Building a Physical AI model factory with NVIDIA Cosmos 3 on SageMaker HyperPod

AWS ML Blog · rss · 2026-09-05

AWS ML Blog publishes an end-to-end guide for building a Physical AI model factory with NVIDIA Cosmos 3 on SageMaker HyperPod. Cosmos 3 uses a single token stream across video/image/action/sound with a Mixture-of-Transformers design (per-layer dual-stream attention joining a reasoner and generator) and train/inference asymmetry. One architecture runs three modes — forward-dynamics world model for synthetic data, inverse-dynamics action labeler, and deployable policy — across Nano (16B), Super (64B), and Edge (4B on-device) tiers. The pipeline shares one persistent GPU node pool, making GPU goodput the governing cost metric; a full robot-policy post-training walkthrough on DROID plus open-source manifests is included.

Original post →

More from Embodied

Embodied channel →