Decart launched Oasis 3, a world model that generates realistic, controllable virtual training environments for robots and autonomous vehicles, aiming to solve the shortage of high-quality synthetic training data.

What happened

Decart, a frontier AI research lab, announced Oasis 3, a video output model designed to accelerate training of operating system models for robots and autonomous vehicles. The model combines photorealistic motion graphics with a physics engine, runs on CoreWeave's specialized cloud infrastructure at 22 frames per second, and generates interactive virtual environments at 512x768x3 resolution with less than 200 milliseconds latency. It is being made available through Decart's API.

Why it matters

Robotics developers face a critical bottleneck—they lack the billions of training examples that text-based AI developers can scrape from the internet. Existing solutions are costly (human teleoperation), messy (open-web videos lack spatial data), or unrealistic (current physics engines cannot capture real-world chaos like oil spills or fragile boxes). This gap between synthetic simulation and reality, called the 'sim-to-real gap,' leaves autonomous systems unable to handle unexpected situations.

What to watch

Oasis 3 allows developers to generate infinite variations of hazardous scenarios—a load falling off a truck, cameras smeared with mud, torrential rain—using simple natural language prompts from multiple camera angles. This represents an affordable way to expose models to millions of different hazards and prepare them for unpredictable real-world conditions.