NVIDIA/physical-ai-data-factory

NVIDIA/physical-ai-data-factory

Description: Synthetic Data Generation Workflows for Physical AI

Language: Shell

License: NOASSERTION

Stars: 13

Forks: 2

Open issues: 0

Created: 2026-05-13T21:41:34Z

Pushed: 2026-06-08T17:24:51Z

Default branch: main

Fork: no

Archived: no

README:

Physical AI Data Factory

Agent skills and workflow documentation for generating labeled synthetic training data for physical-AI inspection and perception models — without waiting for every real defect or rare event to appear in production.

This repository does not ship application code. It ships agent skills: prompt-driven orchestrators that an AI agent (Claude Code, Codex, OpenClaw/Claw, or any compatible agent) follows to run end-to-end synthetic data generation pipelines on NVIDIA OSMO from a provisioned GPU runtime environment.

What's inside

| Skill | What it does | Use cases | |-------|--------------|-----------| | [Defect Image Generation (DIG)](skills/physical-ai-defect-image-generation/SKILL.md) | Synthesizes labeled defect and clean images for Automated Optical Inspection (AOI) by chaining USD rendering (USD-to-ROI), Qwen Image-Edit appearance transfer, and Cosmos AnomalyGen — with optional finetuning. Day 0 (cold start from CAD/USD) and Day 1 (inference + labeling on real photos) paths. | PCBA, metal surface, glass | | [Video Data Augmentation (VDA)](skills/physical-ai-video-data-augmentation/SKILL.md) | Runs video augmentation and auto-labeling (pseudo-labeling) on OSMO using Cosmos, covering flow selection, preflight, submit-time interpolation, monitoring, and output retrieval. | Surveillance / event video |

Each skill is self-contained under skills//, with its canonical OSMO workflow YAMLs in assets/configs/, use-case cookbooks in assets/cookbooks/, and long-form guidance in references/.

Repository layout

skills/ Agent skills (the product). Canonical source.
docs/ Per-workflow environment guides and setup walkthroughs.
.claude/ .codex/ .agents/ Agent-tool entry points (symlinked to skills/).

The .claude/skills, .codex/skills, and .agents/skills directories are symlinks to skills/, so the same canonical skill is discovered automatically by each agent runtime.

Getting started

The fastest path is to provision a GPU runtime environment, pull sample assets, and open an agent (Claw or your preferred coding agent) that drives the workflow for you:

• Defect Image Generation — see

[docs/workflows/physical-ai-defect-image-generation/launchable.md](docs/workflows/physical-ai-defect-image-generation/launchable.md)

• Video Data Augmentation — see

[docs/workflows/physical-ai-video-data-augmentation/launchable.md](docs/workflows/physical-ai-video-data-augmentation/launchable.md)

To use a skill directly with your own agent, point the agent at the relevant skills//SKILL.md and follow its prompts. The skill handles flow selection, preconditions, data handoffs, and OSMO submit/monitor commands.

Prerequisites

Workflows run on OSMO and pull gated assets, so you'll generally need:

• NGC credentials to pull container images from nvcr.io.
• A Hugging Face read token for gated Cosmos / AnomalyGen / Qwen-Image-Edit

models and datasets (accept each model's license first).

• OSMO CLI access with a logged-in profile and an available GPU pool.

The exact secrets, model/dataset links, and resource sizing are listed in each workflow's environment guide.

Contributing

This project is currently not accepting external contributions. The repository is published as-is for reference and reproducibility; issues and pull requests from community members will not be reviewed or merged at this time. If contributions reopen, all commits must be signed off per the Developer Certificate of Origin — see [CONTRIBUTING.md](CONTRIBUTING.md).

Security

To report a vulnerability, contact the NVIDIA PSIRT rather than opening a public issue. See [SECURITY.md](SECURITY.md).

License

Code is licensed under Apache-2.0 and documentation/skill content under CC-BY-4.0 (CC-BY-4.0 AND Apache-2.0). See [LICENSE](LICENSE).