General Intuition raises $320M to train robots on gameplay data

The pitch behind General Intuition is simple enough to be interesting: robotics is bottlenecked by the cost of collecting real-world trajectories, so use the enormous existing corpus of human gameplay footage, with the controller button presses and timing already labeled, as the base training data instead. TechCrunch's Rebecca Bellan reports that the New York lab has raised $320 million at a $2.3 billion valuation, with Khosla Ventures leading and Jeff Bezos, Eric Schmidt and General Catalyst joining, alongside individual researchers from Google DeepMind and MIT. Khosla has now led both the seed and this round in a matter of months.

The demo the company is leaning on: after training on millions of hours of gameplay, they fine-tuned on just eight minutes of real-world robotics data and put the resulting model on a quadrupedal robot. CEO Pim de Witte says the robot ran zero-shot from nothing but a front camera, in an office with people walking around and objects being introduced. "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," he told TechCrunch.

Why this matters if you are not building humanoids yourself: for most of the last two years the assumed cost of admission to competitive robotics has been a fleet collecting real trajectories for years. If a general-purpose model trained mostly on cheap, already-labeled gameplay can bootstrap useful behavior from minutes rather than years of real data, the capital structure of the sector shifts. De Witte's own framing is that General Intuition wants to be the base model everyone else builds on, not the robot maker: "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."

The honest caveats are the ones the reporting does not quite address. A quadruped ambling around an office is a long way from dexterous manipulation, and eight minutes of fine-tuning on a locomotion task tells you little about how the model handles doorknobs, fabric or clutter. The article also does not get into the licensing story around the gameplay corpus, which traces back to de Witte's earlier gaming clip platform Medal; whose footage that is, and who gets paid, becomes a real question once this is an API. Take the specifics as reported, not settled.

If the thesis survives independent benchmarks, the winners are the second wave of robotics startups who no longer need a Tesla-scale data budget to start.