The biggest change in GR00T N1.7 is what it learned from. When NVIDIA released the model as an early-access preview on April 17, the announcement on Hugging Face by NVIDIA’s Edith Llontop and Kalyan Vadrevu framed it around one bet: “human data is the most scalable source of robot intelligence.”

N1.7 was pretrained on EgoScale, which the post describes as 20,854 hours of human egocentric video across more than 20 task categories, from manufacturing and retail to healthcare and homes. The video comes from people wearing head and wrist cameras, with hand tracking. The authors call it “a significant step up from the few thousand hours of robot teleoperation data used to train N1.6.”

The scaling claim

NVIDIA’s headline finding is what it calls the first scaling law for robot dexterity: more human video gives predictable gains, and going “from 1k to 20k hours more than doubles average task completion.” The Hugging Face post does not list the tasks or trial counts behind that figure, so for now it is a company result to check, not a settled one.

The mechanism is spelled out in the GitHub repository. N1.7 represents actions as changes in hand position relative to the current pose, and uses the same representation for human and robot data. NVIDIA says that shared format is what lets lessons from human video carry over to robot control, calling it “a key factor in the model’s cross-embodiment performance.”

The repository is more measured about the overall gain than the launch post. It says N1.7 “delivers comparable performance to N1.6,” with improved generalization and language following that it credits to the human video.

Robot data is still required

Human video gives the model priors. It tells the model nothing about a particular robot’s cameras, joints or grippers. Step one of the repository’s workflow is still to collect robot demonstrations (video, state and action) and convert them to GR00T’s format. According to the Hugging Face post, NVIDIA validated N1.7 on three machines: the Unitree G1, a bimanual YAM manipulator and the AGIBot Genie 1.

For a team adapting the model, the old work remains. Define the task, place cameras consistently, record clean demonstrations and keep some object positions out of training so there is something honest to test on. NVIDIA’s own end-to-end walkthrough warns that “noisy or inconsistent data limits how well the policy can learn.”

In its October 21 release, NVIDIA added a Human 1M Dataset with synchronized three-view video and 3D hand tracking. It is on Hugging Face, with the full set limited to Cosmos Coalition partners. If the scaling law holds, more human video should keep raising dexterity on robots outside NVIDIA’s lab. Results on other companies’ hardware, with task lists and trial counts, would settle it.