Recipes

Step-by-step guides that take you from a standing start to a working result. Each recipe is self-contained with prerequisites, runnable commands, and verification steps.

Recipes are organized two ways: by tier (how much infrastructure you run) and by topic (what task you are doing). New users should start with the tier ladder below; the default path is T0 — Dev, which runs the full training lifecycle loop on one laptop with zero cloud and no required Kubernetes.

NOTE

Only the cloud tiers (T2+) and the topic recipes under Training and Data Collection assume deployed Azure infrastructure. T0 — Dev and the storage-only T1 — Lab do not. If a recipe needs cloud resources, complete the Quickstart first. For the canonical tier definitions, see the tier model.

🪜 Pick a Tier

Each tier states the minimum infrastructure it assumes. Start at the default (T0) and graduate only when a real constraint forces it.

T#	Recipe	Minimum infrastructure	Status
T0	T0 — Dev	One laptop + one robot. ROS 2 + Docker; optional local Kubernetes. No cloud.	Default
T1	T1 — Lab	T0 + one Azure Blob storage account.	Authored
T2	T2 — Pilot	AzureML + storage + registry + MLflow. No k8s.	Recommended
T3	T3 — Production	T2 + single-site local k3s + FluxCD. No Arc.	Advanced (stub)
T4	T4 — Scale	Multi-site Arc + AKS/Flux + gating.	Advanced (stub)
T5	T5 — Operate	T4 + IoT Operations + Fabric RTI (roadmap).	Roadmap (stub)

🗂️ Topic Recipes by Tier

The existing topic recipes assume the tier shown below. They are unchanged by the tier ladder. This table only classifies them.

Topic recipe	Assumes tier	Minimum infrastructure
Configuring Edge Data Recording	T0 — Dev	Jetson / robot, ROS 2. No cloud.
Preparing Datasets for Training	T0–T1	Local for T0; Azure CLI + Blob for cloud datasets.
Your First LeRobot Training Job	T2 — Pilot	Deployed infrastructure, OSMO running.
Your First RL Training Job	T2 — Pilot	Deployed infrastructure, OSMO running.
End-to-End LeRobot Pipeline	T2 — Pilot	Deployed infrastructure, OSMO running.

🚀 Pick a Recipe

Goal	Recipe	Time
Train an RL policy	Your First RL Training Job	30 min
Train a LeRobot policy	Your First LeRobot Training Job	30 min
Run the full train → eval → register pipeline	End-to-End LeRobot Pipeline	60 min
Configure edge recording	Configuring Edge Data Recording	20 min
Prepare a dataset for training	Preparing Datasets for Training	30 min

📖 Recipe Catalog

Training

Recipe	Description	Prerequisites
Your First RL Training Job	Submit an Isaac Lab RL training job on OSMO with SKRL	Deployed infrastructure, OSMO running
Your First LeRobot Training Job	Submit a LeRobot behavioral cloning job on OSMO	Deployed infrastructure, HuggingFace dataset
End-to-End LeRobot Pipeline	Orchestrate train → evaluate → register in one command	Completed basic LeRobot recipe

Data Collection

Recipe	Description	Prerequisites
Configuring Edge Data Recording	Set up ROS 2 edge recording on Jetson with chunking and compression	Jetson device, ROS 2
Preparing Datasets for Training	Download, inspect, and validate datasets for LeRobot training	Python 3.12+, Azure CLI

🔗 Related Documentation

• Tier model (canonical reference): tier IDs, boundaries, and vocabulary
• Getting Started: infrastructure deployment and first training job
• Training Guide: reference documentation for RL and IL workflows
• Data Pipeline: edge recording configuration reference
• Scripts Reference: CLI parameter tables for all submission scripts

🤖 Crafted with precision by ✨Copilot following brilliant human instruction, then carefully refined by our team of discerning human reviewers.