Decart’s New Oasis 3 World Model To Empower The Physical AI Revolution

Nvidia-backed artificial intelligence startup Decart says it has built the foundational infrastructure needed to support the coming robotics revolution, making it available to developers in the shape of Oasis 3, the company’s most advanced world model to date.

For the past few years, the promise of a revolution in general-purpose robotics has seemed tantalisingly close. We’ve been treated to all kinds of exciting hardware demos of robots cleaning people’s homes, strolling around warehouses and stacking items on shelves, and even robotic dogs designed to assist in search and rescue operations. But to see them actually doing these things in the real world? Not yet.

The problem is that the “brains” that power these next-generation robots haven’t progressed as far as the hardware has come along. However, with the recently announced Oasis 3, Decart is providing developers access to an entirely new class of generative model that has been engineered specifically to create physical AI training environments.

Decart’s leadership sees Oasis 3 as a key milestone in the maturation of world models into production-grade engines that will pave the way for major changes in the ways robots are trained and deployed.

More Than Just A Fancy Demo

Oasis 3 is a game-changer for world models, not only because of its advances in spatial computing and photorealism, but also because of the way it makes robotic training so accessible.

Over the last year, we’ve seen a flood of impressive innovations emerge from Big Tech powerhouses, such as Nvidia’s Cosmos 3, which boasts amazing photorealism thanks to its integration with the Omniverse platform. Meanwhile, Google DeepMind has developed Gemini Robotics, which is a family of world models designed to help robots perceive the environment they live in, reason how to complete multistep tasks and then perform them safely.

As enthralling as those innovations are, they haven’t yet enabled intelligent and autonomous robots to scale beyond lab environments, due to technical issues. For instance, latency remains a big problem that prevents humanoid bots and autonomous cars from operating safely at real-world speeds.

Physical mechanics are another issue. While robots might be able to formulate a 10-step plan to clean a kitchen, they’re likely to make a much bigger mess if they apply too much pressure when attempting to pick up an egg and place it into the fridge.

Because it enables closed-loop simulations through a live application programming interface, Decart’s Oasis 3 is not just outputting impressive videos, but realistic environments with ultra-accurate physics that can react instantly, in real time, to robots’ actions and decisions.

Imagine an autonomous drone flying through a park, adjusting its flight path to avoid a collision with a tree. When that happens, Oasis 3 can immediately alter the visual perspective across its synchronised three-camera view. Because it can do this with less than 200 milliseconds of latency, it gives developers the functional prompt-driven sandbox environment they need to teach autonomous systems how to react to almost any situation through active online reinforcement learning.

Decart combined its optimisation stack with CoreWeave’s AI infrastructure and Nvidia’s physical AI rendering layer to deliver some serious specifications. The model outputs realistic environments at a frame rate of 22 frames-per-second and at 512x768x3 resolution. It does this “infinitely,” streaming continuous simulated worlds that won’t cut off or degrade after a few seconds, ensuring the stability required for more sophisticated training runs.

Meanwhile, the tightly synchronised three-camera view ensures that autonomous systems have the spatial awareness needed to properly perceive depth and distances.

Expecting The Unexpected

Oasis 3’s first commercial application is focused on training autonomous vehicles. Developers will be able to access the model via its API and integrate a diverse range of driving environments within their production workflows.

They’ll be able to adjust the geographies, the surface of the road, the weather conditions and introduce all kinds of hazards into the mix. This is vital, because if self-driving cars are going to operate with full autonomy, they’re going to need to react to the unexpected.

Going forward, Decart has much bigger ambitions for Oasis 3. A spokesperson from the company says its infrastructure can be adapted to create simulated training environments for offroad vehicles, humanoid robots that need fine motor skills to manipulate objects and industrial drones that need to navigate obstacles like overhead power lines.

Here, Oasis 3’s domain randomisation capabilities are what sets it apart. Developers will be able to train their autonomous systems in different lighting conditions, throw unexpected clutter in their way, add different textures into the mix and introduce all kinds of rare edge cases.

For instance, a developer training a drone could prompt the system to whip up a gale force wind and see how it handles, then add in some flying debris to teach it the obstacle avoidance skills needed to survive such freak conditions.

If Oasis 3 lives up to its promise, then hardware is truly no longer the main hurdle in bringing robotics into the mainstream. Building physical AI devices has been possible for years, but the next frontier is intelligence. What’s been holding us back from building the kind of world envisioned by science fiction is the infrastructure needed to make that hardware smart enough to exist in the real world. With Oasis 3, world models could finally be ready to close that gap.