NVIDIA tutorial turns one image into training-ready physical AI datasets with Cosmos Data Factory
NVIDIA Developer · youtube · 2026-11-18
NVIDIA Developer published a hands-on tutorial showing how a coding agent drives Cosmos Data Factory to turn a single warehouse image into a training-ready dataset for physical AI.
Key points
- Pick from 5 reference blueprints; the Event Video Generation blueprint uses Cosmos 3 to generate a seed image, then synthesizes rare-event clips (a worker falling in a warehouse)
- Airflow + Kubernetes orchestrate generation and labeling; each clip gets detections, tracks, captions and Q&A annotations tied together by a manifest
- A built-in check verifies the requested event occurred; Cosmos Evaluator can add physical-plausibility scoring on larger machines
- Skills are composable without blueprints — e.g. Curation + Retrieval combined with Auto Labeling to find and annotate footage of workers near forklifts
Code is open-sourced on GitHub, useful for embodied-AI teams needing large-scale synthetic data pipelines.
More from Embodied
- Robotics Startup Genesis Raises $500M Series A at $3.5B Valuation — zhou_xian_ · 2026-11-18
- Genesis Details Path to Reliable Robots: Sim Parallel Practice Plus GENE Brain Model — zhou_xian_ · 2026-11-18
- Berkeley's Ken Goldberg: From 1994's crowdsourced Telegarden to agent-written robot code — lukas_m_ziegler · 2026-11-18
- GPT-6.2 models top AA-Robotics v0.1 with scores of 79 and 78 — ArtificialAnlys · 2026-11-18
- Telenoid: Hiroshi Ishiguro's eerie telepresence robot from Osaka University — ChuckDBrooks · 2026-11-18
- Tesla unveils Powercore transformer: 9 MVA, 98.86% peak efficiency for off-grid data centers, operating 2027 — elonmusk · 2026-11-18