NVIDIA Trained GR00T 2 on 400,000 Hours of Human Video: Five Demonstrations Take a Robot to 95 Percent Success

At the Conference on Robot Learning in Austin, NVIDIA introduced GR00T 2, a 2-billion-parameter robot foundation model built on Cosmos-Predict2.5-2B that generates expected future camera views and the next action sequence from recent images, robot state and a written instruction. One pretrained checkpoint covers parallel-gripper arms, dexterous hands and humanoids, with only a small per-robot head swapped in for each body. NVIDIA says code will be released in December.

The training recipe is the story. Pretraining combined 402,730 hours of egocentric human video, 10,184 hours of handheld UMI gripper demonstrations spanning 2,897 tasks, and just 23,500 hours of robot teleoperation. NVIDIA is publishing 100,000 hours of the human data as Human-1M-100K under a CC BY 4.0 license that permits commercial use, making it one of the largest public human-manipulation datasets released to date, and it has open-sourced the capture hardware designs.

NVIDIA's ablation explains why the mix matters. Scaling the share of human video in pretraining from zero to 100 percent lifted real-robot task progress on 72 Dexter tasks from 50 to 70 percent. Robot data is expensive and human video is cheap, and the curve shows the cheap data converting into real-robot skill.

All headline results are NVIDIA's own, and the technical whitepaper includes a caveat worth reading twice: the 50 evaluation tasks were selected among those the model can learn to at least 80 percent with five episodes. Within that filter, GR00T 2 succeeded 40 percent of the time zero-shot on tasks the target robot had never practiced, 71 percent with one demonstration per task and 95 percent with five, with 37 of the 50 tasks succeeding on every trial. Each adaptation used under three minutes of robot data and a roughly 30-minute LoRA fine-tune on a single RTX 5090.

Cross-embodiment transfer is the second pitch. Trained only on gloved humans doing chemistry-lab demonstrations, with zero teleoperation, a two-handed robot lined a funnel with filter paper in 65 percent of trials. Shown only handheld UMI gripper footage, a YAM arm assembled a toy dump truck end to end in 60 percent of trials. On conveyor sorting, GR00T 2 sorted 86 percent of items against 28 percent for GR00T 1.7 and zero for the pi0.5 model, and two 22-degree-of-freedom Sharpa hands seated GB300 memory modules 85 percent of the time, per NVIDIA's report.

GR00T 2 is exposed as a tool through an MCP server, so a reasoning model acts as System 2 — decomposing tasks, checking progress and writing code — while GR00T 2 handles the hands-on work. Starting from parts dumped on a table, the pair assembled a toy plane, a tractor and a Hot Wheels car in 98, 97 and 97 of 100 trials, according to the company, which also said Foxconn is putting GR00T 2 to work in server manufacturing automation and Siemens is evaluating industrial uses. Both are company statements rather than independently confirmed deployments.

The release rides a simulation stack that has matured quietly. Isaac Lab 3.0 has been generally available since October 20 with the Newton physics backend supporting rigid bodies, deformables and particles, and NVIDIA reports up to 2.5 times the training throughput of Isaac Lab 2.3 on a vision-based dexterous-manipulation benchmark. On the NIST Assembly Task Board suite, a shared policy trained with only a yes-or-no success reward reached 98.2 percent versus 81.4 percent for separate per-task policies, using about 39 times less GPU time, and DexBench v1.0, built with Korean startup RLWRLD, adds 15 manipulation tasks across retail, logistics, assembly and hotel scenes.

For a field where most labs still hand-collect teleoperation data task by task, a reusable foundation checkpoint plus 100,000 hours of licensable human video is a genuine shift in what a small robotics team can attempt. The filtered benchmark selection and vendor-run evaluation temper the 95 percent headline, but the direction is clear: the bottleneck is moving from collecting demonstrations to converting the internet's existing human video into robot skill.