Skip to content

The Physical AI Stack

How the pieces compose

The previous chapters covered individual concepts and technologies. This chapter shows how they compose into a complete system — from simulated warehouse to deployed robot fleet, managed by the same infrastructure patterns you use for conventional workloads.

The layered architecture

┌──────────────────────────────────────────────────────┐
│                  Enterprise / Hub                     │
│                                                      │
│  ┌──────────┐  ┌──────────┐  ┌───────────────────┐  │
│  │ Training  │  │ Model    │  │ Fleet Management  │  │
│  │ Pipeline  │  │ Registry │  │ & Orchestration   │  │
│  │ (KFP v2) │  │ (RHOAI)  │  │                   │  │
│  └────┬─────┘  └────┬─────┘  └────────┬──────────┘  │
│       │              │                 │              │
│  ┌────▼──────────────▼─────────────────▼──────────┐  │
│  │              GitOps (Argo CD)                    │  │
│  │              Policy (ACM)                        │  │
│  │              Observability (Prometheus/Thanos)   │  │
│  └──────────────────────┬─────────────────────────┘  │
│                         │                            │
│  ┌──────────────────────▼─────────────────────────┐  │
│  │            OpenShift (Kubernetes)               │  │
│  │            GPU Operator / NFD                    │  │
│  │            Security (Vault, Sigstore, Mesh)     │  │
│  └────────────────────────────────────────────────┘  │
└──────────────────────────────────────────────────────┘
                          │
              GitOps + Kafka Federation
                          │
┌──────────────────────────────────────────────────────┐
│                  Factory Edge / Spoke                  │
│                                                      │
│  ┌──────────┐  ┌──────────┐  ┌───────────────────┐  │
│  │ Inference │  │ Mission  │  │ Camera / Sensors  │  │
│  │ (KServe)  │  │ Dispatch │  │                   │  │
│  └────┬─────┘  └────┬─────┘  └────────┬──────────┘  │
│       │              │                 │              │
│  ┌────▼──────────────▼─────────────────▼──────────┐  │
│  │         Single Node OpenShift (SNO)             │  │
│  │         GPU Operator (L4 / L40S)                │  │
│  │         ACM Agent (klusterlet)                  │  │
│  └────────────────────────────────────────────────┘  │
└──────────────────────────────────────────────────────┘
                          │
              Local network / fieldbus
                          │
┌──────────────────────────────────────────────────────┐
│                  Robot / Field Devices                 │
│                                                      │
│  ┌──────────┐  ┌──────────┐  ┌───────────────────┐  │
│  │ VLA Brain│  │  Motors   │  │   Cameras         │  │
│  │ (Jetson) │  │ Actuators │  │   Sensors         │  │
│  └──────────┘  └──────────┘  └───────────────────┘  │
└──────────────────────────────────────────────────────┘

Layer 1: OpenShift + GPU infrastructure

The foundation. OpenShift provides container orchestration, the GPU Operator manages NVIDIA hardware, NFD provides hardware discovery, and the security stack (Vault, Sigstore, Service Mesh, NetworkPolicies) provides the trust layer.

This layer is the same whether you are running a web application, a database, or a robot brain. Your OpenShift knowledge transfers directly.

Layer 2: Platform services

GitOps (Argo CD) manages all declarative state. ACM governs multi- cluster policy. Observability (Prometheus, Thanos) provides metrics. These are the operational services that make the AI workloads manageable at fleet scale.

Layer 3: AI workloads

Training pipelines (KFP v2 + Training Operator) produce models. Model Registry tracks versions and lineage. Fleet Management coordinates robots. These are the AI-specific applications that run on the platform.

Layer 4: Edge

Single Node OpenShift at the factory edge runs inference (KServe + vLLM), mission dispatch, and local perception. It is ACM-managed and GitOps-synchronized with the hub.

Layer 5: Physical

The robots themselves — running VLA policies on Jetson hardware, connected to the edge cluster over the factory network.

NVIDIA components mapped to the stack

Stack layer NVIDIA component Role
Simulation Isaac Sim + Omniverse Digital twin, training environment, synthetic data
Simulation Nucleus Asset management for 3D scenes
Simulation PhysX 5 Physics engine
World models Cosmos Predict Scenario generation, pre-dispatch validation
World models Cosmos Transfer Sim-to-real domain adaptation
Perception Cosmos Reason Scene understanding, safety monitoring
Robot brain GR00T N1 VLA foundation model for humanoid control
Motor control SONIC Whole-body controller
Inference NIMs / TensorRT Optimized model serving
Edge hardware Jetson Orin / Thor On-robot compute
GPU infrastructure GPU Operator + GFD Hardware management

Red Hat components mapped to the stack

Stack layer Red Hat component Role
Platform OpenShift Container orchestration, security, networking
AI platform OpenShift AI (RHOAI) Model serving, training, pipelines, registry
Deployment Argo CD (OpenShift GitOps) GitOps state management, model promotion
Multi-cluster ACM Fleet-wide governance, policy enforcement
Edge SNO / MicroShift Factory-edge Kubernetes
Messaging AMQ Streams (Kafka) Fleet telemetry, events, cross-cluster federation
Service mesh OpenShift Service Mesh mTLS, traffic management, observability
Secrets Vault + VSO Credential management
Compliance Compliance Operator STIG scanning, remediation tracking
Supply chain Sigstore + Tekton Chains Image signing, provenance, SBOMs

The data flow

Isaac Sim renders warehouse scene
    │
    ├─ Replicator extracts RGB + depth + segmentation
    │
    ├─ Cosmos Transfer produces photorealistic variations
    │
    └─ Training pipeline (KFP v2) fine-tunes GR00T/VLA on synthetic data
         │
         └─ Model registered in Registry with full lineage
              │
              └─ Promotion PR opened in Git
                   │
                   └─ Reviewed, merged → Argo CD deploys to edge cluster
                        │
                        └─ KServe serves new policy on L4 GPU
                             │
                             └─ Robot executes missions with new policy
                                  │
                                  └─ Telemetry flows back to hub via Kafka
                                       │
                                       └─ Cosmos Reason monitors camera feeds
                                            │
                                            └─ Anomaly detected → Fleet Manager
                                                 replans → cycle continues

Key insight

The physical AI stack is not a separate universe from enterprise IT. It is enterprise IT extended with:

  • Simulation (Isaac Sim) as a first-class workload type alongside web apps and databases
  • Foundation models (GR00T, Cosmos) as a model category alongside NLP and recommendation models
  • Physical feedback loops (robot telemetry, camera feeds) as a data source alongside API logs and metrics
  • Edge clusters (SNO) as deployment targets alongside cloud regions

The platform skills transfer. The infrastructure patterns transfer. What is new is the physical domain knowledge — the concepts in Part 1 and Part 2 of this guide.

Key takeaways

  • The physical AI stack layers: infrastructure → platform services → AI workloads → edge → physical devices.
  • NVIDIA provides the simulation, models, and inference optimization. Red Hat provides the platform, deployment, governance, and security.
  • The data flow connects simulation through training, promotion, serving, monitoring, and back — a closed loop from virtual to physical.
  • Your existing Kubernetes/OpenShift skills are the foundation — physical AI extends them, it does not replace them.