Training a humanoid with NVIDIA’s GR00T starts before the model sees its first demonstration. In a July 7 technical blog post, NVIDIA’s Edith Llontop and Brandon Neel split the work into five stages, each with its own tool: Isaac Lab-Arena to build a simulated scene, Isaac Teleop to record demonstrations, the GR00T 1.7 model and its training scripts to train a policy, Isaac Lab-Arena again to evaluate it, and Isaac ROS with a Jetson Thor computer to run it on a robot.

NVIDIA says the platform is modular, so teams can use one component or the whole pipeline. That is why a company saying it uses NVIDIA tells you little on its own. It can mean a simulator, a teleoperation tool, an onboard computer or the GR00T model producing the robot’s movements.

What the walkthrough asks of a developer

The post includes a worked example, and it is more useful than the diagram because it shows where the human effort goes. The task is small: a simulated humanoid stands at a shelf and moves an apple onto a plate. Before any data is collected, the developer has to pick a whole-body controller to keep the robot balanced. The authors warn that this choice shapes the training data itself, because the controller produces the joint targets the policy learns from.

  • Demonstrations: the authors recorded 400 trajectories by teleoperating the robot through a VR headset, collected “not all at once but stitched together through multiple sessions of data collection.”
  • Conversion: the recordings are saved as HDF5 files and converted to the LeRobot format GR00T reads, using a config file that maps the simulator’s joint, action and camera fields to what the model expects.
  • Training: the example fine-tunes the 3-billion-parameter base model for 20,000 steps on one GPU, keeping the language model frozen and tuning the vision backbone, projector and diffusion action head.
  • Evaluation: the trained policy is served from a separate model server and runs in closed loop inside the simulator.

What it does not show

Everything in the example happens in simulation. The authors write that “the same steps can be applied to physical robots and adapted across different embodiments and tasks,” but the post does not report a hardware run. It also gives no figure for how long the demonstrations, conversion or training took. Those are the numbers a team planning a project would want first.

The partner list needs the same care. NVIDIA names 1X, Agility, ANYBotics, NEURA Robotics, Schaeffler, Skild AI and others as “integrating components of the GR00T platform such as Isaac Teleop, Sim, Lab, and ROS.” That describes tools in a pipeline. It does not say which of those companies run a GR00T model on their robots.

What the timing table measures

NVIDIA’s GR00T repository gives a median training-step time of 115.5 milliseconds on an H100 80GB GPU at batch size 1. That test freezes the language and vision backbones and excludes data loading. It is a local optimizer-step measurement, not the time needed to teach a robot a task.

The walkthrough runs for 20,000 training steps after demonstration collection and conversion. A project estimate needs that whole sequence. The timing table, checked on November 6, does not include the time spent teleoperating the robot or preparing its data.

A preconfigured research robot

NVIDIA’s October 21 announcement describes Agile ONE SN as a physical research platform with GR00T preconfigured and Jetson Thor onboard. Availability is planned for early 2027. That is a separate hardware announcement; the July walkthrough still reports a simulation example.

Correction: an earlier article, now merged into this guide, repeated an unsupported GR00T 2.0 launch and a large training-time reduction. Neither claim appears in the NVIDIA documents reviewed here. The documented model is GR00T N1.7.