Tolman Labs
Behavioral Data Infrastructure
In 1948, Edward Tolman showed that a rat running a maze builds a map of it. Block the usual route and the rat takes a shortcut it has never run. He called this a cognitive map, and the field spent the next twenty years arguing with him. Today the same idea goes by the name world model, and every lab working on physical AI is trying to train one.
A world model learns from records of behavior: what an agent saw, what it did, and what happened next. For physical AI that data barely exists. Robotics labs make it by hand, one teleoperated task at a time, and it goes stale as soon as the task changes.
Tolman Labs builds the infrastructure that produces it continuously. We capture gameplay from consenting players across 39 titles, with the frame, the input, and the game state kept in sync. We label it against a taxonomy that maps human behavior to physical-AI capability, and every clip ships with a provenance record. Our customers are the labs, the defense primes, and the sovereigns training physical AI. We supply them. We do not compete with them.