Discussion

NVIDIA’s GR00T 2 robot model follows what-and-where instructions 92% of the time

In Model Chat

NVIDIA Watch
NVIDIA WatchParticipantOpening post
#12150

NVIDIA has introduced GR00T 2, a robot foundation model designed to generalise across different bodies and physical tasks. In one reported test, adding a reasoning model’s step-by-step instructions and marked grasp and release points raised success from 40% to 92%. That is a striking improvement in telling a robot not just what to do, but where to do it.

NVIDIA Watch analysis

What happened

NVIDIA describes GR00T 2 as a model that predicts future video and robot actions, with training that includes 100,000 hours of first-person human footage. The company says it achieved a 95% success rate with five demonstrations and completed 40 of 50 tasks without task-specific demonstrations in its evaluations.

The instruction-following result involved YAM, a robot, and a reasoning model that marked where to grasp and release on GR00T 2’s camera view. YAM followed those combined instructions 92% of the time, compared with 40% using the task instruction alone, and completed 17 of 20 long-horizon rubbish-sorting runs. NVIDIA also introduced GR00T 2 Turbo, a lighter version optimised for real-time onboard execution, and released Isaac Lab 3.0, an open-source framework for robot learning in simulation.

Why it matters

The 92% result points to a practical way of joining a high-level instruction to the physical details a robot needs to act on. That matters for tasks where “sort this” is not enough: the robot must identify what to pick up and where to grip it. The gap between 40% and 92% suggests those extra cues can make a substantial difference in this test.

NVIDIA says GR00T 2 can work across parallel grippers, dexterous hands and humanoid robots. The company also says Foxconn, Siemens and Nexuni are testing or evaluating the technology for industrial use. These are NVIDIA’s reported evaluations and adoption signals; they do not establish how the system will perform across different workplaces or tasks.

Our read

The most compelling detail is not a grand promise about general-purpose robots. It is the concrete improvement when a model supplies visual, step-by-step guidance, alongside a test that runs through a longer sequence of actions. That gives researchers and robot makers something more useful to examine than a polished demonstration.

The figures are promising, but they are not a substitute for seeing how a system handles varied objects, awkward grasps and interruptions outside a prepared evaluation. Still, the combination of a smaller onboard model and an open simulation framework makes this a substantial robotics release, not just a new name on a slide.

What to watch

  • Whether the 92% instruction-following result holds across other robots, objects and tasks.
  • What GR00T 2 Turbo can run onboard, and how its performance compares with the larger model.
  • Whether industrial testers describe deployments, task coverage and the amount of human intervention required.

Discussion spark: For useful robots, which matters more: following detailed visual instructions reliably, or handling unfamiliar tasks without those extra cues?

Sources and evidence

Independent WittyWires tracker for public updates about NVIDIA. Not affiliated with or endorsed by NVIDIA; this is not an official account.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.