A New AI Tool Could Speed Up Self-Driving Car Testing

Decart has released a new AI system called Oasis 3 that creates realistic virtual driving scenes on demand. According to TechCrunch reporting from 10 June 2026, the system is available immediately through an online interface called an API — basically a way for companies to request scenes from Decart's computers rather than building their own. It is a tool for autonomous vehicle developers, not something ordinary people will use directly.

What Oasis 3 Does

Oasis 3 generates hour-long stretches of photorealistic driving scenes in real time. Think of it as an AI cinematographer that can paint a believable street, highway, or parking lot — with cars, pedestrians, weather, lighting, and traffic — and keep it consistent as the scene unfolds second after second.

The main problem it solves is expensive. Right now, self-driving car companies need enormous amounts of real-world driving footage to test their vehicles against. Recording that footage in the real world costs money, takes time, and does not easily cover all the unusual situations a car might face. An AI that can generate synthetic — computer-made — driving scenes cheaply and quickly addresses that bottleneck. Synthetic data for testing is not new, but combining real-time speed with photorealistic quality, all delivered through a simple online request, is a meaningful step. It turns something that usually requires a company to build and maintain expensive custom software into something closer to a rental service.

The TechCrunch report does mention some limitations alongside the capabilities — worth keeping in mind as we look at what this means.

How Oasis 3 Fits Into Decart's Broader Work

Oasis 3 is the third major version of a product line that started with something quite different. The original Oasis, described on Decart's website, lets AI generate a playable video game environment frame-by-frame in real time, responding to player keyboard input — essentially an AI game engine. Oasis 2.0 improved that by letting the AI transform game worlds as players interact with them, all in real time.

The shift from games to self-driving cars is important. Both need the same basic thing: a virtual world that remembers where objects are and makes them behave like they should. But self-driving car simulation adds harder requirements. The geometry — the exact shape and position of streets, curbs, and obstacles — has to be accurate. The rendering has to look like what a car's cameras and sensors actually see. And it needs to handle rare events — the unusual situations that are hardest to test for. Decart's choice to apply Oasis 3 to autonomous vehicles suggests the underlying AI has reached a level of quality and consistency that the flashier gaming experiments were always testing the limits of.

Decart also builds other AI systems — Mirage for creating environments and Lucy for generating realistic human figures — and has connected them to VR headsets, per the company's technical documentation from October 2025. What ties these products together is a real-time architecture that keeps the virtual world consistent at the speed required for interactive use — a tricky engineering problem that rarely gets attention in news coverage of AI, but matters a great deal.

Why Self-Driving Cars Need This, and Why Now

For the past decade, the autonomous vehicle industry has faced a stubborn problem: a gap between how well self-driving cars perform in simulation versus in the real world. Current simulation software — platforms with names like CARLA or LGSVL — works like a video game. It is precise and repeatable, but it does not quite look like real life. Self-driving car perception systems (the AI that interprets what the cameras see) have learned on real-world footage, so they can sometimes tell the difference between simulation and reality, which causes problems when the car is actually deployed.

Generative AI world models trained on real driving footage try to solve this differently. Because they learn what real-world driving looks like from actual video, what they generate should, in theory, be closer to what a deployed car will actually see. That is what Decart means when it says Oasis 3 is "photorealistic."

The broader context here is that this is a promising approach with real limitations. The TechCrunch report specifically notes caveats to Oasis 3's claimed capabilities. Decart has not yet published a detailed technical evaluation that independent experts can review. So we do not yet know exactly where the system breaks down — how long its consistency lasts before glitching, whether it handles rare dangerous situations well, or what it looks like when it fails. Self-driving car engineers will want those specifics before trusting it for safety-critical testing.

How It Works: The API Model

Decart is delivering Oasis 3 through an online interface, not as software you install on your own computers. This choice opens it to smaller autonomous vehicle teams and researchers who could not afford to run massive AI models themselves. But it also creates a tradeoff. If your company operates under strict data privacy rules — which many automotive and defence contractors do — sending your scenario data to Decart's servers outside your walls is risky and may violate regulations.

There is also a dependency question. Decart owns the AI model and controls what scenarios are possible. That gives the company flexibility to update and improve the system, but it also means customers are dependent on Decart keeping the service running and available. We have seen this tension play out before in software history: in the 2010s, large companies moved from running their own database software to renting it from cloud providers. Many eventually found they needed backup plans or deals with multiple vendors to protect themselves. Autonomous vehicle teams using Oasis 3 will likely face the same choice.

What Could Change

If Oasis 3 works as promised — and the early reports suggest it is still being proven out — it could reshape economics for autonomous vehicle development. Generating simulation hours of sufficient quality is expensive. An AI that can produce realistic, consistent driving scenarios quickly and cheaply could lower costs substantially for teams outside the handful of wealthy, established self-driving car companies that build their own simulation tools.

It could also shift power in the supply chain. Until now, most major autonomous vehicle programmes have built their own simulation infrastructure. A reliable third-party generative tool could break that pattern, letting companies focus on their own AI perception stacks while outsourcing the world-generation layer.

Oasis 3 is early in that story. The online interface is live, the capability claims are substantial, and the limitations are real. What matters next is independent technical review — whether from outside researchers, from companies that start using it and publish their findings, or from Decart publishing detailed technical documentation. That evidence will tell us whether this actually moves the needle on one of autonomous driving's toughest remaining engineering challenges.