GR00T 2 Clears 97% Assembly Benchmark

The important advance is not raw dexterity alone—it is recovery. GR00T 2 shows how a reasoning layer can supervise low-level robot skills through repeated observe, act, verify loops.

• –Completed 98 toy-plane, 97 toy-tractor, and 97 Hot Wheels assemblies out of 100
• –Outperformed the same policy without an agent, which achieved 71–84 successes
• –Used code-as-policy recovery skills when parts dropped or progress stalled
• –Began each trial with parts dumped into clutter, testing robustness beyond clean demonstrations
• –The results are promising but remain NVIDIA-reported evaluations on narrow toy tasks, not proof of factory-ready autonomy