CoreWeave 推出 Physical AI Launchables，為機械人及自動駕駛開發提供工作負載範本

Before a robotics team can test a new policy, someone has to make the simulator, data, storage, and compute work together. CoreWeave is targeting that preparation work with Physical AI Launchables, a catalog of tested workload templates announced on October 21 for robotics, autonomous vehicles, and simulation.

The templates package environments for five tools—NVIDIA Isaac Sim, CARLA, MuJoCo, NVIDIA AlpaSim, and NVIDIA Cosmos 3—so engineers can configure and start a workload with a command rather than assemble its infrastructure from scratch. The launch also includes skills that let coding agents operate the same workflows.

For developers, the important question is what becomes repeatable. Launchables aims to preserve a working environment between experiments while letting teams change the scenes, datasets, models, and settings that determine the result.

Moving Setup Work Beneath the Experiment

A physical AI experiment involves more than running a model. A driving evaluation needs a simulator and scenarios; a robot-learning job needs an environment, a policy, and a way to collect outcomes. Those components also need appropriate storage, compute resources, and execution limits.

CoreWeave says its engineers have built and tested the configurations behind each template. Teams can adapt those configurations to their own work, using reserved capacity for ongoing programs or burst jobs for individual experiments. Each supported tool also has a blank template that supplies the environment without prescribing the workload.

“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,” said Richard Ahlfeld, CoreWeave’s senior vice president of physical AI, in the announcement.

The distinction is useful: CoreWeave supplies the packaging and execution layer, while the simulation and modeling software comes from NVIDIA and other developers. Teams still decide what to test and whether the results are meaningful.

What the First Templates Can Run

CoreWeave’s technical launch post makes the catalog more concrete. Its examples cover robot manipulation, simulated driving, synthetic data, and reinforcement learning:

• NVIDIA Isaac Sim: Robot-arm pick-and-place, labeled warehouse scenes, and vehicle-dynamics experiments.
• CARLA: Driving-scenario batches that produce video for inspection.
• MuJoCo: Parallel reinforcement-learning experiments and kitchen-manipulation tasks.
• NVIDIA AlpaSim: Closed-loop driving evaluation.
• NVIDIA Cosmos 3: Generating video scenarios that may be difficult or expensive to collect.

These are starting configurations rather than a fixed set of demonstrations. Engineers can change workload parameters or edit application code while retaining the tested environment beneath them. The launch also includes NVIDIA Cosmos Data Factory skills for data augmentation, labeling, curation, and retrieval; CoreWeave says additional related templates are planned.

Coding Agents Get the Same Controls as Engineers

Launchables includes CoreWeave-specific skills for Claude Code, Codex, and other coding agents. These give an agent guidance about relevant settings and access to the commands used to launch, monitor, and retrieve a run.

The release emphasizes a distinction between a job completing and an experiment succeeding. A process can exit without establishing that a driving policy handled its scenarios well or that a robot achieved its task. Returning reports, metrics, media, and diagnostics through a consistent interface gives both engineers and agents material to assess after execution.

That can make automation more useful across repeated experiments. An agent can help operate a workflow, but the team still needs appropriate evaluations and success criteria. A functioning deployment does not establish that a model is ready for a physical system.

Keeping GPU Time Focused on GPU Work

The announcement also addresses how jobs consume resources. Where a workload permits it, templates separate CPU preparation and post-processing from GPU execution, releasing GPU capacity after that stage finishes. Templates declare resource requirements and concurrency limits to help coordinate parallel runs.

This matters because the experiment’s cost includes the time expensive hardware spends waiting for other work. Separating stages can improve allocation, although the actual benefit depends on the workload and how much preparation, simulation, training, and analysis it requires.

CoreWeave has not provided a general Launchables benchmark or a universal setup-time reduction in the supplied announcement. Its immediate proposition is a tested configuration and a common way to operate it, rather than a measured performance gain that applies to every robotics project.

How Launchables Fits CoreWeave’s Physical AI Stack

The release extends CoreWeave beyond raw GPU capacity into the software workflows that consume it. On its physical AI platform page, the company describes infrastructure spanning GPU and CPU compute, networking, storage, Kubernetes orchestration, and Slurm on Kubernetes. Those components support the different demands of data preparation, simulation, and model training.

Its broader cloud platform connects that infrastructure with CoreWeave Forge, which brings together tools for observing, evaluating, and improving models and agents. Launchables adds tested engineering workloads above those layers. The strategy is to reduce the number of environments a team must assemble as it moves between data, experiments, and evaluation.

CoreWeave also named Cadence, Convergent Science, Encord, nTop, PhysicsX, and Voxel51 as software partners working with its infrastructure. Those relationships span data and computer-vision tooling through engineering simulation; they should not be read as a claim that every partner product is already a launch template.

A Faster Start Still Needs a Sound Evaluation

The opportunity is straightforward: when an environment survives from one experiment to the next, teams can spend more time comparing models and investigating failures. Shared templates can also make it easier for colleagues to reproduce a configuration before changing a variable.

The harder questions remain in the domain itself. Simulated scenes must represent useful conditions, synthetic data must be appropriate for the task, and evaluations must expose failures that matter outside the simulator. Launchables can simplify the path to running those tests; it does not remove the need to design them carefully.

CoreWeave directs interested teams to its Physical AI Launchables overview for the initial catalog and further information. For robotics and autonomous-vehicle developers, the practical appeal is a shorter path from an experiment idea to results they can inspect.