AWS publishes a sample physical-AI toolchain; the repo marks edge deploy as planned

Amazon described a Physical AI Toolchain on AWS twice in one week. The AWS Physical AI Blog is dated 7 October 2026. The Amazon news article, bylined Amazon staff, is timestamped 8 October 2026. Both point at sample code. The blog links aws-samples/sample-the-physical-ai-toolchain-on-aws. The GitHub API lists that repo as created on 9 July 2026 and pushed on 8 October 2026. The API licence field, the LICENSE file, and the README all say Apache-2.0.

The README calls it reference architectures and infrastructure-as-code for running NVIDIA’s physical-AI software on AWS. It does not release a new model. GR00T, Cosmos, and the other networks it wires in are NVIDIA’s.

The loop the article describes

Amazon’s article lists five stages. Synthetic data generation builds training scenes so a team collects less real footage. Model training learns from human demonstrations and from practice in simulation. Simulation and validation tests behaviour before any real hardware. Edge deployment puts an optimized model on the machine, so it can decide without a constant cloud link. Continuous improvement sends operational data back as new training data.

Sim-to-real is the handoff from the simulator to the robot. A policy can look finished in software, where mass, friction, and cameras are the ones the scene file chose. Sim-to-real is the further work of making that behaviour hold on a physical machine. The README’s name for the pillar is “Sim-to-Real / HIL” (hardware-in-the-loop): domain adaptation, safety monitoring, digital-twin sync, and deployment scoring.

The article names Amazon SageMaker for training, EC2 GPUs for simulation, AWS IoT Greengrass for the edge, and Amazon Bedrock AgentCore for orchestration. On the NVIDIA side it names Isaac Sim, Isaac Lab, Isaac GR00T, and Cosmos. The README adds NVIDIA OSMO on Amazon EKS, and Terraform as the path that deploys most components.

Available, preview, and planned

The README status table, on the 8 October push, splits those names:

Component	README status
Foundation, Cosmos, Isaac Lab, Isaac GR00T, DreamZero, Isaac Sim, OSMO	Available
Strands Agents	Available; hosted AgentCore runtime planned
Isaac Lab Arena evaluation	Preview
Edge deployment to Jetson via EKS Hybrid Nodes and Greengrass	Planned

Cosmos is available in this table. Edge deployment is planned, even though the news article lists it as a stage and names Greengrass. Strands Agents, the natural-language layer, is available, with the hosted AgentCore runtime still planned. DreamZero, which the README describes as a NVIDIA world-action model fine-tuned with LoRA on SageMaker, is available and is absent from the article’s NVIDIA list.

The article says the sample can help manufacturers “launch their own physical AI capabilities in weeks rather than the years it would take starting from scratch.” That timing is the article’s line. The README’s example costs (a GR00T smoke test around $2, OSMO around $5 an hour) are the sample’s own estimates, and it says terraform destroy tears the stack down.

Start from the table. An available row is Terraform you run in your account, with GPU quota and an NVIDIA NGC key, which the README lists as a prerequisite. The planned edge row is not a fleet deployment you can turn on from the repo today.