AWS launches open-source Physical AI Toolchain for robotics

AWS says the toolchain addresses the challenge of connecting the components needed to make trained models work reliably in the real world.[1] Its described workflow runs from synthetic data generation to model training, simulation and validation, edge deployment, and continuous improvement.[1] Amazon describes the toolchain as purpose-built for industrial automation, autonomous mobility, and humanoid robotics.[3] The project builds on a framework AWS unveiled in December 2025.[2] AWS’s technical blog material also covers NVIDIA Cosmos 3 world models and Isaac Lab, which the evidence describes as a reinforcement-learning framework.[2]

The AWS side of the workflow uses Amazon SageMaker for model training and AWS IoT Greengrass to distribute models to edge devices.[1] NVIDIA components include Isaac Sim, Isaac Lab, Isaac GR00T, and Cosmos.[1] Companies can use these parts separately or combine them into an end-to-end workflow, giving teams the option to adopt selected pieces rather than assemble the entire toolchain.[1] The toolkit’s GitHub materials include reference architectures, Terraform Infrastructure-as-Code templates, and deployment automation samples.[2] AWS says operational data collected by deployed robots can return to the cloud to improve models.[1]

AWS estimates that running the GR00T full-training sample costs approximately $79.[2] It estimates that the Cosmos Predict workload costs around $37 per hour when run on p5.48xlarge instances.[2] These figures describe particular samples and workloads; they do not establish a general price for using the toolchain. AWS says training these models requires significant data, while little data is available.[1] It also says the new toolchain is not a direct replacement for RoboMaker.[1] A cloud-based robotics simulation platform was shut down in 2025, but the evidence does not identify it as RoboMaker.[1]