Decart releases Oasis 3: a photorealistic driving simulator that runs for hours via API

Decart, a startup building generative world models, has opened API access to Oasis 3, a system that generates photorealistic driving scenes in real time for testing self-driving cars.

What Oasis 3 can do

Oasis 3 is a so-called world model: a neural network that doesn't just draw static images but simulates a dynamic physical environment. In practice, the model generates continuous video streams of driving through a city, on a highway or in a parking lot, complete with other vehicles, pedestrians, traffic lights and changing lighting conditions. The model runs in real time and can synthesize hours of driving without stopping.

Developers get access through the API and can:

• Create custom routes and driving scenarios
• Test autopilot behavior in rare and dangerous situations
• Generate data for fine-tuning perception networks
• Check object-detection systems without going out on a real road
• Scale testing without expanding a vehicle fleet

This can significantly cut development cost and time: real-world road tests remain expensive, slow and legally complicated in practically any jurisdiction.

Why world models matter for self-driving cars

Traditionally, companies developing autopilots choose between two extremes: real miles, which are expensive, slow and require a large fleet, or classic graphics simulators such as CARLA. Simulators are fast, but their imagery is far from reality. Between the two lies a gap that engineers call the sim-to-real gap: a neural network trained on unrealistic data often gets lost in the real world, failing on details the simulator never reproduced.

Generative world models are an attempt to close that gap. Photorealistic imagery, with reflections on wet asphalt, rain on the windshield and sun glare in the camera lens, gives an autopilot data as close to reality as possible.

"Our goal is to give self-driving developers a tool to scale testing without scaling the vehicle fleet,"

Decart said.

Caveats and limitations

Decart does not hide the fact that Oasis 3 has weak spots. Despite its impressive realism, the system has known limitations:

• The model may produce visual artifacts during abrupt lighting changes or in extreme weather scenarios
• The accuracy of other road users' behavior still lags behind specialized physics engines
• Generated scenes do not guarantee full fidelity for rare edge cases
• For highly safety-critical testing, Oasis 3 data will require additional verification

Still, API availability means that small teams and startups get technology that used to be the preserve of large labs such as Waymo or Toyota Research Institute.

What this means

Oasis 3 opens a new stage in the democratization of autonomous-system development: photorealistic synthetic data becomes available through a simple API call. If world models mature to a level of accuracy sufficient for production, they could radically change the economics of automotive AI: part of the costly real-world testing will move to the cloud, and the barrier to entry for new players will drop noticeably.