Physical AI Infrastructure Resilience
Set up, validate, scale, and recover Physical AI infrastructure for synthetic data generation workflows.
The skill specifies target confirmation, component scoping, secret-output avoidance, secretKeyRef use, and stopping on failed gates. However, Terraform and Azure/Kubernetes actions have substantial external effects; setup.sh can rewrite tfvars and run a full apply with auto-approve after IP drift, without explicit per-action confirmation, a general rollback procedure, or complete data-flow disclosure. Points are deducted accordingly.
The main flow, preflights, layered verification gates, and failure messages are reasonably clear, and the Azure preflight checks commands, versions, authentication, providers, and configuration. However, this is a static review with no executed key paths, and evidence for broad component coverage, abnormal-input handling, and recovery testing is thin, so the score remains below the static ceiling.
Audience, Azure/MicroK8s targets, OSMO, inference backends, workload choices, and positive/negative triggers are fairly explicit. Nevertheless, the description is very long, boundaries and environment compatibility are incomplete, Chinese-language interaction is not addressed, and reachability from mainland-China networks for Azure, NVIDIA NGC/NVCF, and upstream services is not discussed.
The component/stage structure supports progressive disclosure and includes version, Apache-2.0 licensing, author, tags, and maintenance-oriented rules. However, the recommended Instructions and Examples sections are absent; the benchmark reports no examples, an overlong description, and duplicate content. Changelog, maintenance ownership, and update path are also insufficiently explicit.
The flow covers cluster, OSMO, inference, workload, and verification stages, and could reduce blind retries through its gates. However, the supplied files contain no reproducible execution results; the benchmark verdict is FAIL and Tier 3 evidence is unavailable. Output completeness and benefit over manual operation therefore remain weakly verified.
Concrete scripts, configuration, verification gates, and a static benchmark report provide some traceable evidence. There is no independent execution in this review, no comprehensive committed test evidence covering the key paths, and no Tier 3 benchmark signal, so higher verifiability cannot be awarded.
- Do not run Terraform apply, Azure provider registration, Helm sync, or other cluster mutations without explicit user confirmation; in particular, setup.sh can enter an IP-drift branch that performs a full Terraform apply with auto-approve.
- Before production use, verify network reachability, credential scope, quotas, and cost for Azure, NVIDIA NGC/NVCF, Helm repositories, and upstream OSMO services; mainland-China availability is not demonstrated.
- Add and execute tests and rollback drills covering the selected components' key paths; the supplied benchmark verdict is FAIL and provides no Tier 3 execution results.
What does this skill do, and when should you use it?
This NVIDIA skill manages infrastructure for Physical AI synthetic data generation across local MicroK8s and Azure AKS. It composes Kubernetes, inference endpoints, OSMO, and workload stages into a reproducible environment. Supported inference paths include NIM Operator, NVCF, Azure AI Foundry, or no inference backend, with workload-specific readiness checks. It is a strong fit for teams that need gated deployment and recovery guidance, provided they have the required platform tools, credentials, quotas, and access.
Reads only the component references required for the selected target, along with workflow specifications and environment-derived values. Runs checked-in preflight and deployment scripts for Kubernetes, Helm, Terraform, Azure, inference services, OSMO, and workload submission when applicable. Verifies cluster reachability, node readiness, GPU capacity, inference endpoint health, OSMO pool and storage state, and final workflow completion. Treats Pending, Unknown, ImagePullBackOff, unbound PVCs, zero Ready replicas, and failed terminal workflows as layer failures requiring diagnostics before retry.
- A Physical AI engineer needs resilient VDA infrastructure on Azure AKS with NIM Operator and wants deployment gates for quota, GPU, inference, OSMO, and workflow readiness.
- A platform engineer needs a local MicroK8s environment with OSMO and NVCF for synthetic data generation.
- An infrastructure operator needs preflight checks for Azure access, caller CIDR, GPU quota, storage, API keys, and OSMO login before provisioning.
- A workflow operator needs structured recovery guidance after scheduling failures, image pull failures, storage failures, or failed OSMO workflow states.
What are this skill's strengths and limitations?
- Supports both local MicroK8s and Azure AKS deployment paths.
- Defines explicit stage gates and rejects invalid cluster/inference combinations.
- Includes credential-handling rules, preflight discipline, layered failure diagnosis, and recovery guidance.
- Covers NIM Operator, NVCF, Azure AI Foundry, OSMO, and several Physical AI workload routes.
- Requires kubectl, MicroK8s or Azure CLI/Terraform, and OSMO or inference credentials.
- The SKILL.md does not provide a complete standalone command-by-command deployment tutorial; execution depends on component references and checked-in scripts.
- Azure deployments depend on quota, identity, storage, network, and access conditions.
- Its scope excludes standalone OSMO log summarization and workload-only operations.
How do you install this skill?
Install the skill from the NVIDIA/skills collection with the skills CLI:
npx skills add nvidia/skills --skill physical-ai-infrastructure-setup-and-resilient-scaling --yes
To target Codex explicitly:
npx skills add nvidia/skills --skill physical-ai-infrastructure-setup-and-resilient-scaling --agent codex --yes
The source does not document a separate package or standalone manual-install procedure for this skill.
How do you use this skill?
After installation, give the agent a concrete target prompt such as: "Set up resilient Physical AI infrastructure for VDA on Azure AKS with NIM Operator." The skill should select one Kubernetes, OSMO, inference, and workload option, load only the needed references, run preflights, and enforce the Kubernetes, inference, OSMO, and workload gates. Do not use it for OSMO log summarization alone unless infrastructure setup, scaling, validation, or recovery is also requested.