Amazon has introduced the Physical AI Toolchain on AWS, an open-source stack for building machines that act in the physical world. The company says it is designed for industrial automation, autonomous mobility and humanoid robotics.
The toolchain brings together steps from creating training data to running AI models on machines. Amazon says users can generate varied training scenarios in AI-created environments, reducing the need to collect data in the real world. They can then train models using human demonstrations and practice in simulated environments.
Before a model reaches hardware, the stack lets users test machine behavior in virtual environments. It also supports deploying optimized models to machines in the field, where they can make decisions without a constant cloud connection.
Read nextAWS Launches Open-Source Strands Box to Restrict AI AgentsThe toolchain uses AWS and NVIDIA technologies
The stack uses Amazon SageMaker for model training, Amazon EC2 GPU instances for simulation, AWS IoT Greengrass for deployment to machines and Amazon Bedrock AgentCore for orchestration. It also incorporates NVIDIA Isaac Sim for simulation, Isaac Lab for reinforcement learning, Isaac GR00T for humanoid machine training and Cosmos for generating synthetic worlds.
Amazon says the toolchain draws on its robotics operations and is built on AWS using NVIDIA’s physical AI stack. After deployment, the process described by Amazon feeds operational data from machines back into the toolchain to create new training data.
