Skip to main content
← Labs
Massive World Streaming & Persistent Universe

Project Atlas

Research Initiative

A next-generation streaming architecture for seamless, large-scale Roblox worlds through intelligent asset streaming, persistent world states and predictive loading.

The size ceiling

Roblox streams parts in and out around the player, which works well for worlds of moderate size and starts to show at scale. Large maps hit memory limits, streaming boundaries become visible as pop-in, and world state outside the streamed region effectively stops existing.

Project Atlas asks what it would take to build worlds an order of magnitude larger without those artefacts, and without requiring a high-end machine to play them.

Predictive streaming

Reactive streaming loads content once the player is close enough to need it, which guarantees a race between the loader and the player. Predictive streaming uses heading, speed and route to load what the player is about to need, converting a hard deadline into a soft one.

The research question is how much of that prediction can be done cheaply and how gracefully the system behaves when the prediction is wrong, since a confidently incorrect prefetch is worse than none.

Persistence beyond the session

A persistent world means changes outlive the player who made them and the server they were made on. That requires deciding what is worth persisting, how conflicting changes reconcile, and how a world stays coherent when thousands of people are editing it.

Regional simulation is the likely approach: the world is partitioned, each region maintains its own state, and only the regions with observers run at full fidelity while the rest advance on a coarser model.

More from Labs