All posts

Blog

AWS releases an open-source Physical AI Toolchain built with NVIDIA robotics software

FindAmNow
AWSAmazonNVIDIAphysical AIroboticsopen source

AWS has released the Physical AI Toolchain on AWS, an open-source set of reference architectures and code for training, simulating and deploying robots, built on AWS services and NVIDIA Isaac, GR00T and Cosmos. It is a development stack, not a ready-made robot.

Amazon Web Services has released the Physical AI Toolchain on AWS, an open-source collection of reference architectures, infrastructure-as-code and deployment automation for building robots and autonomous machines. It combines AWS cloud services with NVIDIA's physical AI software. Coverage of the launch appeared this week, including Interesting Engineering on 11 October 2026.

What it is

Amazon describes it as a stack aimed at three areas: industrial automation, autonomous mobility and humanoid robotics. The code is published in the aws-samples GitHub repository. Teams can deploy the whole thing or pick individual parts for an existing pipeline.

It is organised around five stages:

  1. Synthetic data generation – creating simulated training scenarios instead of recording every situation in the real world
  2. Model training – learning from human demonstrations and simulated practice
  3. Simulation and validation – testing behaviour virtually before touching hardware
  4. Edge deployment – running optimised models on the machine itself
  5. Continuous improvement – feeding operational data back into training

The components

On the AWS side, Amazon names Amazon SageMaker for training, Amazon EC2 GPU instances for simulation, AWS IoT Greengrass for edge deployment and Amazon Bedrock AgentCore for orchestration. The GitHub repository marks some pieces, such as the AgentCore runtime, as planned rather than available.

On the NVIDIA side, it integrates Isaac Sim (simulation), Isaac Lab (reinforcement learning), Isaac GR00T (humanoid robot training), Cosmos (synthetic world generation) and OSMO for workflow orchestration. Engineering.com reports it also supports common formats and tools such as PyTorch, ROS 2, the LeRobot data format, URDF and ONNX.

Prerequisites listed in AWS's technical material include an AWS account with approved GPU capacity, an NVIDIA NGC API key and a Hugging Face token.

Amazon's pitch

"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," said Uwem Ukpong, vice president of AWS Industries, as quoted by Assembly magazine.

Amazon says the design draws on its own robotics operations, which it reports include more than one million robots, and claims customers can launch physical AI capabilities "in weeks rather than the years" it would take from scratch. That timeline is Amazon's claim and will vary with hardware and use case.

Companies named in coverage as building in this space include NEURA Robotics, RLWRLD and Config.

What it is not

This is a development platform, not a robot or a model. As Interesting Engineering points out, it does not guarantee that a model trained in simulation will work reliably on a factory floor. The gap between simulation and reality, including friction, lighting changes and sensor noise, remains the hard part. GPU costs for simulation and training are also the customer's to carry.

Why it matters

Big cloud providers are moving from text and image AI toward machines that act in the physical world. A free, open-source starting point lowers the setup work for robotics teams, but it also ties that work closely to AWS and NVIDIA. For smaller manufacturers, the useful takeaway is that the building blocks are now documented and public, even if the cost of running them is not small.

Sources

Source