A Guth Labs publication ↗
Back to News

NVIDIA details Cosmos dataset workflows and expanded sensor support

Miniature cream car on a curved green glass road between two blank glass panels on a round cream platform
AI-generated editorial illustration

NVIDIA detailed an expanded Cosmos data-preparation workflow on November 18, combining five reference blueprints and six reusable agent skills for physical AI training datasets.

Its technical post describes workflows for driving scenarios, image and video augmentation, rare-event videos and manufactured-defect images. Coding agents can coordinate services on a cluster, while Cosmos Evaluator checks generated material for requested attributes and physical plausibility.

The update says Cosmos Curator now accepts multicamera video, inertial sensor and GPS inputs, with initial lidar support. Further lidar devices and radar remain planned additions.

Cosmos Dataset Search is now open source under Apache 2.0, according to NVIDIA. It combines scene metadata, spatial relationships and temporal behavior to find useful video segments, with adaptive agent-based reranking in early access.

NVIDIA also reports throughput and cost measurements, but says its estimates cover compute only, excluding costs such as setup, storage and human review. The practical goal is to make rare or difficult-to-capture training examples easier to produce, label and trace back to their source.

Sources and citations

Revision history

Published .