gen‑ai.news
← Back
Video

Decart’s new world model can simulate hours of photorealistic driving — with some caveats

Decart has released Oasis 3, the latest iteration of its real-time world model, which is built to simulate photorealistic driving environments over extended periods. The system is aimed primarily at autonomous vehicle development, where the ability to generate diverse, high-fidelity road scenarios without physical test drives can significantly reduce both cost and risk. The model is now available through an API, allowing developers and research teams to integrate it into their own workflows and tooling.

World models like Oasis 3 work by learning the visual and physical dynamics of an environment from large datasets, then generating coherent, continuous video output that responds to inputs - in this case, the kinds of conditions and decisions a self-driving system might encounter. The promise is that a model trained on real driving footage can stand in for real-world testing across a wide range of edge cases, from unusual weather to complex intersections, without requiring a vehicle on the road.

The "some caveats" noted in the announcement are worth paying attention to. Sustained photorealistic coherence over hours of simulated driving is a hard problem - world models tend to accumulate errors or visual drift the longer they run, and there are open questions about how faithfully they reproduce the long-tail scenarios that matter most for safety validation. Whether Oasis 3 has meaningfully closed that gap, or whether the hours-long capability comes with constraints on scene complexity or interactivity, will likely depend on how developers stress-test it in practice.

Decart previously drew attention with an earlier version of Oasis focused on interactive game-world generation, making the pivot toward automotive simulation a notable shift in focus. Positioning Oasis 3 as an API product suggests the company is targeting enterprise and research users rather than consumers, and it places them in a competitive space alongside other groups working on generative simulation for robotics and autonomous systems. How the model holds up under rigorous AV development requirements - where the bar for reliability is high - remains to be seen.

Enjoy this story? Get the next one in your inbox.

Twice a week: the most important stories in generative image and video AI, distilled into a 2-minute read.

Free. Unsubscribe any time. No spam, ever.

Your next read

No image
Video

Snap spins off AI video team into new company, Dotmo, due to costs

Snap is spinning off its internal AI video team into a new independent company called Dotmo, with the move driven primarily by the high costs of developing generative video technology in-house. The staff involved are departing Snap to focus solely on AI video work under the new entity. It marks another instance of Snap shedding an internal unit rather than continuing to absorb the expense of frontier AI development.

Video

Amazon, Nvidia, and AMD bet $310 million on AI startup building 3D world models

Odyssey ML has raised $310 million from Amazon, Nvidia, and AMD, pushing its valuation to $1.45 billion. The startup is focused on building 3D world models - AI systems that can understand and generate structured representations of physical space. The round also draws in notable backers including Google chief scientist Jeff Dean and CIA-linked venture fund IQT.

No image
Video

Meet Qwen-RobotSuite: Three Embodied AI Models for VLA Manipulation, Video World Modeling, and Navigation

The Qwen team has released Qwen-RobotSuite, a collection of three specialized models targeting different challenges in embodied AI: physical manipulation, world modeling, and navigation. Each model draws on existing Qwen language and vision foundations while introducing architecture and training choices tuned for robotics tasks. The release comes with benchmark results and details on the data pipelines used to train each system.