cuRoboV2 features and Open Source (Apache 2.0)

[balakumar-s]

• Open source (Apache 2.0) repository.
• Bring features from cuRoboV2 research: https://arxiv.org/abs/2603.05493
• Major refactor for cleaner code, breaks most existing API.

---

This discussion was created from the release cuRoboV2 research features.

[sirius-open]

Thanks for your amazing update. Can it be used in isaac-sim? May I ask is there any example program for issac-sim？thank you.

[balakumar-s]

Yes, cuRobo can be used with Isaac Sim.

That said, we do not currently include Isaac Sim examples in the codebase. We are focused on a few core features right now and plan to add Isaac Sim examples in the coming months.

Installation should be easier now because we removed the dependency on PyTorch for the compilation step.

For now, the available interactive examples use Viser. If you are interested in contributing an Isaac Sim example, please start a new discussion and we would be happy to help guide the integration.

[siangfrank]

Thanks for the major V2 update and the Apache 2.0 switch!

We’re currently assessing cuRoboV2 for our robotics stack. Could you clarify how it relates to nvidia-isaac/cumotion? Are they powered by the same backend? Since V2 introduces several breaking changes, knowing the roadmap for these two repos will help us better plan our integration. Thanks!

[balakumar-s]

cuRobo v2 is a research effort. It's Python, Apache 2.0, maintained by a small research team, and open to community contributions. We explore new ideas in fast motion optimization. For example: 1ms multi-camera depth-to-ESDF integration, humanoid-scale planning and retargeting, reactive control, batched trajectory generation, and torque-limited planning.

cuMotion is a product. It's C++, built and supported by a dedicated engineering team inside NVIDIA, and shipped as part of Isaac. It focuses on industrial use cases with the testing and stability that production deployments need.

The two repos do not share a backend. cuRobo v2 is Python built on PyTorch, Warp, and CUDA Python. cuMotion is a separate C++ implementation with its own codebase. That said, several algorithms in cuMotion were invented in cuRobo and hardened in cuMotion, so there is real technical heritage between the two. cuMotion exists because some users need a hardened, well-tested implementation of specific algorithms for production, and a research codebase is not the right place to carry that burden. Building cuMotion as a separate product lets the research side keep moving fast while production users get the stability they need.

Which one should you use?

• cuRobo v2: doing research, working with humanoids, generating training data, or running pilot deployments.
• cuMotion: deploying a robot in production when the features you need are already available in cuMotion. If you want a feature for production that’s not yet available, please raise an issue in the cumotion repository.

Why did the API change from cuRobo v1 to cuRobo v2?

We did not anticipate the widespread use of v1 and focused its code on functionality. With v2, we have simplified the API to make it easier for users and LLMs to work with. This is a onetime refactor.

I'm on cuRobo v1. Should I update to v2 or move to cuMotion?

Both require a migration, so the question is which target fits your use case:

• If you are doing research, working with humanoids, or generating training data, update to cuRobo v2. The new API is simpler than v1 and you keep access to ongoing research features.
• If you are deploying a robot in production and the features you need are already in cuMotion, move to cuMotion. You get the testing, stability, and support that production deployments need.

cuRoboV2 roadmap: We will update the roadmap soon here: #620

[siangfrank]

Thanks for clarifying! Knowing they use separate backends and target different use cases really helps our evaluation. We'll watch #620 for the roadmap updates. Awesome work on this!

[ProNeverFake]

Nice work, congrats!