NVIDIA's New Robot Brain Skips the Teleoperation, Learns from Watching Us

NVIDIA has released GR00T N1.7, an open-source model designed to be the foundational intelligence for humanoid robots. Available now with a commercial license, it’s built for immediate use in industrial settings like assembly lines and warehouses.

The model represents a shift in how robots learn. Instead of relying primarily on thousands of hours of tedious robot teleoperation, GR00T N1.7 was pre-trained on over 20,000 hours of video from cameras worn by humans. The idea is simple: a robot with two arms and a first-person view can learn the basics of manipulation by watching people perform tasks. NVIDIA reports this approach revealed a clear scaling law—more human video data leads to predictable gains in a robot's dexterity, more than doubling task completion rates.

Architecturally, the model separates reasoning from action. One component interprets visual scenes and language instructions to plan steps, while a second, faster system translates those plans into precise, real-time motor commands. This allows a robot to handle multi-step jobs and perform delicate, contact-rich manipulation with individual finger control.

For engineers, it’s a ready-to-deploy tool. It supports several popular robot platforms out of the box and can be fine-tuned for custom hardware. The model is a direct upgrade from the previous version, requiring only a path change in existing deployments. GR00T N1.7 is available for download on Hugging Face and GitHub.

Source: Hugging Face Blog