CoreWeave packages robotics and self-driving development tools into cloud-ready templates

Robotics and autonomous-vehicle teams can start simulation and training runs from tested configurations rather than rebuild their computing environments, CoreWeave says. On October 21, 2026, at NVIDIA GTC Berlin, it introduced Physical AI Launchables: workload templates designed to start with one command, using the scenes and data engineers bring to the job.

Five tools, with room to change the workload

CoreWeave engineers built and tested the configurations for software used in robotics, autonomous vehicles and simulation. The catalog covers NVIDIA Isaac Sim, CARLA, MuJoCo, NVIDIA AlpaSim and NVIDIA Cosmos 3. Its packaged workloads span data generation, evaluation and grading, and policy training—the work of teaching a system how to act.

Each supported tool also comes with a blank template, so teams are not restricted to CoreWeave’s prepared workloads. They can change settings or application code without rebuilding the environment underneath.

Teams can use reserved computing capacity for a sustained program or run a burst job to test one idea. CoreWeave handles packaging, deployment and execution beneath the models, environments and evaluations the engineers work with.

CoreWeave has built their own version optimized for their AI cloud and ships it with Physical AI Launchables, so these workflows run natively on their infrastructure and developers can focus on the physical AI problem rather than the orchestration underneath.

Nakul Arora, NVIDIA director of product management, in CoreWeave’s announcement

Agents read the results, not just the job status

Every template ships with agent skills for Claude Code, Codex and other coding agents, tuned for CoreWeave’s cloud. These are instructions for operating the workload: CoreWeave says an agent knows which settings are worth changing, issues the same command an engineer would and reads the output separately from checking whether the job finished.

Operations are documented and responses are machine-readable, allowing scripts and coding agents to launch runs, monitor progress and retrieve results. Reports, metrics, media and diagnostics return through the same interface.

Launchables also includes Cosmos Data Factory agent skills for augmenting video, image and text data with generative AI. They cover automatic labeling, data curation and retrieval. CoreWeave says more Cosmos Data Factory workflows are coming as Launchable templates.

Keeping setup work off the GPUs

The templates also control how work uses computing resources. CoreWeave describes three mechanisms:

• Finish CPU-only preparation and post-processing before GPU execution begins, where the workload allows it.
• Release GPU capacity as soon as the GPU stage ends.
• Declare resource and concurrency limits in each template to keep parallel runs from competing with one another.

Alongside the templates, CoreWeave is working with Cadence, Convergent Science, Encord, nTop, PhysicsX and Voxel51 to run their products on its infrastructure. Those relationships span vision and data tools through engineering simulation, with the stated goal of letting teams stay in software they already use.

CoreWeave’s announcement directs readers to request a briefing. It does not specify Launchables pricing or a general-access date.