Atlas
Michael Sargis · 2025 · Research project
A small world model with a very unreasonable sense of ambition.
The idea
Give a model one camera frame and a proposed action. Ask it what happens next. Then ask it again, before a real robot has to find out the expensive way.
Atlas compresses a scene into a compact representation, predicts how that representation changes, and decodes possible futures. The planner compares those futures before choosing an action.
Results
The model has four billion parameters, runs on a laptop, and predicts 20 seconds of physical interaction. It can plan around a glass on a table. It cannot be trusted with your favorite glass.
The hard part
A plausible video is not proof that the physics is right. The evaluation needs to distinguish a convincing-looking spill from the actual path the water takes. I care more about the moment that breaks the prediction than the 19 seconds before it.