NVIDIA unveils GR00T 2 world-action model: 40% zero-shot success, human demos transfer to robots

yukez · x · 2026-11-10

NVIDIA GEAR introduces GR00T 2, a new generation of foundation world-action models that jointly predict future video and robot actions—imagining how physical interactions unfold before acting. One model runs on parallel grippers, dexterous hands, and humanoids; code opens in December.

Key capabilities:

X-Adaptability: skills shown only by humans—bare hands, a handheld UMI gripper, or the wearable DexUMI device—transfer to robots that never practiced them. A two-fivel-fingered-hand robot lined a lab funnel with filter paper in 65% of trials; a YAM arm assembled a toy dump truck end-to-end in 60% of trials from UMI demos alone.
Zero-shot & few-shot: 40% success on 50 tasks never seen on the target robot; 5 demos (<3 min of robot data) per task to master new skills.
Data Pyramid: training relies mostly on non-robot data—400k+ hours of egocentric human video, 10k hours of UMI demos, 23.5k hours of teleoperation.

GR00T 2 Turbo runs onboard via Jetson Thor, driving the full body in real time. Through the WAM API, a reasoning model plans and verifies progress while GR00T 2 executes; agents can even steer it by marking grasp targets and drop locations in the camera view.

Original post →

NVIDIA unveils GR00T 2 world-action model: 40% zero-shot success, human demos transfer to robots

More from Embodied