Insider Brief
- CoreWeave launched Physical AI Launchables, a catalog of tested templates designed to let robotics, autonomous-vehicle and simulation teams run common workloads without rebuilding cloud environments for each job.
- At launch, the service supports Nvidia Isaac Sim, CARLA, MuJoCo, Nvidia AlpaSim and Nvidia Cosmos 3, with additional agent skills for Claude Code, Codex and other coding agents.
- CoreWeave said the templates can reduce setup and orchestration work by managing resource limits, GPU use, parallel jobs and results, while partners including Cadence, Encord, PhysicsX and Voxel51 are also bringing physical AI software to its infrastructure.
CoreWeave has launched Physical AI Launchables, a catalog of tested templates designed to let robotics, autonomous-vehicle and simulation teams run common software workloads without rebuilding their cloud environments for each job.
Unveiled at Nvidia GTC Berlin, the new offering packages software configurations that CoreWeave engineers have built and tested for physical AI workloads. Teams can point a template at their own scenes and data and start a run with a single command, according to the company.
According to CoreWeave, the system is intended to reduce the setup work that can come before robotics and AV development, including configuring compute and storage resources and preparing software environments. The templates can be used with reserved capacity for longer programs or for shorter burst jobs used to test an idea.
“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,” senior vice president of physical AI Richard Ahlfeld said in the announcement. “Physical AI Launchables gives them a tested starting point for the tools they already use, so more of their time goes to the work only they can do.”
Physical AI Launchables works with CoreWeave’s existing infrastructure, monitoring tools and Forge platform for developing models and agents. The new layer focuses on the engineering software used to build robots, autonomous vehicles and digital twins.
Templates for Physical AI Tools
At launch, Physical AI Launchables supports Nvidia Isaac Sim, CARLA, MuJoCo, Nvidia AlpaSim and Nvidia Cosmos 3, covering workloads such as data generation, evaluation, grading and policy training. Each tool also includes a blank template so teams can configure workloads beyond those already packaged by CoreWeave.
CoreWeave said the templates cover tasks including data generation, evaluation and grading, along with policy training. Each tool also has a blank template so developers can configure workloads beyond those already packaged by CoreWeave.
Launchables includes Cosmos Data Factory agentic skills for generative AI data augmentation across video, images and text, as well as auto-labeling, data curation and retrieval. CoreWeave said additional Cosmos Data Factory workflows are planned as templates.
The company is also shipping its own agent skills with each template for Claude Code, Codex and other coding agents. Those agents can adjust settings, start runs, monitor jobs and read the resulting output. CoreWeave said its version is tuned to run on its cloud infrastructure.
What the Templates Handle
CoreWeave said the templates are designed to reduce the infrastructure work around physical AI development:
- Reuse working environments: Developers can change settings or application code without rebuilding the underlying setup for each experiment.
- Limit unnecessary GPU use: When workloads allow it, CPU-based preparation and post-processing are completed outside the GPU stage so capacity can be released sooner.
- Run jobs in parallel: Templates define resource and concurrency limits to keep simultaneous workloads from competing with one another.
- Automate runs: Scripts and coding agents can launch jobs, monitor progress and retrieve results.
- Centralize results: Reports, metrics, media and diagnostic information return through the same interface.
CoreWeave said those features let teams work with the models, datasets, environments and evaluation tools they already use while the platform handles packaging, deployment and execution underneath.
The company is also working with physical AI software providers to run their products on CoreWeave infrastructure. Partners include Cadence, Convergent Science, Encord, nTop, PhysicsX and Voxel51, covering areas ranging from vision and data tooling to engineering simulation.
