CoreWeave Physical AI Launchables gives your team a catalog of ready-to-run Templates for robotics and autonomous vehicle programs. Each Template runs on a configuration tested and maintained by CoreWeave and delivers an exact report of results.
Each Template carries a manifest that declares what it needs and what it returns, the software at pinned versions, a configuration tested on CoreWeave, the execution path that runs it, and the results it delivers. A workload that runs once is a demo. A qualified Template is a foundation your team can build on.
Physical AI teams should be free to focus on building better models, policies, and products. Rely on Templates we maintain instead of a toolchain you build and manage yourself. Templates address the key challenges your team faces every day.
Every Template ships tested on CoreWeave, with the environment, configuration, execution, and result collection already packaged. Your engineers work on the task, not the tool.
Runtimes, drivers, and licenses change. Keeping Templates working is our job, not yours. Every Template follows the same conventions for inputs, launch, and results, so setup carries over to the next workload.
Every Template reports whether it ran, whether all the outputs arrived, and how they scored. If anything is missing, the Template tells you. Then you judge the results against your own criteria.
Physical AI teams work in a loop. They generate data in simulation, curate it, train a policy, and evaluate the result, then carry what they learn into the next round. Every stage runs on its own tools with its own setup. Templates cover the loop one workload at a time, starting with generation, evaluation, and policy training today.
The catalog is organized by the software your team already uses, with Templates underneath for specific jobs: a scene to generate, a scenario set to score, a policy to train. Every tool also ships a blank: the tested environment, with the workload left to you.
Robot and vehicle simulation with photoreal rendering, used for manipulation evaluation and synthetic data generation.
Open-source driving simulator for scoring a policy across a scenario matrix.
Physics simulation for contact-rich manipulation and GPU-parallel reinforcement learning.
Closed-loop driving simulation, where each decision re-enters the environment and compounding errors surface.
World-model video generation for scenarios that are rare or costly to capture.
Once you’ve run one, you can run them all.
Each Template lists what’s been tested and where supported behavior ends.
.svg)
Point the Template at your data, then set the run config: which scenarios, how long, how many repeats, what counts as a pass, and what comes back.
All it takes is one command, whether a person issues it or an agent. You get an estimate of GPU-seconds required before anything starts.
Execution, outputs, and results come back as three separate facts. Inputs are recorded, so any run can be set up again from its record.
Define driving scenarios across towns, weather, and traffic. Run them in parallel. Get back recordings and per-scenario results.
A finished job isn’t proof that every scenario ran, or that any of them passed. 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. Delivered separately, they tell your team exactly what to do next.
Everything you just watched a person do, an agent can do too
NVIDIA publishes reference AI agent skills across its physical AI software, and other tools your team uses are doing the same. Knowing what to try is one thing, but having somewhere to run it is another.
A Template arrives ready on CoreWeave, and ships with CoreWeave’s own agent skills. Every operation is documented and every response machine-readable, so your agent can launch the run itself.
The Template reports what the run produced separately from whether the job finished. So your agent can change one thing, run again, and know whether it helped.
The right GPU depends on the work. For rendering-heavy simulation and video workloads, that’s often NVIDIA RTX PRO 6000 Blackwell Server Edition, with RT Cores for rendering and hardware video encoding. It’s our default. We choose the hardware and configuration for each Template.
for simulation and rendering
for rendering and hardware video encoding where the workload needs it
Straight answers on what’s available, what’s qualified, and what stays yours.
Go deeper with blogs, technical guides, and more.
Tell us which tool and workload you want to run. Your account team will set up a Template on your own data.
CoreWeave Physical AI Launchables sits inside CoreWeave for Physical AI, alongside the infrastructure, the field engineering team, and the partner ecosystem around it.
