AWS publishes an open-source Physical AI Toolchain with NVIDIA stack

What did Amazon announce?

The About Amazon post, bylined Amazon Staff, says physical AI is moving onto factory floors, warehouses and roads, and that customers were spending too much engineering time on infrastructure. Uwem Ukpong, vice president, AWS Industries, is quoted: “We built the Physical AI Toolchain on AWS because customers told us that too much of their engineering effort was going to infrastructure instead of innovation.” The post says the stack is “integrated with the NVIDIA Physical AI stack” and “covers the complete physical AI development lifecycle.”

Customer examples in the post are vendor lines: NEURA Robotics on cognitive humanoids, RLWRLD on an 8.1-billion-parameter manipulation model, and Config on more than 200,000 hours of robot action data. David Reger of NEURA and Amit Goel of NVIDIA are quoted. Amazon also cites a Global Startup Trends Report claiming one in seven startups is building physical AI and that 72% of builders say cloud computing is essential. We did not open that report for this story.

A separate AWS Documentation page, “Guidance for Physical AI for Robotics on AWS,” describes a seven-step loop from Isaac Sim and Isaac Lab on GPU EC2 through Greengrass, IoT Core and SageMaker retraining. That guidance is architecture copy. It is not the same artifact as the Terraform samples. For the difference between an agent and a scripted control loop, see what an AI agent is.

What is actually in the GitHub repository?

The README calls the project “a curated collection of reference architectures, Infrastructure as Code, and deployment automation.” It describes three compute classes: high-bandwidth GPU clusters for foundation-model training, elastic mid-tier GPUs for simulation, and edge GPUs (NVIDIA Jetson Thor / AGX) for real-time inference. The flywheel table has four pillars — synthetic data, model training, software-in-the-loop simulation, and sim-to-real / hardware-in-the-loop — plus an agentic orchestration layer.

The component table is the useful part. Available, with Terraform or CDK paths: Foundation (S3, ECR, IAM, VPC), Cosmos (Predict V2V and Transfer 2.5), Isaac Lab on SageMaker and Batch, Isaac GR00T N1.6 fine-tuning, DreamZero LoRA fine-tuning in a standalone aws-samples repo, Isaac Sim workstations, OSMO 6.3 on EKS, and Strands Agents with the strands-robots SDK (MuJoCo by default, real hardware opt-in). Isaac Lab Arena Evaluation is Preview. Edge Deployment — “Model packaging to Jetson via EKS Hybrid Nodes + Greengrass” — is Planned. Amazon Bedrock AgentCore hosted runtime for Strands is also described as planned.

The LICENSE file is Apache License 2.0. GitHub’s API returns license key apache-2.0. Authors listed in the README are AWS solutions architects. A pick-and-place example uses a UR3 arm, a Robotiq 2F-85 gripper and 27 teleoperation episodes.

How does that differ from the marketing pillars?

The About Amazon post lists five pillars: Synthetic Data Generation, Model Training, Simulation and Validation, Edge Deployment, and Continuous Improvement. It names SageMaker, EC2 GPU instances, IoT Greengrass, Bedrock AgentCore, Isaac Sim, Isaac Lab, Isaac GR00T and Cosmos as if they were the running set. The README agrees on the NVIDIA training and simulation pieces and disagrees on edge: Greengrass appears in the architecture diagram’s “Deploy” sentence and in a “When to Use” row that points at jetson-edge-deployment, but the status column for Edge Deployment is still Planned.

If you need a shipped local-compute product rather than a robotics sample, that is a different AWS-and-NVIDIA story than this toolchain; see the RTX Spark hybrid-intelligence note. This repo is cloud IaC plus a planned edge path.

What does it cost, and what tokens do you need?

The README’s estimated-cost table is AWS’s own: about $2 for a GR00T smoke test on ml.g5.12xlarge for 15 minutes; about $79 for a full GR00T train (11 hours); about $10 / $93 for DreamZero smoke and 1,000-step fine-tunes on ml.g7e.24xlarge; about $37 an hour for Cosmos 3 Predict on p5.48xlarge Capacity Block; about $8 an hour for Cosmos Transfer 2.5 on g6e.12xlarge Spot; about $1.86 an hour for an Isaac Sim g6e.4xlarge; about $10–30 for Isaac Lab RL; about $5 an hour for a full OSMO stack. Resources are said to tear down with terraform destroy.

Prerequisites are an AWS account with GPU quota, AWS CLI v2, Python 3.11+, an NVIDIA NGC API key for container pulls, and a Hugging Face token for weights. Production path wants Terraform 1.5 or newer. “No Docker required locally — containers build in AWS CodeBuild.” We did not provision any of those resources.

What should readers not assume?

Amazon writes that it has “deploy[ed] more than 1 million robots across its operations network” and that the toolchain is “built with expert guidance inspired by those learnings.” That is first-party operations copy. The sample is not Vulcan, Proteus, or a humanoid you can order. It is also not a computer-use agent for a desktop; for that class of product see computer-use agents.

The GitHub repository existed in July. The 8 October post is the public launch note we are dating. A last push on 8 October does not mean every component listed as Available has been independently reproduced here. We have not applied the Terraform, pulled NGC containers, or run the UR3 example.