Oasis 3: Decart orients its world model toward robot training

Decart adapts its Oasis 3 world model to robotics, generating 3D environments at 22 fps to train autonomous vehicles and humanoids on NVIDIA hardware.

Decart Lab is evolving Oasis, its interactive world model born in gaming, into robotics.

Oasis 3 generates interactive and photorealistic environments from a simple text prompt, in which a robot operates in a closed loop: it perceives, acts, and the world reacts in real time to its actions. The targeted issue is the bottleneck of robotic training, as classic simulators require hundreds of hours of expertise per environment while producing video game visuals that transfer poorly to the real world, and struggling to cover rare events, such as a child crossing a snowy street.

Designed natively for robots, the model takes their actions as input rather than a keyboard and outputs three synchronized camera views, maintained spatially consistent by a geometry-aware architecture. Generation relies on an image-by-image autoregressive diffusion, conditioned on actions, which produces unlimited duration sessions by correcting drift over time. Optimized by the in-house DOS stack, touted as a hundred times more efficient, the system runs at 22 frames per second at 512×768 across three cameras, with a latency of less than 200 ms, on CoreWeave infrastructure and NVIDIA's physical AI ecosystem.

Decart claims it as the first API-accessible world model upon launch, with autonomous vehicles as its initial application area, before humanoids and industrial robotics. The company, founded in 2023, has raised over $450 million from Sequoia, Benchmark, Radical Ventures, and NVIDIA.