Photo related to: NVIDIA announces GR00T 2 and Isaac Lab 3.0 at CoRL
Photo: blogs.nvidia.com
Key figure: According to NVIDIA, GR00T 2 achieved 71% average success across 50 tasks with one demonstration per task, rising to 95% with five demonstrations per task.
According to NVIDIA, GR00T 2 achieved 71% average success across 50 tasks with one demonstration per task, rising to 95% with five demonstrations per task. Source: blogs.nvidia.com, 2026-11-09. Source: VERXERA card from blogs.nvidia.com
  • NVDANVIDIA · NASDAQ$178.67▼ 4.27% day

Prices from Twelve Data, captured Nov 9, 5:38 PM ET; not updated after publication. Day change is against the prior close. Not investment advice.

What matters

In its November 9, 2026 announcement, NVIDIA introduced GR00T 2, Isaac Lab 3.0 and open benchmarks at the Conference on Robot Learning. The company reported robot-task performance results and a simulation-training throughput comparison, alongside a choice of physics solvers for researchers.

Verified facts

  • NVIDIA said it was introducing GR00T 2, Isaac Lab 3.0 and open benchmarks at the Conference on Robot Learning, known as CoRL.
  • NVIDIA reported that GR00T 2 completed 40 of 50 evaluated tasks at least once without task-specific demonstrations. That measure records whether a task was completed at least once, rather than a repeated-success rate.
  • According to NVIDIA, GR00T 2 achieved 71% average success across 50 tasks with one demonstration per task, rising to 95% with five demonstrations per task.
  • NVIDIA reported that Isaac Lab 3.0 with Newton achieved up to 2.5 times the training throughput of Isaac Lab 2.3 with PhysX in a vision-based dexterous manipulation benchmark.
  • NVIDIA said Isaac Lab 3.0 allows researchers to switch between the NVIDIA PhysX and Newton physics solvers.

Analysis

The reported measures answer different questions: task completion without demonstrations, average success with demonstrations, and simulation-training throughput. They should not be treated as interchangeable measures of robot reliability. The throughput comparison also changes both the Isaac Lab version and the physics solver; it does not isolate either change’s contribution. Its stated scope is one vision-based dexterous manipulation benchmark, not every training workload.

Editorial opinion

VERXERA opinion: NVIDIA’s November 9, 2026 announcement offers a concrete basis for examining robot learning and simulation together, rather than treating either as a complete performance verdict. The company reports 71% average success across 50 tasks with one demonstration per task and 95% with five, alongside up to 2.5 times the training throughput for Isaac Lab 3.0 with Newton versus Isaac Lab 2.3 with PhysX. That connection matters because task success and training throughput describe distinct dimensions of the announced tools. The separate result of completing 40 of 50 tasks at least once without task-specific demonstrations must retain its narrower meaning. These company-reported results do not establish consistent performance across other tasks or workloads. Deeper source reading is worthwhile for readers whose work matches those evaluation settings, not as a general guarantee of robot capability.

This is VERXERA’s editorial interpretation, separated from the facts and analysis above. The reader keeps the final judgment.

Counterevidence and limits

This report rests on NVIDIA’s announcement; the performance figures and product descriptions are the company’s statements.

Known uncertainties

How consistently GR00T 2 repeats the tasks completed at least once, and how broadly the reported throughput advantage applies beyond the named benchmark, remain open questions.