The strongest number in NVIDIA's robot announcement is 95%. GR00T 2, the world-action model NVIDIA introduced at the Conference on Robot Learning in Austin on November 9, averages 95% success on 50 dexterous tasks after five demonstrations of each. Read quickly, that sounds like a robot close to solving new work on its own.

The model's own whitepaper supplies the condition, in its own words. The 50 tasks, called Dexter-50, have no training data from the target robot — a bimanual machine with two 22-degree-of-freedom Sharpa hands; they appear only in data from a different robot or in human video. The paper then defines the set, and concedes the consequence: the tasks "were chosen among those the model learns with five episodes (at least 80% success), so the five-episode rate is high partly by construction." The whitepaper discloses the condition. The blog post's headline figures omit it. Neither fact makes the 95% false; together they make it a statement about a selected set, and NVIDIA does not say how large the pool was that the 50 were drawn from.

Three numbers, and they are not interchangeable

Zero-shot, GR00T 2 averages 40% success across the Dexter-50 tasks. Separately, it completes 40 of the 50 tasks at least once without task-specific demonstrations — a different measure: completing a task once in a run of trials is not a success rate. And the 95% figure only arrives after adaptation: one demonstration per task lifts the average to 71%, five lift it to 95%.

The distribution underneath the average is worth the space. Zero-shot, the model completes 40 of the 50 tasks at least once and succeeds in every evaluated trial on four of them. With five demonstrations, 37 of the 50 reach 100% measured success — a finite-trial result, not proof those tasks never fail. The paper's own failure analysis says that in most zero-shot failures the model performs the right steps in the right order and its new hands fall short on one of them, on precision. A little robot data closes that gap, which is a plausible story — on tasks chosen because a little data closes the gap.

The transfer underneath the headline

Strip the selected benchmark away and the paper reports additional cross-embodiment results. Shown a chemistry-lab task only by people, never on the robot, GR00T 2 lines a funnel with filter paper and wets it in 65% of trials. Shown a toy dump-truck assembly only through a handheld gripper — with 41 to 75 handheld demonstration episodes per task — a different robot arm completes it end to end in 60% of trials. Scaling the human video in pretraining from none to the full mixture raised real-robot task progress on 72 other tasks from 50% to 70% — progress, a graded measure, not binary success. Those are NVIDIA's own measurements, with no independent evaluation yet, but they are not conditioned on a five-demonstration screen in the same way.

The near-perfect numbers elsewhere in the release need the same care. In trials where parts are dumped onto the table before each run, supervised fine-tuning completes the three toy assemblies in 69, 51 and 23 of 100 trials. With a reasoning agent supervising GR00T 2 — sequencing skills, watching for dropped parts, repairing the scene and handing control back — the same assemblies reach 98, 97 and 97 of 100. These results evaluate the model together with agentic supervision and recovery, so they cannot be presented as standalone-policy success rates.

What "open" buys today

NVIDIA calls GR00T 2 an open frontier model, and parts of the release are genuinely open: Isaac Lab 3.0 is open source, Isaac Lab-Arena's 0.4 developer preview arrives at CoRL, and 100,000 hours of the human training data are on Hugging Face. For GR00T 2 itself, the November 9 blog post, research page and whitepaper give no download date, no price and no production timeline. The named customers are at the testing stage: Nexuni is testing laundry handling, Foxconn is applying the model to server-manufacturing automation, Siemens is evaluating industrial uses.

Isaac Lab 3.0's speed claim needs the same treatment. NVIDIA reports up to 2.5 times the training throughput for Isaac Lab 3.0 with Newton versus Isaac Lab 2.3 with PhysX, on one vision-based dexterous-manipulation benchmark. That changes both the software version and the physics engine; it is not an isolated version-speed comparison, and not a general speedup for every workload.

For a robotics team, the practical read is narrow and useful. The adaptation procedure is fully specified: five demonstrations, under three minutes of data, about half an hour of LoRA fine-tuning on one RTX 5090, the same settings for every task and no per-task tuning. A team that can reproduce that procedure can assess transfer on its own hardware — once the model is actually available to download, which NVIDIA has not yet dated. What the launch does not support is planning around the 95% as a field reliability figure. It was measured on tasks selected to respond to exactly the treatment being measured.

Sources

  1. [1] NVIDIA Blog, "Practice Makes Precise: At CoRL, NVIDIA and Partners Reveal How Robots Rehearse Delicate Work in Simulation", November 9, 2026 — primary source for the launch, the 40-of-50 at-least-once line, the 71% / 95% averages, partner testing and the Isaac Lab 3.0 with Newton throughput comparison; read first-hand by the deskRead source
  2. [2] NVIDIA Research, "GR00T 2: Scaling Robot Foundation Models for Generalization, Dexterity, and Adaptation", project page, November 2026 — primary source for the architecture, pretraining mixture and cross-embodiment resultsRead source
  3. [3] NVIDIA, GR00T 2 whitepaper (PDF, 135 pages) — Figure 2 caption and Section 3.1 for the five-episode, at-least-80% task-selection definition and the "high partly by construction" acknowledgment; Appendix A.3 and Figure 24 for the 40% / 71% / 95% averages, the per-task distribution and the LoRA protocol; the desk verified the paper's text directlyRead source
  4. [4] The Robot Report, "NVIDIA launches GR00T 2 and Isaac Lab 3.0 at CoRL", November 9, 2026 — independent trade coverage confirming the launch and the Isaac Lab-Arena 0.4 developer previewRead source
  5. [5] Unite.AI, "NVIDIA Introduces GR00T 2 for Robots Learning From Human Video", November 9, 2026 — secondary read of the whitepaper; its report of a December code date could not be confirmed in NVIDIA's own material and is not used in the articleRead source