CoreWeave launches one-command templates for Isaac Sim, CARLA, MuJoCo and Cosmos 3

What CoreWeave launched

CoreWeave, the GPU cloud company, has launched Physical AI Launchables, a catalog of ready-to-run templates for robotics, autonomous driving and simulation work. It was unveiled on 21 October 2026 at NVIDIA GTC Berlin, and the press release sums up the pitch in one line: teams start a run with one command instead of rebuilding their environment first.

The company frames the problem as the week before the run. In robotics and self-driving teams, someone usually has to stand up a simulator, wire storage, install drivers and wait on a job scheduler before any real experiment starts, and that work is repeated for every new tool or project. CoreWeave says its own engineers built and tested a configuration for each workload, so the environment is already there.

Richard Ahlfeld, CoreWeave's senior vice president of physical AI, said robotics and AV teams "should be spending their time" on the hard engineering, and that Launchables gives them "a tested starting point for the tools they already use".

The templates at launch

At launch the catalog covers five tools, and the CoreWeave blog post lists the specific templates under each. Every tool also ships a blank template: the tested environment with nothing in it, for teams whose project is not on the list.

NVIDIA Isaac Sim, the robot simulator, gets four ready-made jobs: a Franka arm pick-and-place with episode video, a forklift-and-pallet scene that produces labeled synthetic data, vehicle variants with driving and braking trajectories, and a warehouse scene with workers and incidents that outputs labeled multi-camera data.

CARLA, the open-source driving simulator, gets a scenario matrix that sweeps driving scenarios across towns, weather and traffic and returns video. MuJoCo, the physics engine used widely in robot learning, gets a GPU-parallel reinforcement learning sweep on MuJoCo Warp and a kitchen pick-and-place job using RoboCasa.

NVIDIA AlpaSim gets a closed-loop driving template that evaluates NVIDIA's Alpamayo driving model over reconstructed real scenes, the same model family Bolt picked for its robotaxis this week. NVIDIA Cosmos 3 gets a scenario-generation template that uses Cosmos 3 Nano to make text-to-video clips of rare or costly situations. Voxel51's FiftyOne dataset tool is listed as coming soon.

What a template actually does

A template is more than a container image. CoreWeave says each one declares what it needs and what it returns, so you can tell whether it fits before you run it. Changing a run means passing name and value pairs on the command line, for things like scenarios, duration, parallelism or which model is under test, and pointing it at your own scenes and data. For bigger changes you can edit the template's application code directly; the blog shows Isaac Sim variations, each hand-edited and launched in about 30 seconds.

Three design points stand out. Templates finish CPU-only preparation and post-processing outside the GPU stage where the workload allows, and release the GPUs as soon as the GPU work ends, which matters on a cloud that bills for GPU time. Each template sets its own resource and concurrency limits so parallel runs do not compete with each other. And results, including reports, metrics, media and diagnostics, come back through one interface.

The product page makes a useful distinction on results. A CARLA campaign reports three things separately: whether execution completed, whether outputs came back for every scenario, and how the results scored against your driving criteria. A finished job, CoreWeave notes, is not proof that every scenario ran or passed. Templates can run against reserved capacity for a long program or as a burst job to test one idea.

Built for coding agents too

CoreWeave is pitching the catalog as much to AI coding agents as to engineers. Every operation is documented and every response is machine-readable, so a script or an agent can launch a run, watch it and fetch the results. Each template ships with CoreWeave's own agent skills, which the company says work with Claude Code, Codex and other coding agents.

NVIDIA publishes reference agent skills for its physical AI software that are built to run on any cloud. CoreWeave's version is tuned for its own infrastructure. Nakul Arora, a director of product management at NVIDIA, said that lets these workflows "run natively" so developers can "focus on the physical AI problem rather than the orchestration underneath".

The catalog also includes NVIDIA's new Cosmos Data Factory agentic skills for augmenting video, image and text data, auto-labeling, curation and retrieval, which we covered when NVIDIA expanded Cosmos with Data Factory blueprints. CoreWeave says more Data Factory workflows will arrive as templates. Because the template reports its output separately from whether the job finished, an agent can change one variable, rerun and tell whether it helped.

Partners and the GPU choice

Not every team wants to leave its existing tools, so CoreWeave is also lining up software partners whose products run on its compute while keeping their own interface. Named partners are Cadence, Convergent Science, Encord, nTop, PhysicsX and Voxel51, spanning data and vision tooling through engineering simulation.

Encord chief executive Eric Landau said combining CoreWeave compute with Encord's data platform makes it easier for teams to work with multimodal data. Voxel51 chief executive Brian Moore said the pairing helps "accelerate the data flywheel". Voxel51 recently added a data agent and a free tier to FiftyOne.

On hardware, the product page says CoreWeave picks the GPU for each template, and for rendering-heavy simulation and video work its default is the NVIDIA RTX PRO 6000 Blackwell Server Edition, which has RT cores for rendering and hardware video encoding. The blog points to a solution brief comparing AlpaSim closed-loop driving on that card with the H100.

What is missing

CoreWeave did not publish prices in the announcement, blog or product page. Access is through a briefing request: the page asks teams which tool and workload they want, and says an account team will set up a template on their data. That makes this closer to a managed onboarding offer than a self-serve catalog for now.

There is also no public list of which software versions, drivers or licenses each template is qualified against, beyond a frequently asked questions section on the product page. And CoreWeave is explicit that templates remove infrastructure work, not engineering judgment: the checks tell you a run completed and scored, not that your robot policy is correct.

How it compares

The big clouds are chasing the same customers. Amazon Web Services earlier this month packaged its own physical AI toolchain, and simulation stacks built on MuJoCo keep improving, as Google's Gemini 4 Argon robotics demos showed. CoreWeave's angle is narrower: pre-tested jobs for the specific simulators robotics and AV teams already use, on GPUs it chooses for them.

That fits CoreWeave's broader push beyond raw GPU rental, including its Forge platform for improving models and agents. For a robotics startup without a platform team, the value is skipping setup. For a large lab that already has its own cluster tooling, the appeal will depend on price, which is still unknown.

What to watch

Watch for published pricing or a self-serve option, the arrival of the Voxel51 FiftyOne template and more Cosmos Data Factory workflows, and any customer that reports how much setup time the templates actually saved. Independent numbers from the AlpaSim solution brief, especially cost per simulated mile on RTX PRO 6000 versus H100, would also help teams judge whether the default hardware choice is the right one for them.