Roblox turns text prompts into drivable cars with 4D generation

Most generative 3D tools handed to creators today do little more than spit out digital statues: they look fine, but they just sit there doing nothing. Roblox wants to fix that.

The platform has rolled out a beta for "4D generation," an AI capability powered by its proprietary Cube foundation model. The extra dimension here is interactivity. Instead of generating a static mesh, the model creates fully functional 3D objects that behave in real time.

Once a creator turns the feature on inside their experience, players can simply type a text prompt to spawn a working car, hop inside, and drive it down the road.

Breaking cars into schemas

The hardest part of generating a drivable car on the fly isn’t the exterior styling—it’s getting the moving parts to behave. Roblox approached this using what it calls schemas, which are modular rule sets that deconstruct an object into functional sub-components.

For the beta, Roblox rolled out two initial schemas: Body-1, which outputs single-mesh static items, and Car-5. The Car-5 schema forces the model to generate five separate mesh parts: a chassis and four distinct wheels.

Once the geometry is generated, Roblox’s system runs a retargeting pass. It analyzes the unique dimensions of whatever vehicle the AI dreamed up, attaches the driving scripts, and scales the physics so the wheels spin and steer naturally without clipping through the fenders.

Roblox says its long-term plan is an open-vocabulary schema system covering thousands of real-world objects, eventually scaling up to generating entire interactive scenes with code and animations.

Early players drove playtime up 64%

Community developers are already seeing what happens when players get their hands on this kind of sandbox power. Developer Laksh tested the tech in his game Wish Master, where users can type in whatever they want to see it materialize.

Players ended up conjuring everything from drivable sports cars to functional airplanes and flying dragons, generating over 160,000 4D objects during the early access phase.

The engagement numbers are striking: Laksh noted that players who actively engaged with 4D generation logged an average 64% increase in playtime.

It wasn't entirely smooth sailing, though. Laksh pointed out that players regularly asked for things that don't exist or used phrasing the team hadn't anticipated, forcing ongoing iterations to decipher player intent.

Why it matters

This is one of the smartest applications of generative AI in gaming we've seen so far.

While the rest of the industry has been racing to generate hyper-realistic 3D textures, Roblox recognized that visual fidelity means nothing in a game if the asset can’t move. By framing generation around modular schemas and dynamic scripting, they solved utility first.

For players, it shifts the Roblox loop from playing in pre-built sandboxes to improvisational creation inside the live game. Typing a sentence to spawn a personalized getaway vehicle makes user-generated content frictionless, which explains that massive 64% spike in session length.

Still, we think Roblox has a steep hill to climb before delivering on its grand vision. Right now, Car-5 is essentially a glorified dynamic template. Expanding from fixed schemas to thousands of arbitrary interactive items without breaking physics, glitching through maps, or ruining game balance will be a massive engineering challenge.