Decart has launched Oasis 3, a new AI world model designed to generate long-running, realistic driving simulations for autonomous vehicle and robotics testing. The company says the system can create interactive road environments from text prompts and make them available through an API, moving the technology beyond a closed research demo and into developer hands.
The launch highlights a fast-growing race in AI: building models that do not just generate text, images or short videos, but simulate environments that machines can act inside. For autonomous vehicles, that could mean testing difficult road scenarios without waiting for them to appear in real-world driving data.
From Road Data to Generated Worlds
Oasis 3 is Decart’s newest model for what the company calls physical AI. It is aimed first at autonomous vehicles, but the broader target includes robotics, warehouses, manufacturing, drones and other systems that must operate in the real world.
The model works by generating driving environments from prompts. A user could ask for a snowy highway, a wet city street at dusk, a foggy rural road, a muddy camera view or a tunnel with poor lighting. Oasis 3 then creates a simulated road scene that can respond to actions over time, rather than producing a fixed video clip.
That is important because autonomous systems need more than ordinary road footage. They need repeated exposure to rare and dangerous situations. Decart’s own launch material describes examples such as “the child chasing a ball into a snowy street” and “the truck running a red light at dusk.” These are exactly the scenarios developers want to test, but they are hard to collect, schedule or recreate safely in the real world.
Decart argues that traditional simulation has not solved the problem fully. Classic simulators can model roads and traffic, but many still look artificial, require heavy manual design work and struggle to match the complexity of real-world physics. The company says Oasis 3 is meant to reduce those gaps by generating realistic visuals and interactive driving scenes at scale.
API Access Puts the Model in Developers’ Hands
One of the more important parts of the announcement is API access. Decart says Oasis 3 is available to developers from launch, which means customers can start building the model into simulation and testing workflows instead of only watching a product demonstration.
The company’s documentation lists Oasis 3 Preview at $0.02 per second for real-time world model and driving simulator use. That means a 60-second session would cost about $1.20 before any enterprise discounts or custom arrangements. The pricing signals Decart’s aim to make real-time simulation a usage-based infrastructure product, not only a research showcase.
The company says Oasis 3 runs at 22 frames per second at 512×768×3 resolution with less than 200 milliseconds of latency. It also supports synchronized three-camera driving views, so a simulated vehicle can receive observations closer to a real sensor setup than a single forward-facing video feed.
The system is powered by Decart’s Optimization Stack, known as DOS, which the company presents as the infrastructure layer behind its real-time models. Decart says DOS is built to reduce the cost and latency of running video, world model and agentic AI workloads.
A Fresh Funding Push Behind the Launch
Oasis 3 arrives shortly after Decart announced a $300 million funding round led by Radical Ventures. The company says it has now raised more than $450 million in total. Investors and partners include NVIDIA, Sequoia Capital, Benchmark, Zeev Ventures, Adobe Ventures, Toyota Ventures, eBay Ventures, Atreides Management and Valor Equity Partners.
The investor list shows why the launch is being watched beyond autonomous driving. Toyota Ventures connects Decart to mobility and robotics. NVIDIA’s involvement links the company to the physical AI and accelerated computing ecosystem. Adobe and eBay point to possible uses in media, commerce and interactive visual experiences.
Decart is not building only Oasis. Its other major model line, Lucy, is focused on real-time immersive video and world transformation for areas such as virtual try-on, advertising, livestreaming, gaming and social platforms. Oasis is the physical AI counterpart, built for robots and autonomous systems that need to train in interactive environments.
In a company statement tied to the funding announcement, Decart CEO and co-founder Dean Leitersdorf said, “World models are the key to moving AI from the virtual world into the physical world.” That line reflects the company’s larger thesis: language models helped AI reason in text, but robots and vehicles need models that understand structure, motion, space and consequences.
The Technology Still Faces Hard Limits
Despite the promise, Oasis 3 is not a complete answer to autonomous driving simulation. The biggest challenge is consistency over long sessions. AI-generated environments can look realistic at first, but they may drift over time. A road layout, location, building or background detail can slowly change in ways that make the scene less stable than the real world.
That matters because autonomous vehicle testing requires more than visual quality. A driving simulator must preserve the logic of the environment. If a vehicle passes an intersection, turns around and returns, the intersection should still be there. If another car appears in a lane, it should behave like a physical object. If the model generates a city, the streets and road rules should remain coherent over time.
Another limitation is physical interaction. World models can produce convincing motion, but collisions, vehicle behavior and unusual traffic events are difficult to model correctly. The issue is especially important for safety testing because the most valuable driving scenarios are often rare, dangerous or chaotic. A system trained mostly on normal driving footage may have less data for crashes, near misses and unusual human behavior.
This is a broader problem in autonomous vehicle development. The industry has long depended on real-world miles, synthetic data and hand-built simulation. Generative world models could add a new layer, but they still have to prove that they can produce scenes that are not only realistic to the eye, but reliable enough for safety-critical evaluation.
Competition Is Moving Quickly
Decart is entering a crowded and technically demanding field. Google DeepMind has been developing Genie, a general-purpose world model for interactive environments. World Labs, founded by AI researcher Fei-Fei Li, is working on spatial intelligence and models that understand and generate 3D worlds. NVIDIA has also been pushing physical AI and autonomous vehicle simulation through its Cosmos world foundation model work.
The attention around these efforts comes from the same industry shift. AI companies are moving from static generation to interactive generation. A chatbot can answer a question. A video model can generate a clip. A world model must keep producing a coherent environment while an agent acts inside it.
That is a much harder technical problem. It requires memory, physics, geometry, low latency, multi-camera consistency and the ability to react to actions without losing the scene. For autonomous vehicles and robots, those requirements are not cosmetic. They are central to whether the model can be trusted for training and evaluation.
A Step Toward AI-Generated Testing Infrastructure
Oasis 3 gives Decart a stronger position in the emerging market for AI-generated simulation. The model’s biggest advantage is not only that it can create realistic driving scenes, but that it can do so through an API and run interactively for extended sessions.
Still, the launch should be read as a major step, not a finished replacement for existing simulation systems. The current limits around drift, memory and physical consistency show why world models remain an active research problem. For developers, Oasis 3 may be most useful as an additional tool for scenario generation, early testing and long-tail exploration, rather than the sole foundation for safety validation.
The larger direction is clear. Autonomous vehicles and robots need to experience more of the world before they enter it. Decart is betting that much of that experience will be generated, not collected. Oasis 3 brings that vision closer, but the road from photorealistic simulation to dependable real-world training is still under construction.
Comments