Decart has launched Oasis 3, a generative model that creates photorealistic driving scenes in real time for testing autonomous vehicles, according to TechCrunch reporting published on 10 June 2026. The tool is available now via an API — a cloud service that developers can call up to generate synthetic driving data without building their own simulation software.

What Oasis 3 Actually Does

Oasis 3 generates hours of continuous, realistic-looking driving scenarios. The key problem it aims to solve: training self-driving cars is expensive and logistically difficult. Real-world testing requires collecting enormous amounts of video from actual roads in every condition and corner case you can imagine. Synthetic data — fake but realistic scenes created by AI — can fill in gaps and let teams test edge cases that are rare in the real world.

Synthetic simulation is not new. But what Oasis 3 offers is a combination that did not exist before: photorealistic rendering, real-time output, and availability as a simple API service. Where teams previously needed their own custom simulation platforms, they can now call an API and get back synthetic driving video. That is a significant simplification.

The TechCrunch report mentions important caveats — limitations and gaps in what the system can do. Those caveats are worth keeping in mind.

The Bigger Picture: Decart's Evolution

Decart started with a different product line. The original Oasis, as described on Decart's site, was a real-time AI system that rendered playable video game environments frame by frame, learning from thousands of hours of gameplay video. Instead of using a traditional game engine, it learned how pixels should change when a player moves — and generated the visuals on the fly using a neural network.

Oasis 2.0 extended this to video-to-video transformation: the AI watched live gameplay footage and altered it in real time based on player input.

Now Oasis 3 applies the same underlying idea to autonomous vehicles. Both gaming and AV simulation share a core challenge: maintaining a coherent world where objects stay in place and behave realistically frame after frame. But AV adds harder demands: precise geometry, sensor realism (simulating what a camera or LiDAR actually sees), and handling rare, dangerous scenarios.

The fact that Decart is now targeting autonomous driving suggests the underlying technology has matured beyond the flashy game demos and reached a level of consistency and accuracy that AV developers actually need.

Decart's other tools — Mirage for environment creation and Lucy for generating realistic human figures — have also been integrated with VR and AR headsets, per the company's documentation from October 2025. The connecting thread across all these products is the ability to maintain a believable, stable world while generating video in real time. That is a hard systems problem — getting the AI fast enough and coherent enough to produce frames at the rate your app needs — and it is less glamorous than the machine-learning algorithm itself, but it is what actually makes products work.

Why AV Simulation Matters Now

The autonomous vehicle industry has been fighting a problem for years: the gap between simulation and reality. Traditional simulation engines like CARLA or LGSVL let engineers test in a controlled, repeatable way. But the environments they generate do not look quite like real sensor footage. A self-driving car's perception software — trained on real-world camera images — can sometimes tell the difference between simulated and real data. When it encounters the real world, it has to guess. That gap is called distributional shift, and it is a major headache.

Generative world models trained on real driving video take a different approach. Because they learn from actual footage, their outputs look more like what a deployed car will actually see. That, in theory, cuts down on the mismatch problem.

The larger context here is worth pausing on. We have seen this pattern before, when new infrastructure tools emerge in a maturing industry. In the 2000s and 2010s, cloud platforms like AWS lowered the barrier to building and scaling software. Teams that could not afford their own data centers suddenly could compete. Now generative simulation could play a similar role in AV development: making expensive validation tools accessible to smaller teams and research labs.

Important Caveats

The TechCrunch report flags significant questions about Oasis 3 that have not yet been answered. Decart has not published detailed technical documentation, model specifications, or independent test results. That means it is not yet clear exactly where the system's limitations lie — how long it can generate coherent video before it starts glitching, how well it handles rare and dangerous driving scenarios, or what breaks when the system encounters something far outside its training data.

For AV engineers considering this tool, those specifics matter. If you are building a validation system that has safety implications, you need independent evidence that the tool does what it claims. Right now, that evidence does not exist in the public record.

API versus Owning Your Own Software

Decart is offering Oasis 3 as a cloud service, not as software you download and run yourself. That choice has real consequences. For smaller teams without the resources to run large AI models in-house, the API is a game-changer — you do not need expensive hardware or expertise in running machine-learning systems. You just make a call to Decart's servers and get back synthetic video.

But for large companies, especially those in regulated industries like automotive or defence, sending scenario data to a third-party service raises concerns. Data sovereignty — the question of where your data lives and who has access to it — becomes an issue. Compliance and security teams will want to review whether this is acceptable.

There is also the question of control. With Decart hosting the model, the company controls what scenarios can be generated and what gets patched or updated. Customers depend on Decart to keep the service running and evolving. We have seen this before, when large enterprises moved from owning their own Oracle or SAP systems to using cloud-based software. Many eventually built hybrid setups or worked with multiple vendors to avoid getting locked into a single provider. AV teams will likely face the same decisions.

What It Means If This Works

If Oasis 3 does what Decart claims — and if those claims hold up under scrutiny — the effect on AV development could be significant. Generating hours of high-fidelity simulation video is expensive right now. A tool that can produce photorealistic, physically plausible driving scenarios at scale would bring down that cost in ways that matter for smaller teams.

It would also shift how the autonomous vehicle industry works. Right now, the major AV programmes — Waymo, Cruise, and others with deep pockets — build their own simulation systems and keep them in-house. A third-party generative simulation service could loosen that grip and let more teams access the same tools. Over time, that could reshape the AV toolchain from a vertically integrated pile of custom software to a modular ecosystem where different teams build different layers.

The timeline here matters. Oasis 3 exists and the API is live, but maturity is still accruing. The real test will come when independent researchers or early customers publish their findings, or when Decart releases detailed technical documentation. Those are the data points that will show whether this technology actually moves the needle on one of autonomous driving's hardest engineering problems.