AWS introduces open-source tools to train, simulate and deploy intelligent machines

AWS has introduced the Physical AI Toolchain on AWS, an open-source stack that connects its cloud services with Nvidia software to develop intelligent machines. Aimed at industrial automation, autonomous mobility and humanoid robotics, it supports physical AI: machines that sense their surroundings, reason about what is happening and act in the real world.

Less infrastructure work before the machine moves

The catalyst was customer frustration with the work needed before building a useful machine. AWS Industries vice president Uwem Ukpong said customers were spending too much engineering effort on infrastructure. AWS’s response combines architecture guidance, deployment automation and ready-to-use code, rather than offering only a model or a simulation tool.

AWS says the design draws on lessons from Amazon’s robotics operations, where the company says it has deployed more than 1 million robots. That experience informs the toolchain’s guidance for building machines that operate autonomously in real environments. Customers still supply the hardware and knowledge of their own industry.

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. We want to flip that.

Uwem Ukpong, vice president, AWS Industries

From generated worlds to real machines

The workflow starts with synthetic data generation: creating training scenarios in AI-generated environments to reduce reliance on costly real-world data collection. Models then learn from human demonstrations and practice in simulated environments. Simulation and validation provide a place to test machine behavior before deploying it to physical hardware.

After testing, optimized models move to machines in the field. This is edge deployment: decisions happen on the machine without constant cloud connectivity. Operational data then flows back into the development process to generate new training data, closing the loop between what a machine learns and what it encounters.

The software behind the workflow

• AWS components: Amazon SageMaker handles model training; Amazon EC2 GPU instances support simulation; AWS IoT Greengrass supports edge deployment; and Amazon Bedrock AgentCore provides intelligent orchestration.
• Nvidia components: Isaac Sim provides simulation; Isaac Lab supports reinforcement learning, or learning through practice and feedback; Isaac GR00T supports humanoid machine training; and Cosmos generates synthetic worlds.

Adoption does not require the whole stack

Customers can manage the full development cycle from a single control plane, or select only the simulation, training or deployment components that fit existing workflows. Fleet-management capabilities support provisioning, securing and updating thousands of machines over the air as deployments scale into production.

AWS says its guidance, reference code and automation can help manufacturers launch physical-AI capabilities in weeks rather than the years required to start from scratch. That is the company’s acceleration claim, not an independently measured result. Each deployment still needs tailoring to the customer’s hardware, operating environment and intended task.