1X World Model: a simulator learned from EVE robot data to evaluate policies
In one sentence 1X Technologies presents a world model trained on thousands of hours of video and action data from its EVE humanoid: given the current state and a sequence of actions, it generates video of what will happen next. It serves as a learned simulator for evaluating robot policies. A public challenge launches alongside, with over 100 hours of data released.
To know whether a robot can really perform a task you have to try it in the real world, and doing so across thousands of different situations is slow and expensive. One alternative is a "world model": a model that looks at the scene, knows which actions the robot is about to take, and imagines how the world will evolve.
1X Technologies trained such a model on data collected over a year by its EVE humanoid robots, performing manipulation tasks in homes and offices and interacting with people. The model learns directly from raw sensor data and generates video of what will happen. The stated goal is to use it as a learned simulator to evaluate robot policies across millions of scenarios; in the future it could also help train them.
1X also shows the limits: objects can change shape or color or vanish when occluded, physics is sometimes violated (a plate stays suspended in the air), and in front of a mirror the model does not "recognize" EVE.
Alongside the announcement the company launches the 1X World Model Challenge and releases over 100 hours of video under Apache 2.0 plus pretrained baseline models.
Companies
1X Technologies
Tools
—
Tags
Sources