A cohesive, end-to-end development environment for physical AI on AWS brings together simulation, training, and governance in a single pipeline, built as an AWS CDK deployment that delivers a ready-to-use golden AMI with Isaac Sim, Isaac Lab, NVIDIA drivers, and DCV remote desktop.

Isaac Lab on AWS is designed as a production-oriented platform that dramatically reduces setup time, offering a streamlined, governed, and scalable development-to-deployment workflow for physical AI policies.

The package includes sample projects that demonstrate end-to-end workflows—training, data collection, and imitation learning—across three tasks, showing interactions with the desktop, simulations, and submission to SageMaker without code changes.

Experiment tracking and model governance are embedded: training runs feed MLflow, and completed policies are versioned and registered in the MLflow model registry with provenance data and promotion via staging and production aliases.

The solution supports two deployment topologies—from public GPU workstations (Instance Mode) to private, multi-user auto-scaling fleets (Fleet Mode)—all from a single codebase with configurable access and security.

Security, governance, and cost efficiency are central, featuring private networking, encryption, per-user storage, auto-scaling fleets, a one-time golden-AMI bake to avoid boot provisioning costs, and cost controls via per-job caps and allowlists.

A central SageMaker AI training plane handles managed training, enabling researchers to submit their project directory directly to training jobs without Docker rebuilds, with metrics streamed to MLflow for centralized tracking.

Prerequisites and guidance cover GPU quotas, tooling, pre-deployment validation, teardown procedures, cost considerations, and getting started through AWS accounts and partner programs such as the Physical AI Fellowship.