Nvidia GR00T 2 Teaches Robots From Human Videos: Success on 50 Tasks Rises From 40% to 95%

Imagine you need to teach a robotic arm to slice a banana. The standard approach today goes like this: an operator guides the manipulator by hand for hours while the system records every movement. This is teleoperation. It is expensive, slow and tied to a specific robot model: change the hardware, and you have to collect the data all over again. In its GR00T 2 report, Nvidia tries to remove the human operator from this scheme, or at least greatly reduce their involvement.

A Model That Draws Its Own Future

The previous version, GR00T 1, belonged to the VLA (vision-language-action) class. It took a camera image and a text command as input and produced motor signals as output. GR00T 2 is built differently. Nvidia calls it a world-action model, or WAM. It simultaneously computes the next movement and, through diffusion, generates a short video of the near future. If a hand picks up an object, the model "draws" in advance how the frame will change.

This video forecast works as a bridge between human and robot. The model was trained on a cluster of more than a thousand GB300 chips, including roughly 400,000 hours of video in which people perform various tasks with a camera on their heads. To teach the robot something new, a person only needs to demonstrate the steps using a sensor glove, a special handheld gripper or a finger exoskeleton. During training, these recordings are mixed with general motion data from the target robot. Then, in its video forecast, the model draws a robotic hand in place of the human one and derives suitable control commands from that picture.

In simple terms: it watched how a human does it, mentally put itself in their place and worked out which motors to turn.

What the Tests Showed

A robot with five-fingered hands lit a match, sliced a banana and tossed parcels onto a conveyor. There was no teleoperation for these specific tasks. But success fluctuated from 10% to 80% depending on the task, and 10% is still a demo, not a product. The system learned to drive nails from a single video of a human hand, and did so through computer simulation.

The strongest figure concerns fine-tuning. On 50 selected tasks, where the model already coped reasonably well after a few examples, less than three minutes of ordinary training raised success from 40% to 95%. An important caveat: the tasks were chosen from those that were already going well, so this number cannot be extrapolated to any job.

For long assembly operations, Nvidia places a language model above GR00T 2, for example OpenAI's GPT-6 Astra. GR00T handles grasping and movements. The language model breaks the task into steps, marks grasp points on the image and launches auxiliary routines if a part is dropped. The parts of toy models lay on the table in a jumble, and the combination assembled them in at least 97% of attempts. Without the corrective language model and fine-tuning, assembling a toy car succeeded only 23 times out of 100.

For bipedal robots, there is a lightweight version, GR00T 2 Turbo, which processes fewer image details. It runs directly on the humanoid's onboard Jetson Thor computer. Such a robot scooped popcorn, cleared tables and wiped up liquid with a towel under its foot. The last task succeeded in slightly fewer than half of attempts.

Where It Still Doesn't Work

The authors themselves list the barriers to reliable real-world operation. The model does not yet process raw data from tactile sensors and sometimes misses subtle errors, such as a screw that is not fully tightened. The language-model planner slows the system considerably. When sorting trash, the robot sat idle about a third of the time: on average it waited more than seven seconds for the planner's next calculation.

What This Means for Practitioners

The main shift is not in records but in the economics of data. If videos of people and cheap gloves replace part of teleoperation, the main cost item in robotics stops being tied to specific hardware. For founders, this is a reason to calculate the cost of collecting demonstrations in a new way.

For developers, it is a reason to look closely at the "action model plus separate LLM planner" scheme and to build the planner's latency into the speed budget from the start. Product managers should remember: the 10–80% figures and the 23% without a planner are laboratory results. Reliable scenarios on a factory floor or in a warehouse are still far off, especially where force and touch must be sensed.