synthetic-data
Generate synthetic training data using NVIDIA Cosmos world foundation models for SDG pipelines
- License 许可
- MIT
- 最近更新
- 2026-11-12
- Stars
- ★ 137
- Forks
- 71
01 前言 Front Matter
- name
- synthetic-data
- description
- Generate synthetic training data using NVIDIA Cosmos world foundation models for SDG pipelines
03 SKILL.md 正文
Synthetic Data Skill
Generate photorealistic training data using NVIDIA Cosmos world foundation models — Cosmos Transfer, Cosmos Predict, and Cosmos Reason.
Overview
The Synthetic Data domain provides SDG pipelines that transform simulation-rendered frames into photorealistic training data, predict future environment states, and curate output for training quality.
Pipeline Stages
| Stage | Model | Purpose |
|---|---|---|
| Transfer | Cosmos Transfer 2.5 | Convert Isaac Sim renders to photorealistic images |
| Predict | Cosmos Predict 2.5 | Generate future frame sequences from observations |
| Reason | Cosmos Reason 2 | Assess data quality and filter training samples |
Workflow Submission
SDG workflows can be submitted via OSMO or AzureML:
- OSMO workflows:
synthetic-data/workflows/osmo/ - AzureML jobs:
synthetic-data/workflows/azureml/
Key Files
| File | Purpose |
|---|---|
synthetic-data/README.md |
Domain overview and directory structure |
synthetic-data/workflows/osmo/sdg-pipeline.yaml |
End-to-end OSMO SDG pipeline |
synthetic-data/cosmos/configs/README.md |
Model configuration reference |
synthetic-data/specifications/synthetic-data.specification.md |
SDG pipeline specification |
synthetic-data/specifications/cosmos-integration.specification.md |
Cosmos model integration specification |
Environment Requirements
All NVIDIA Cosmos containers require:
| Variable | Value |
|---|---|
ACCEPT_EULA |
Y |
PRIVACY_CONSENT |
Y |
NVIDIA_DRIVER_CAPABILITIES |
all |
GPU Requirements
Each Cosmos model stage requires a minimum of 1x A100 (40 GB) GPU. H100 (80 GB) recommended for production workloads.
04 同仓库更多 Skills
osmo-lerobot-training
Submit, monitor, analyze, and evaluate LeRobot imitation learning training jobs on OSMO with Azure ML MLflow integration and inference evaluation - Brought to you by microsoft/physical-ai-toolchain
infrastructure
Deploy and manage Azure infrastructure for the Physical AI Toolchain including Terraform IaC, Kubernetes setup, GPU configuration, and network topology
fleet-intelligence
Monitor robot fleet telemetry via Azure IoT Operations, drift detection, Grafana dashboards, and Fabric analytics
fleet-deployment
Deploy trained robot policies to edge fleets via FluxCD GitOps, image automation, and deployment gating
environment-deployment
Generate, transfer, and consume environment-specific Azure, AKS, OSMO, ACR, and Azure ML deployment bundles. Use when: discovering deployed environment details; creating OSMO image manifests or platform values; preparing a HiL host; uploading or downloading deployment files through Azure Key Vault; or deploying with generated environment configuration.
e2e-tests
Autonomously launch, monitor, retry, and report all Physical AI cloud E2E tests across independent Azure ML and OSMO GPU infrastructure.