# AWS Physical AI Toolchain is open source, and its own repository marks edge deployment as planned

URL: https://onticpost.com/aws-physical-ai-toolchain-robots/
Published: 2026-10-16
Updated: 2026-10-15
By: Hill, Onticpost

> AWS Physical AI Toolchain code is Apache 2.0, but its training stages run on paid AWS GPUs and its repository lists the edge deployment stage as planned.

Amazon Web Services introduced the AWS Physical AI Toolchain on October 7, 2026, in a post on its own blog, and Amazon’s newsroom called it an open-source stack for building intelligent machines the next day. The code sits in a public AWS repository under the Apache License 2.0. Its main README lists ten components: eight are marked available, one is a preview, and one, the stage that puts trained models onto robots, is marked planned. Amazon’s announcement names AWS IoT Greengrass for edge deployment among the services that power the toolchain. Onticpost has no robot hardware and has not run the toolchain, so every figure below comes from AWS, Amazon or GitHub documents we saved.

**Takeaway points**

- AWS introduced the Physical AI Toolchain on October 7, 2026, as a set of reference architectures, Infrastructure as Code and deployment automation, and its repository README lists ten components, of which eight are marked available.
- The code is Apache 2.0, but every training and simulation stage is written to run on paid AWS services, with AWS’s own estimates running from about $1.86 an hour for an Isaac Sim workstation to about $37 an hour for Cosmos 3 Predict.
- Edge deployment and the hosted Amazon Bedrock AgentCore runtime are both marked planned in the repository, and we have not run any part of the toolchain.

## What is in the AWS Physical AI Toolchain?

The AWS Physical AI Toolchain is a collection of Terraform modules, containers, notebooks and workshop guides that set up NVIDIA’s robot software on AWS. The blog post, written by four AWS solutions architects and two NVIDIA staff, calls it “a curated collection of reference architectures, Infrastructure as Code, and deployment automation”. It is not a robot, a model or a new simulator. The blog says plainly that it “provides infrastructure patterns and architectural guidance, not finished robot behaviors.”

The repository’s README sets out the parts and the state of each.

| Component | What the README says it does | Status in the README |
| --- | --- | --- |
| Foundation | Shared S3 buckets, ECR repos, IAM roles, VPC, SSM parameters | Available |
| Cosmos | NVIDIA Cosmos world generation and data augmentation (Transfer 2.5) | Available |
| Isaac Lab | Reinforcement learning training, 4096 parallel environments, on SageMaker and Batch | Available |
| Isaac GR00T | Fine-tune the GR00T N1.6 vision-language-action model on SageMaker and Batch | Available |
| DreamZero | Fine-tune DreamZero, a 14B World Action Model, with LoRA on SageMaker | Available |
| Isaac Sim | GPU workstation for physics simulation | Available |
| OSMO | NVIDIA OSMO 6.3 orchestration on EKS | Available |
| Strands Agents | Natural-language robot control with the Strands Agents SDK | Available |
| Isaac Lab Arena Evaluation | GR00T closed-loop evaluation and gated SageMaker registration | Preview |
| Edge Deployment | Model packaging to Jetson via EKS Hybrid Nodes and Greengrass | Planned |

The toolchain ships with one worked example. The README says it uses a UR3 arm with a Robotiq 2F-85 gripper and includes 27 real teleoperation episodes for a pick-and-place task. The blog says the toolchain is “robot-agnostic and task-agnostic”, and that a team brings its own URDF robot description, its own teleoperation data and its own task.

Uwem Ukpong, vice president of AWS Industries, gave the reason for it in Amazon’s newsroom piece: customers had told AWS that “too much of their engineering effort was going to infrastructure instead of innovation.” Software for factory robots is opening up from several directions this year. We reported that [Intrinsic releases Intrinsic Core as open source under Apache 2.0](https://onticpost.com/intrinsic-core-open-source-robot-software/), and that [Qualcomm agrees to buy the company behind MoveIt in the Qualcomm PickNik deal](https://onticpost.com/qualcomm-acquires-picknik-moveit/).

## Which simulators and frameworks does the AWS toolchain name?

The toolchain is built around NVIDIA’s robot stack, and the simulators it names are NVIDIA Isaac Sim, NVIDIA Isaac Lab and, in one component, MuJoCo. The blog lists five NVIDIA parts by role: GR00T for vision-language-action training, Isaac Sim for physics simulation, Isaac Lab for reinforcement learning, Cosmos for synthetic data generation and OSMO for workflow orchestration. It names Jetson and RTX hardware as the targets for running models on a robot.

Some of the other parts are not from NVIDIA. The blog says the toolchain ingests raw recordings in Zarr, converts them to the LeRobot format, trains with PyTorch and Hugging Face, defines reinforcement learning tasks with Gymnasium, describes robots with URDF, exports to ONNX and hands models to ROS 2 for control on the robot. The README’s pipeline adds TensorRT for the final export.

The Strands Agents component uses MuJoCo by default. Its README says one call starts a MuJoCo simulation by default, with no GPU and no hardware, and that real servos move only when a user passes the real mode. For the hosted version on Amazon Bedrock AgentCore, which is still planned, it names Claude, reached through Amazon Bedrock, as the reasoning model.

The models the toolchain fine-tunes are named exactly in the component documents. The GR00T README describes GR00T as a 3B-parameter model and supports both N1.6 and N1.7, where the main README names only N1.6. The DreamZero model is described as 14B. The Isaac Lab Arena component also pairs GR00T, OpenVLA and MolmoAct2 with the LIBERO benchmark.

## What is open source in the AWS Physical AI Toolchain, and under which licence?

The repository’s own code is under the Apache License 2.0, version 2.0 of January 2004, according to the LICENSE file at its root and GitHub’s record of the repository. The README repeats “Apache 2.0”, and so does the Isaac Sim component’s README.

The Isaac Lab Arena component carries its own NOTICE file. It lists the outside projects that component “integrates or adapts”: NVIDIA Isaac-GR00T, NVIDIA Isaac Lab Arena and AllenAI LeRobot under Apache License 2.0, and OpenVLA-OFT and LIBERO under the MIT License.

The licence covers the AWS repository. It does not make the NVIDIA software the toolchain installs part of the same package. The README requires an NVIDIA NGC API key “for container image pulls” and a Hugging Face token “for model weight downloads”. The Isaac Sim component runs from an “NVIDIA Isaac Sim Marketplace AMI”, which the README says needs a subscription, “free, one-time per account”. The Isaac Lab Arena README asks for a Hugging Face token with “the GR00T N1.6 model terms” accepted, and an NGC key “with the required NVIDIA terms accepted”. None of the documents we saved states the licence terms of that machine image or of the NVIDIA containers.

The words used differ between the two announcements. Amazon’s newsroom piece calls the toolchain “open-source” twice, once in its own text and once in a quote from Amit Goel of NVIDIA. The AWS blog post does not use the term and calls it “publicly available”.

GitHub’s record shows the repository was created on July 9, 2026, three months before the announcement, and last pushed on October 9, 2026.

## What in the AWS toolchain runs only on paid AWS services?

Every training, simulation and data generation stage is written for paid AWS compute. The README’s prerequisites start with “AWS account with GPU quota (SageMaker + EC2)”, and it notes that no Docker is needed locally because “containers build in AWS CodeBuild”. The table below is AWS’s own estimate of cost, from the README.

| Component | AWS’s estimated cost | Instance named |
| --- | --- | --- |
| GR00T training, smoke test | About $2 | ml.g5.12xlarge for 15 minutes |
| GR00T training, full | About $79 | ml.g5.12xlarge for 11 hours |
| DreamZero fine-tune, smoke gate | About $10 | ml.g7e.24xlarge for 25 minutes |
| DreamZero fine-tune, 1000 steps | About $93 | ml.g7e.24xlarge for 4 hours 11 minutes |
| Cosmos 3 Predict | About $37 an hour | p5.48xlarge, Capacity Block |
| Cosmos Transfer 2.5 | About $8 an hour | g6e.12xlarge, Spot |
| Isaac Sim workstation | About $1.86 an hour | g6e.4xlarge |
| Isaac Lab reinforcement learning | About $10 to $30 | ml.g5.xlarge for 2 to 4 hours |
| OSMO, full deployment | About $5 an hour | EKS, RDS and ElastiCache |

The Isaac Sim README adds a cost the main table leaves out: a 512 GB disk at about $40 a month, which “persists when instance is stopped”. It also warns that the instance “bills per second while running”. That workstation is a g6e.4xlarge with an L40S GPU and 48 GB of video memory, in AWS’s description.

One part runs without AWS. The Strands Agents README says its simulation path “needs no AWS resources”, and its default control plane runs on a laptop or workstation against MuJoCo.

## Do Amazon’s announcement and the AWS repository agree?

They do not agree on what is finished. Amazon’s newsroom piece lists “AWS IoT Greengrass for edge deployment, and Amazon Bedrock AgentCore for intelligent orchestration” among the services powering the toolchain. The repository marks both as not yet built.

| Point | Amazon newsroom, October 8 | AWS repository |
| --- | --- | --- |
| Edge deployment with Greengrass | Named as part of the toolchain | Edge Deployment row: “Planned” |
| Bedrock AgentCore | Named for orchestration | Strands README: deployment to AgentCore is “Planned” |
| Number of pillars | “five core pillars” | “four pillars” with an orchestration layer at the center |
| Time to get going | “in weeks rather than the years” | The blog says “in hours rather than weeks” |

The AWS blog post sits between the two. It says that “final deployment pushes optimized models to NVIDIA Jetson hardware using AWS IoT Greengrass”, in the present tense, while the repository it links to lists that stage as planned.

The README also points readers to a component that is not there. Its table of scenarios says that for edge deployment a team should “Use jetson-edge-deployment to package and push models”. The list of the repository’s 594 files and folders that we saved from GitHub contains no path with “jetson” or “greengrass” in it.

There is a smaller mismatch on the GR00T smoke test. The main README gives about $2 on ml.g5.12xlarge for 15 minutes. The GR00T component README gives the same about $2 and about 15 minutes, on ml.g5.xlarge.

The blog also says each component “is an independent Terraform module”. The README shows that DreamZero is deployed with the AWS CDK, from a separate standalone repository.

## What has AWS not published about the toolchain?

AWS has not given a date for the edge deployment stage or for the hosted AgentCore runtime, and it has not named any company running the toolchain in production. Amazon’s newsroom piece describes three robotics firms building on AWS and quotes the chief executive of one, NEURA Robotics, saying the toolchain “helps us accelerate exactly that cycle”. It does not say how NEURA uses it, or whether any robot trained with it works on a production line.

AWS gives no measure of what a trained policy achieves. The repository supplies one example task and an evaluation pipeline that the main README marks as a preview, with a pointer to that component’s validation status. The main README makes no claim about success rates on real robots.

The licence terms of the NVIDIA software the toolchain pulls in, from the Marketplace image to the NGC containers, are not set out in any of the AWS documents we saved.

Amazon says it has deployed “more than 1 million robots” across its operations network and that the toolchain draws on those lessons. That is Amazon’s claim. The newsroom piece does not say which of its own robots, if any, were trained with this toolchain.

We have not run the AWS Physical AI Toolchain, on AWS or on our M4 Pro with 48 GB of memory, and we have no robot to run it on. Our account rests on the AWS blog post of October 7, 2026, Amazon’s newsroom piece of October 8, 2026, and the repository’s README, LICENSE, NOTICE and component documents, saved on October 15, 2026. Robots that learn on site are reaching factories by other routes: [Boston Dynamics Atlas robots start training at a new center inside Hyundai’s Georgia plant](https://onticpost.com/boston-dynamics-atlas-metaplant-application-center/).

## Sources

- [AWS blog, Introducing AWS Physical AI Toolchain, October 7, 2026](https://aws.amazon.com/blogs/physical-ai/introducing-aws-physical-ai-toolchain/)
- [AWS repository, README and Apache 2.0 LICENSE](https://github.com/aws-samples/sample-the-physical-ai-toolchain-on-aws)