Decart launches Oasis 3, a real-time world model API for simulating photorealistic driving environments.
The release of Oasis 3 via API shifts AV simulation from rigid, heavily scripted game engines to generative, real-time world models. While long-term temporal consistency remains a known caveat with autoregressive generation, the ability to programmatically stress-test edge cases in photorealistic fidelity significantly lowers the barrier for AV validation.
Decart has officially launched Oasis 3, a real-time world model designed to generate photorealistic driving environments, now accessible to developers via API. Unlike traditional simulation engines that rely on explicitly modeled 3D assets and physics engines, Oasis 3 leverages generative AI to predict and render the next frame of a driving sequence in real time. This allows for the simulation of hours of continuous driving data based on programmatic inputs.
Technical Details Oasis 3 represents a shift toward neural simulation. By exposing this as an API, Decart is allowing autonomous vehicle (AV) developers to integrate generative world models directly into their testing pipelines. The model processes control inputs (like steering and acceleration) and outputs corresponding photorealistic video frames. However, the caveats associated with this approach are typical of autoregressive video generation: maintaining strict temporal consistency, physics adherence, and spatial memory over extended periods can degrade. While it can simulate hours of driving, edge-case hallucinations or slight physics drift are likely constraints engineers must account for when validating safety-critical systems.
Why It Matters For AV engineering teams, simulation is the bottleneck. Traditional platforms like CARLA require massive manual effort to build environments and script edge cases. Oasis 3 fundamentally alters this workflow, enabling teams to procedurally generate infinite variations of weather, lighting, and traffic conditions using a generative model. Even with minor physical inconsistencies, the sheer volume and photorealism of the generated data provide a high-value sandbox for training perception and planning stacks.
What to Watch Next The immediate metric for success will be how easily AV teams can integrate the Oasis 3 API into their existing CI/CD validation loops. Watch for updates addressing temporal hallucination mitigation and whether Decart introduces fine-tuning capabilities, allowing developers to condition the world model on their own proprietary fleet data.