Nvidia’s GR00T 2 lifts selected robot tasks from 40% to 95%

By Rebecca Stone

Nvidia reports that its GR00T 2 AI foundation model raised success on 50 selected robot tasks from 40% to 95% after less than three minutes of conventional training. The result came after the model had already performed well on those tasks from a few examples.

The result addresses a costly bottleneck in robot training. Nvidia says engineers normally spend hours remotely controlling machines through teleoperation, a process that is expensive, slow, and requires separate data for every robot model. GR00T 2 seeks to replace some of that labor with human movement transferred from video.

How does GR00T 2 reduce teleoperation?

Unlike prior approaches, including Nvidia’s predecessor GR00T 1, GR00T 2 does not directly translate camera images and language commands into motor movements. Nvidia introduced the model in a new research report as a way to shorten the training process for new robot tasks.

Nvidia calls the design a "world-action model." It calculates the next movement and a short video of the immediate future at the same time. In a picking example, the predicted video shows how the camera image should change when a robotic arm takes an object.

Nvidia says the paired prediction should let the model transfer movement sequences between different robot form factors. The model can generate what the target robot should do next while producing the control commands for that action, rather than treating each body as a separate training problem.

The approach differs from a pipeline that selects a motor command and stops there. GR00T 2 first predicts an immediate visual outcome, then derives the command sequence needed to make that outcome occur.

Where did the training data come from?

Nvidia trained GR00T 2 on a computing cluster containing more than 1,000 GB300 chips. The training set included about 400,000 hours of video showing people performing varied tasks while wearing head-mounted cameras. The footage supplies general human motion rather than robot-specific demonstrations.

To add a new task, researchers say a person need only demonstrate the steps through data gloves, specialized hand grippers, or a finger exoskeleton. Those recordings are joined during training with general motion data from the target robot. GR00T 2 then generates the robotic arm in place of the human arm in its future-video prediction.

From that prediction, the system derives matching control commands. The approach lets Nvidia test learned behavior on different kinds of robots and seek a shorter route from human demonstration to machine action.

The human demonstration supplies the task sequence, while motion data from the target robot informs how the action should map to that machine. Nvidia combines both during training rather than asking a human recording alone to produce a command that fits every robot.

What did the robots learn?

Nvidia’s lab tests covered manipulation and a task learned from one hand movement:

• A robot with five-fingered hands lit a match, sliced a banana, and tossed packages onto a conveyor belt without teleoperation for those specific tasks.
• Success rates ranged from 10% to 80%, depending on the task. Nvidia reported the broad range rather than a single aggregate score.
• From a single video of a human hand movement, GR00T 2 learned to hammer nails through computer simulation.
• On 50 selected tasks that GR00T 2 already handled well after a few examples, less than three minutes of conventional training lifted success from 40% to 95%.

Nvidia separated the headline gain from the model’s broader performance. The 95% result came from selected tasks the model already handled well after a few examples; rates across the reported tests ranged from 10% to 80%.

How does the language model change assembly?

For longer assembly work, Nvidia paired GR00T 2 with a higher-level AI language model, identified in the source as OpenAI’s GPT-6 Astra. GR00T 2 controls the gripping movements. The language model divides the job into individual steps and color-codes grip points in the camera image.

It also launches helper routines when a part falls off. In the researchers’ test, loose parts for toy models sat on a table in no particular order. The combined system assembled them in at least 97% of trials.

Without the corrective language model and additional fine-tuning, the toy car assembly succeeded in only 23 of 100 attempts. The team also trained precision tasks such as plugging in network cables or chargers through repeated trial and error on the real device.

In the toy test, the parts’ random placement made step order and grip selection central problems. The language model supplied the high-level corrections, while GR00T 2 translated the planned actions into gripper movement.

The gap shows that corrective planning and additional fine-tuning together mattered in the tested setup. It also shows why Nvidia treated the language model as a separate layer from the motion system.

What does GR00T 2 Turbo change?

For bipedal robots, Nvidia developed GR00T 2 Turbo, a slimmed-down variant that processes fewer image details. It runs directly on the Jetson Thor onboard computer of a humanoid robot.

In Nvidia’s tests, the humanoid machine scooped popcorn, cleared tables, and wiped up liquids with a towel under its foot. The towel task succeeded in just under half of all attempts.

What still blocks reliable real-world use?

The research team says significant hurdles remain before the system can handle reliable real-world work. GR00T 2 does not yet process direct sensor data for touch, and it sometimes misses subtle errors such as half-turned screws. Those gaps matter because the robot must recognize both contact and imperfect assembly states.

The research team also reported that the higher-level language model slows the robot considerably. During trash sorting, the machine stood idle for roughly one-third of the total run because it waited more than seven seconds on average for the planner’s next computation.

Nvidia’s figures therefore show two different kinds of progress in one system: video-based training can raise the success of learned motions quickly, while multi-step language planning can leave the robot idle during trash sorting.

Nvidia also separates motion competence from system-level reliability. GR00T 2 can perform a grip sequence after a short demonstration, but the language layer can wait on a computation before the next action and does not supply the missing touch channel.

Where does the research effort stand?

The Decoder reports that other laboratories are also seeking ways to teach robots new tasks with less teleoperation. Its AI Radar #5 covered efforts from Neolabs. Nvidia’s work extends that research line by using future-video prediction to connect human demonstrations with control commands for another robot body.

The work does not remove all robot-specific data. It shifts part of the demonstration burden from directly driving a machine to recording human motion, while still combining that recording with general motion data from the target robot during training.

The immediate stakes are those Nvidia itself identifies: faster planning, direct touch processing and reliable detection of subtle errors. The team has not yet shown that GR00T 2 can meet those conditions for dependable real-world use.

Until those gaps close, the 95% result after less than three minutes of extra training remains a laboratory result rather than evidence that the system can work reliably in the field.