companies fine-tune for their own machines, whether self-driving cars, warehouse robots, or humanoid platforms.
“We’re not going to build a self-driving car company,” de Witte said. “We’re going to make it 10 times easier for the next person to build a self-driving car company.”
The approach mirrors the shift that reshaped AI in 2022–2023: instead of thousands of narrow models each trained for a specific task, a single general-purpose model serves as the starting point for all downstream applications. General Intuition is betting the same pattern will repeat in the physical world.
Sources: TechCrunch (July 8)