README.md

sidebar_position	1
title	Getting Started
description	Entry point for deploying the Physical AI Toolchain
author	Microsoft Robotics-AI Team
ms.date	2026-09-19
ms.topic	overview
keywords	getting-started quickstart deployment onboarding

Deploy the Physical AI Toolchain and submit your first training job. This hub guides you through setup, deployment, and verification.

The default path starts on a laptop, not in the cloud. Begin with Start Here — T0 Dev, then graduate to higher tiers only when your scale demands them.

🪜 Choose Your Tier

Adoption is modeled as six graduated tiers (T0-T5). Each tier states the minimum infrastructure needed to complete the full training lifecycle: capture demonstrations on a robot, train an imitation policy, validate it, and run that policy back on the robot. Each tier is a legitimate stopping point. Start at T0 and graduate only when a concrete trigger forces it.

Tier	When to start here	Graduate when…	Quick start
T0 — Dev ⭐	Default. One laptop, one robot; zero cloud and no required Kubernetes.	No local GPU; the task needs many training iterations as conditions vary; or a second person needs the data.	Tier 0 — Dev
T1 — Lab	One site, a few robots, a shared GPU box; first cloud storage.	Training scale or team size outgrows one GPU box; dataset governance and catalogs become necessary.	Tier 1 — Lab
T2 — Pilot ✅	Recommended production. One site at scale; cloud training default.	The robot count or update cadence makes hand-updating each robot error-prone and version skew real, while everything is still at one reachable site.	Tier 2 — Pilot
T3 — Production	Advanced. Single-site declarative deploy (local k3s + Flux, no Arc).	Robots span multiple sites, or sites become unreachable from a single operator network.	Tier 3 — Production
T4 — Scale	Advanced. Multi-site fleet delivery; Arc reachability broker.	You explicitly want production signals to drive retraining and fleet-wide health analytics. This is a deliberate decision, not an automatic consequence of scale.	Tier 4 — Scale
T5 — Operate	Roadmap. Fleet intelligence for drift detection and retraining.	Available after the roadmap implementation lands.	Tier 5 — Operate

⭐ default · ✅ recommended production

For the tier-by-tier infrastructure boundaries see the Architecture Overview. Jump to T0 — Dev, T1 — Lab, T2 — Pilot, T3 — Production, T4 — Scale, or T5 — Operate. See the canonical Tier Model for the authoritative tier table and vocabulary.

Note

Roadmap honesty. T5 (Operate / fleet intelligence) is on the roadmap and not yet available. The fleet-intelligence domain is currently specified, with implementation planned. Today's shipping capability spans T0-T4.

Start Here — T0 Dev

The default starting path is one laptop and one robot, with zero cloud and no required Kubernetes. You close the full capture -> train -> validate -> run loop entirely on local hardware. Plain processes are the baseline; local Kubernetes is optional for workloads such as GPU offload.

1. Clone the repo and prepare a Python 3.12+ environment with uv. The repository-wide ./setup-dev.sh first requires Azure CLI, Terraform, kubectl, Helm, and jq; it is not a Python-only T0 bootstrap. See the Tier 0 recipe for the local workflow and check each component's prerequisites.
2. Record ROS 2 bags to local disk on the robot or laptop.
3. Copy data from robot to laptop with cp or rsync.
4. Run the dataviewer in local mode on the laptop.
5. Run lerobot-train on the laptop using CPU or a local GPU.
6. Keep training outputs on local disk; hosted experiment tracking enters at T2.
7. Use evaluation/sil/scripts/run-local-lerobot-eval.py for LeRobot evaluation. Isaac Lab evaluation uses evaluation/sil/play.py and requires the Isaac Lab runtime.
8. Launch the inference node as a plain process or container. No Flux, gating, or GitOps is required for T0.

Edge infra: ROS 2 and Docker only. Cloud infra: none. See the Tier 0 — Dev recipe for the step-by-step walkthrough.

Tip

Graduate when you have no local GPU, the task needs many training iterations as conditions vary, or a second person needs the data. At that point, step up to Tier 1 — Lab (first cloud storage) or jump straight to the recommended production path, Tier 2 — Pilot (cloud training). The Quickstart covers the cloud (T2 — Pilot) path end to end.

🚀 Guides

Guide	Description
Start Here — T0 Dev	Default local-first path: laptop + one robot, no cloud
Choose Your Tier	Pick a tier and see its graduation triggers
Quickstart	Cloud path (T2 — Pilot): clone to the first cloud training job
Architecture Overview (coming soon)	System topology, components, and data flow
Glossary (coming soon)	Term definitions for Azure, NVIDIA, and OSMO

⏱️ Time and Cost

The local default path (T0 — Dev) has no cloud cost. It runs entirely on your laptop. The estimates below apply to the cloud path (Quickstart, T2 — Pilot and up).

Item	Estimate
Total deployment time	~1.5-2 hours
Quick validation cost	~$25-50
GPU VM rate	~$3.06/hour (A100)

Note

Run terraform destroy when finished to stop incurring costs. See Cost Considerations for detailed estimates.

📋 Prerequisites Summary

T0 does not require Azure resources or Kubernetes services by default. Local components require Python ≥3.12, uv, ROS 2, and their own runtime dependencies; the dataviewer frontend also needs Node.js.

Install kind, kubectl, and Helm only when selecting an optional local Kubernetes profile such as GPU offload; the repository-wide setup-dev.sh additionally checks cloud and Kubernetes CLI tools even for a local-only workflow. The additional cloud tools below are required only for the cloud path (Quickstart, T2 — Pilot and up).

Tool	Version	Required for
Python	≥3.12	All tiers (incl. T0)
Terraform	≥1.9.8	Cloud path (T2+)
Azure CLI	≥2.65.0	Cloud path (T2+)
kubectl	≥1.31	AKS cloud path, Kubernetes tiers, and setup-dev.sh
Helm	≥4.2	AKS cloud path, Kubernetes tiers, and setup-dev.sh
jq	See prerequisites	setup-dev.sh and deployment scripts

For the cloud path, an Azure subscription with Contributor + User Access Administrator roles, GPU quota for the VM SKU selected in node_pools or managed compute, and an NVIDIA NGC account are required. The root Terraform GPU pool defaults to Standard_NV36ads_A10_v5; request quota for the family you configure, not a different GPU family. See Prerequisites for full details.

📚 Related Documentation

Resource	Description
Contributing Guide	Development workflow and code standards
Deployment Guide	Detailed deployment reference
Cost Considerations	Pricing breakdown and optimization