Decart Launches Oasis 3: A New Era for Autonomous Vehicle Simulation
The autonomous vehicle industry has long faced one of its most persistent bottlenecks: how do you safely and efficiently test a self-driving system without putting it on a real road? The answer, increasingly, lies in simulation — and Decart is pushing that answer further than almost anyone else. The company has officially launched Oasis 3, a real-time AI world model capable of generating photorealistic driving environments that can run for hours without interruption. Now available via API, Oasis 3 is positioned to become a foundational tool for developers and autonomous vehicle companies looking to accelerate their testing pipelines.
This launch marks a significant leap forward in what generative AI can do for the transportation and mobility sector. But as with any emerging technology of this scale, there are important caveats worth understanding before assuming Oasis 3 is a plug-and-play solution for every AV developer.
What Is a World Model — and Why Does It Matter for Autonomous Vehicles?
Before diving into the specifics of Oasis 3, it helps to understand what a "world model" actually is in the context of artificial intelligence. A world model is a type of generative AI system that doesn't just produce a static image or a short video clip. Instead, it continuously generates and updates a simulated environment in response to inputs — much like how a video game engine renders a world dynamically as a player moves through it, except the entire environment is generated by a neural network rather than pre-built assets.
For autonomous vehicles, this distinction is enormously consequential. Traditional simulation platforms rely on hand-crafted 3D environments, manually placed traffic agents, and rule-based scenario scripting. These approaches are expensive to build, slow to scale, and often fail to capture the true diversity and unpredictability of real-world driving. A world model like Oasis 3 can generate novel, photorealistic driving scenarios continuously and in real time, dramatically expanding the volume and variety of test cases available to AV engineers.
What Oasis 3 Brings to the Table
Decart's Oasis 3 builds on the company's earlier work in generative world modeling, but the scale and fidelity of this version represent a meaningful upgrade. The system is designed to produce photorealistic video of driving environments that is visually convincing at the level of detail that matters for perception system evaluation — things like lighting conditions, road surface textures, weather variation, and the behavior of surrounding vehicles and pedestrians.
Several capabilities stand out in Oasis 3's feature set:
- Real-time generation: Unlike offline simulation tools that render scenarios ahead of time, Oasis 3 generates environments on the fly, enabling interactive testing where AV systems can actually respond to the simulated world as it unfolds.
- Extended runtime: One of the headline achievements is the system's ability to sustain coherent, high-quality simulation over hours rather than seconds or minutes — a critical requirement for evaluating long-horizon driving behavior.
- Photorealistic output: The visual fidelity of Oasis 3 is aimed squarely at closing the "sim-to-real gap," the persistent challenge of making simulated environments realistic enough that a perception model trained or tested on them performs well in the physical world.
- API access: Oasis 3 is now available to developers via API, meaning teams can integrate it into existing workflows, build custom tooling on top of it, and scale usage based on their testing needs without managing the underlying infrastructure.
The Caveats: What Oasis 3 Cannot Yet Replace
For all the genuine excitement surrounding Oasis 3, the technology comes with limitations that responsible adopters need to weigh carefully. Photorealism is not the same as physical accuracy, and there is an important distinction between a simulation that looks real and one that behaves in ways that are reliably consistent with the physical world.
World models trained on video data learn to predict what a scene should look like next based on patterns in that data. This makes them extraordinarily powerful at visual generation, but it also means that rare or unusual physical events — a tire blowout, a sudden structural collapse, an object behaving in a genuinely novel way — may not be handled correctly because such events were underrepresented in the training distribution. AV developers testing safety-critical edge cases need to be aware that a photorealistic environment is not automatically a physically rigorous one.
Additionally, questions around scenario reproducibility, ground-truth data availability, and integration with existing simulation stacks will vary depending on a team's existing infrastructure. The API model offers flexibility, but it also requires engineering investment to use effectively at scale.
Why This Still Represents a Major Step Forward
Despite those caveats, the arrival of Oasis 3 as a commercially accessible product is a landmark moment for the AV industry. The ability to generate hours of photorealistic, interactive driving footage on demand dramatically lowers the cost and time associated with testing perception systems, training data augmentation, and scenario diversity evaluation.
For startups and research teams without access to large fleets of test vehicles or proprietary simulation platforms, Oasis 3 via API opens a door that was previously closed. For larger AV programs, it offers a complementary layer of simulation capacity that can run in parallel with physics-based tools, increasing overall testing throughput.
The Bigger Picture: Generative AI Reshaping Transportation
Decart's Oasis 3 is part of a broader shift in how the transportation industry is beginning to use generative AI — not just for content creation or customer-facing applications, but as infrastructure for development and validation. As world models become more capable, more controllable, and more physically grounded, their role in the AV development stack will only deepen.
For developers ready to explore what real-time world modeling can do for their autonomous vehicle programs, Oasis 3 and its API represent a compelling starting point — with clear eyes about both its current strengths and the work still ahead.