NVIDIA reported that its GR00T 2 robot model completed 40 of 50 evaluated tasks at least once without seeing a demonstration specific to those tasks. For people deciding how much training a new job will require, two other numbers in its Conference on Robot Learning announcement are more revealing.
Given one demonstration per task, the model averaged 71 percent success across the 50 tasks in NVIDIA’s evaluation. With five examples, the average rose to 95 percent. The company trained the model on both human and robot data, including 100,000 hours of human activity viewed from the participant’s perspective. The absence of a task-specific example therefore does not mean the model learned without prior experience.
NVIDIA identified laundry handling at Nexuni and server manufacturing at Foxconn among the applications under investigation or development. It did not present the 95 percent result as a measured production-line success rate at either company.
When a first success becomes a useful skill
Counting tasks completed even once tells a reader how broad the robot’s repertoire might be. Counting successful trials tells a prospective operator how often it might need help. A robot that gets a task right once in ten tries would count toward the first measure alongside one that succeeds ten times out of ten. Treating those two robots alike would obscure the operating cost that matters on a shift.
The jump between one and five examples makes the marginal value of another demonstration a concrete question for an operator. Whether those examples cover the awkward cases matters too: a garment bundled differently or a connector partly hidden from view may occupy little of a benchmark yet consume much of a technician’s time.