CoreWeave launches Physical AI Launchables: one-command templates for robotics and AV workloads
CoreWeave announced Physical AI Launchables at NVIDIA GTC Berlin on October 21, 2026: a catalog of robotics, autonomous vehicle and simulation workloads packaged as engineer-tested templates that let teams start a run with one command instead of rebuilding environments first.
At launch, templates span five tools - NVIDIA Isaac Sim, CARLA, MuJoCo, NVIDIA AlpaSim and NVIDIA Cosmos 3 - covering data generation, evaluation and grading, and policy training across the physical AI loop. Every tool also ships a blank template so teams are not limited to packaged workloads.
Launchables includes new Cosmos Data Factory agentic skills for generative AI data augmentation across video, image and text, plus auto-labeling, data curation and retrieval, and ships agent skills tuned for CoreWeave's cloud for Claude Code, Codex and other coding agents.
'Robotics and autonomous vehicle teams are doing some of the most demanding engineering in this industry, and they should be spending their time on it,' said Richard Ahlfeld, CoreWeave's senior vice president of physical AI. NVIDIA's Nakul Arora noted CoreWeave built its own optimized version of NVIDIA's reference agent skills.
Templates are designed to keep GPUs on GPU work: CPU-only preparation and post-processing finish before GPU execution where the workload permits, and capacity releases when the GPU stage ends. Partners building on CoreWeave include Cadence, Convergent Science, Encord, nTop, PhysicsX and Voxel51.
For engineering leads, Launchables attacks the setup week before the first run rather than model quality; the trade-off to evaluate is how far CoreWeave-tuned templates constrain portability back to other clouds or on-prem clusters.