Key Takeaways

  • Ming-Yu Liu, Vice President of Cosmos Lab at NVIDIA, defines an open model by three components: open weights hosted on Hugging Face, training frameworks on GitHub, and shared datasets.
  • The open release approach spans three NVIDIA lines: Cosmos for physical simulation and world models, Nemotron for language, and Project GR00T for robotics.
  • NVIDIA publishes GitHub recipe cookbooks with each release so developers can post-train base models on proprietary data without rebuilding pipelines.
  • Liu argues that full physical automation requires adaptable robot arms that no single lab can build in isolation.

The Three Parts of an Open Release

Most AI labs drop weights on Hugging Face, issue a press release, and call the work open. The weight file gives you inference, but it leaves you blind to how the model learned. You cannot easily patch failure modes, retrain on custom distributions, or understand data biases when the underlying pipeline remains locked inside a corporate repo.

NVIDIA takes a wider approach across its model portfolio. As Liu explains, “Open model, it's not just have the model open weight allow you to use. We also provide the training framework, so that you can take this open model and your own data, and post-train to something more tailored for your use case.”

This structure covers three main assets: Cosmos for physical simulation, Nemotron for general reasoning, and Project GR00T for humanoid and robotic intelligence. Alongside the weights, NVIDIA posts the code repositories on GitHub, writes technical blogs explaining architectural choices, and adds practical recipe cookbooks. In select cases, the team releases the actual training data. As Liu notes, “We even open-source data that we created to help you to use those data if you want to build your own open model.”

Why Physical Robotics Requires Shared Pipelines

Digital AI models live in software environments where edge cases trigger errors or bad text outputs. Physical AI models control actuators, robot arms, and heavy machines in factories and warehouses. When a physical model fails, it breaks hardware or stops an entire production line.

Getting a robot arm to handle varied, real-world tasks demands high adaptability. Liu points out that no single vendor can map every factory floor, warehouse layout, or mechanical tool alone. “To achieve full automation, we need our robot be smarter, even the robot arm smarter and more, I would say, more configurable,” Liu says. “I think it's difficult for one company to do it alone.”

When developers can inspect both the training pipeline and the recipe cookbooks, they can take a foundation model like Cosmos or GR00T and fine-tune it directly on their own sensor streams and specialized hardware configurations. This shifts the engineering task from starting at zero to running verified post-training routines on top of proven baselines.

What to Do With This

Pick one internal automation bottleneck where off-the-shelf vision or language models fail. Clone the official GitHub repository for Cosmos or Nemotron, open the recipe cookbook, and run a test post-training script against 500 samples of your own proprietary operational data this week.