What does this skill do, and when should you use it?
This NVIDIA skill orchestrates physical-AI video augmentation and auto-labeling workflows on OSMO. It handles flow selection, credential preflight, cache readiness, inference-path decisions, submit-time interpolation, monitoring, and output retrieval. It supports auto-labeling-only, augmentation-plus-labeling, and two end-to-end flows. Adoption requires the OSMO CLI, an available GPU pool, relevant credentials, and a correctly identified dataset backend.
Selects auto_labeling, augmentation_and_al, e2e, or e2e_super_resolution from user intent; runs credential, control-plane, readiness, and pre-submit checks; prepares model caches when needed; derives storage_url from the active dataset backend; submits and monitors OSMO workflows; and downloads complete run artifacts. For augmented flows, it also renders side-by-side evidence and summarizes augmentation and pseudo-labeling outputs.
- A data team labeling only source videos can use the auto_labeling flow.
- A team creating augmented videos and labeling those outputs can use augmentation_and_al.
- A team prioritizing throughput for the full pipeline can use the parallel e2e flow.
- A team requiring a super-resolution gate before augmentation can use e2e_super_resolution.
- A user without an input dataset can validate the setup with NVIDIA's default VDA demo assets.
How do you install this skill?
- Execution automatically submits cloud workflows, deploys or repairs NIMs, downloads models, and incurs GPU, storage, and network costs; authorization, budget, and rollback should be confirmed in the target environment.
- Credentials are passed into NVIDIA containers and VLM/LLM/Hugging Face-related workflows; the external Google Drive, Google Storage, and Hugging Face sources require separate supply-chain, network-reachability, and integrity review.
- The benchmark contains only one task and does not establish multi-video behavior, retry robustness, permission isolation, or actual output quality.
- Shell / CLI
- Network access
- Local filesystem
OSMO CLIBashPython 3Gitjq
Install the specific skill from the NVIDIA catalog with:
npx skills add nvidia/skills --skill physical-ai-video-data-augmentation --yes
To target a specific agent, for example Codex:
npx skills add nvidia/skills --skill physical-ai-video-data-augmentation --agent codex
The README does not document additional single-skill installation steps.
tmp="$(mktemp -d)"
git clone --depth 1 https://github.com/NVIDIA/skills.git "$tmp"
mkdir -p ~/.claude/skills
cp -R "$tmp/skills/physical-ai-video-data-augmentation" ~/.claude/skills/
rm -rf "$tmp"Generated from the source repository and skill path; it copies only this skill's folder. If the author's install steps above differ, follow those first. To scope it to one project, replace ~/.claude/skills with that project's .claude/skills.
How do you use this skill?
Once installed, send your agent any of these to trigger it:
- Run video data augmentation on OSMO and auto-label the augmented videos.
Give the agent a concrete request such as: “Run video data augmentation on OSMO and auto-label the augmented videos.” If no flow is specified, it defaults to augmentation_and_al; if no video is supplied, it uses nvidia/video-data-augmentation-demo. Before execution, the environment needs HF_TOKEN, a logged-in OSMO CLI with a matching data credential profile, and at least one ONLINE GPU pool. The skill pauses for missing credentials, an ambiguous storage-backend prefix, or an unavailable GPU platform.
What are this skill's strengths and limitations?
- Covers the workflow from preflight through output download.
- Provides four explicit workflow modes and automatic model-cache remediation.
- Derives storage configuration from the active dataset backend instead of assuming a stale scheme.
- Requires side-by-side comparison evidence for augmented flows, not merely raw output URLs.
- The repository README describes daily synchronization plus signatures, evaluation datasets, and benchmark-report governance for published skills.
- Requires OSMO CLI access, a logged-in profile, matching data credentials, and an ONLINE GPU pool.
- Usually requires a Hugging Face token; some optional NGC operations may also require an API key.
- Execution depends on network access, GPU scheduling, and model caches; cold-start runs can take about 45–80 minutes.
- The supplied SKILL.md provides no standalone test suite or concrete platform-coverage evidence.
- It is an execution orchestrator and is explicitly not intended for container-internal tuning-only questions.
How does this skill compare with similar options?
Side by side with related skills; every score comes from the same FSRS standard.
| Skill | FS score | Stars | Last updated | License |
|---|---|---|---|---|
| OSMO Physical AI Video Augmentation Orchestrator this page ✓ NVIDIA · Official | 40 · Not recommended | ★ 3.8k | today | Apache-2.0 |
| Physical AI Infrastructure Resilience ✓ NVIDIA · Official | 47 · Use with care | ★ 3.8k | today | Apache-2.0 |
| n8n Binary & File Handling | 53 · Use with care | ★ 6.5k | 4d ago | MIT |
| VSS Video Embedding Deployment ✓ NVIDIA · Official | 57 · Use with care | ★ 3.8k | today | Apache-2.0 |
| VSS Archive Video Search ✓ NVIDIA · Official | 51 · Use with care | ★ 3.8k | today | Apache-2.0 |
How did FollowSkills review this skill?
The documentation identifies required secrets, data paths, credential purposes, input-preservation rules, and a prohibition on scaling down or deleting existing NIMs; it also requires deriving storage_url from the active backend. However, it automatically submits workflows, deploys or repairs NIMs, installs dependencies, and accesses external model sources without per-action confirmation, least-privilege detail, rollback, or complete data-flow disclosure, so points are deducted.
The workflow, parameters, preflight, cache checks, monitoring, and some failure feedback are relatively structured, and commands use set -euo pipefail. Static review cannot reproduce key paths because the referenced scripts are not supplied; execution also depends on online images, runtime apt/pip installation, and a configured OSMO environment, with limited abnormal-input coverage, so the score remains below the static ceiling.
Triggers, use cases, four flows, defaults, and several blocking conditions are clearly described. However, the core capability depends on OSMO, GPUs, and NVIDIA/Hugging Face/Google download sources; Chinese-language support and mainland-China network reachability are not addressed, and non-OSMO environments are largely out of scope, so points are deducted.
SKILL.md includes purpose, prerequisites, a script inventory, flow tables, examples, limitations, version/owner metadata, and explicit licensing. The supplied benchmark nevertheless reports a missing recommended Examples section and an invalid author format; maintenance responsibility and a changelog/update path are also incomplete in the skill materials, so points are deducted.
The stated goal is end-to-end VDA orchestration from preflight through submission, monitoring, retrieval, and side-by-side evidence, with relatively direct output paths and completion artifacts. The BENCHMARK reports strong results for one task, but coverage is only a single positive task with no negative tasks, and static review cannot verify actual artifact quality or cost-benefit, so points are deducted.
Pinned model revision hashes, explicit configurations, and one NVSkills-Eval report provide limited audit evidence. The report covers only one task and the supplied files lack a committed test suite or CI evidence covering key paths; independent reproduction is therefore unavailable from this static material, so points are deducted.
Open a dimension to read why it scored that way
Evidence confidence：Low — Mostly static review, author material or a limited demo; useful for discovery, not high-risk decisions.
See the full review method →