Amazon Devices & Services has deployed a simulation-first manufacturing solution this month that pairs Amazon-created software with NVIDIA digital twin technologies to move toward zero-touch production. The system trains robotic arms to inspect a range of products for quality auditing and to integrate new goods onto the line without hardware changes. Amazon´s approach relies heavily on synthetic data and a modular workflow, aiming to replace time-consuming physical prototyping with virtual, repeatable experiments.
The technical pipeline stitches together several NVIDIA stacks and Amazon cloud services. Amazon imports CAD models into NVIDIA Isaac Sim on the Omniverse platform, then generates more than 50,000 synthetic images per device to train object- and defect-detection models. NVIDIA Isaac ROS produces robotic trajectories, while cuMotion and the nvblox library support fast collision-free planning on NVIDIA Jetson AGX Orin modules. FoundationPose, a foundation model trained on 5 million synthetic images, provides pose estimation and object tracking that can generalize to unseen objects. On the cloud side, AWS accelerated model development using Amazon EC2 G6 instances and AWS Batch; Amazon Bedrock and Bedrock AgentCore handle higher-level task planning and ingest multimodal product specifications.
That collection of tools is designed to enable what Amazon describes as ´zero-shot manufacturing´ and a broader move to generalized manufacturing. Lines can switch from auditing one product to another with software updates alone. The modular design already supports defect detection during production and is built to integrate more advanced reasoning components in the future, including NVIDIA Cosmos Reason for deeper analysis and decision-making. Eliminating the need for physical prototypes reduces cost and shortens the time it takes to bring new devices to consumers.
The deployment at an Amazon Devices facility demonstrates a path from simulation to the real world, with simulated stations matching real ones for training and validation. The result is faster on-boarding of new products, more flexible robotic operations, and a clearer route to autonomous, software-driven manufacturing pipelines that scale across products and stations.
