AWS has published a reference architecture for continuously improving robotics and autonomous-vehicle models with Nvidia Cosmos 3 on SageMaker HyperPod. The accompanying repository includes infrastructure templates and job manifests for building the cluster and scheduling the pipeline.
The design treats physical AI development as a loop rather than a one-time training job. It generates synthetic sensor data, post-trains perception and control models, and evaluates them in closed-loop simulation before repeating with new observations. Cosmos 3 handles video, images, actions and sound as one token stream, and can operate as a world model, action labeler or deployable policy.
Because the same model family covers the major stages, teams can time-share one persistent GPU pool instead of provisioning separate clusters for generation, training and evaluation. HyperPod supplies resilient managed compute while Amazon EKS schedules the individual workloads. AWS argues that the useful metric is GPU “goodput”—actual pipeline progress per reserved GPU-hour—because organizations pay for committed capacity even when a stage is idle.
Cosmos 3 spans Nano and Super tiers, plus a smaller Edge version for devices. The blueprint is aimed at infrastructure teams, not a turnkey robotics product: users still need data, evaluation criteria, capacity planning and safety validation for their application. Its practical contribution is a runnable pattern for keeping expensive accelerators occupied across a multi-stage development cycle.