General Intuition leverages video game data to train smarter robotics AI

General Intuition is training foundation models for physical AI using millions of hours of video game data, aiming to reduce the need for extensive real-world robotics data. The startup recently raised $320 million.

AI & Culture Desk
Updated Nov 13, 2026 2 min read
General Intuition leverages video game data to train smarter robotics AI

General Intuition is developing foundation models for physical AI, leveraging millions of hours of video game data. The company's CEO, Pim de Witte, argues this approach can create general models capable of transferring intuition about movement and interaction across various environments, contrasting with older methods that built specialized models from scratch. This strategy aims to overcome the significant cost and time barriers associated with collecting large, real-world robotics datasets.

The startup built its foundation model by training on extensive video game data, which includes information on human controller inputs and timing. According to de Witte and lead investor Vinod Khosla, this action data is crucial for developing human-like intuition regarding spatial-temporal reasoning. General Intuition believes this approach will eventually reduce the necessity for collecting hundreds of thousands or millions of hours of real-world data, suggesting only a few minutes might suffice for specific tasks.

General Intuition's core product is the generalization capability of its model itself. De Witte stated that the model's base level of reasoning about space and time is what will enable robotics companies to move away from collecting vast amounts of real-world data. The company's thesis is that the emergence of such general models will make much of the current specialized work in robotics redundant.

The company recently secured $320 million in funding at a $2.3 billion valuation, based on this strategy. General Intuition has demonstrated its model's ability to play video games for extended periods and, after fine-tuning on just eight minutes of real-world robotics data, to power a quadrupedal robot. This fine-tuning was performed using only the robot's front camera, without additional sensors.

De Witte described the robot's performance in an office environment with dynamic objects and people as a significant surprise, calling it "a sign of what’s to come." The company's long-term objective is not to manufacture robots but to provide the foundational AI models that other robotics firms can use to build their machines. De Witte explained this by stating, "We’re not gonna build a self-driving car company. We’re gonna make it 10 times easier for the next person to build a self-driving car company."

This approach mirrors the evolution of natural language processing, where general-purpose models like OpenAI's GPT series, Claude, or Llama became the starting point for many applications, replacing the need for organizations to build specialized models from the ground up. General Intuition's work suggests a similar paradigm shift is possible for embodied AI, where simulated or game-based data serves as a proxy for physical experience.

Sources

  1. 01 This startup thinks robotics is about to have its ChatGPT moment — TechCrunch
#robotics #foundation-models #video-game-data #ai-training #general-intuition