NVIDIA Releases Robot Foundation Model GR00T 2, Plans to Open Source in December
At CoRL, NVIDIA released its next-generation robot foundation model GR00T 2, which uses a world-action model architecture to simultaneously predict future frames and robot actions. Its pretraining used approximately 403,000 hours of first-person human video, and NVIDIA plans to open the model weights, code, and adaptation solutions in December. The model achieves cross-embodiment skill transfer through X-Adapt, reaching a 40% zero-shot success rate on 50 new tasks and 95% after five demonstrations, requiring less than 3 minutes of robot data. The article also notes that Anthropic's Claude Fable 5.2 and OpenAI's GPT-6 Astra are extending toward robot manipulation, with general-purpose models and open-source foundation models squeezing embodied model startups from both sides, while engineering capabilities such as precision manipulation, tactile feedback, and real-time response remain their moat.
Why read this
SelectedAI score 78
The article reviews the specific data behind GR00T 2's training on human video and cross-embodiment transfer, and compares it with concurrent progress in general-purpose large models, presenting the two-sided pressure facing embodied AI startups.
Topics
Body · AI translation
[[**NVIDIA releases the GR00T 2 robot foundation model, which can learn from human videos and cross-robot data and is planned to be open-sourced; general-purpose large models are simultaneously extending into robotics, leaving embodied-model startups caught in a pincer attack, while real-world engineering capability remains a barrier.**
**Key points**
**1. GR00T 2 achieves joint prediction of actions and environment**
It adopts a world action model that simultaneously predicts future frames and robot actions, enabling human videos to be used for training; pretraining includes 403,000 hours of human video, and progress on 72 real-robot tasks rises from 50% to 70%.
**2. Cross-embodiment transfer and few-shot adaptation lower the data threshold**
X-Adapt lets robots learn from human or other-robot demonstrations, achieving a 65% success rate in chemistry experiments; zero-shot success on 50 new tasks is 40%, reaching 95% with five demonstrations, with less than 3 minutes of data required.
**3. Hierarchical control and the Turbo version support complex tasks and real-time control**
The upper-layer reasoning model handles planning and checking, while GR00T 2 executes actions and recovers from anomalies; in continuous assembly of three types of toys, 97 to 98 of 100 attempts succeed per type, and Turbo's single action prediction takes about 124 milliseconds.
**4. General large models and open-source foundation models squeeze embodied startups from both sides**
Anthropic says Claude Fable 5.2 scores 64.9% on RoboDojo dual-arm tasks, while GPT-6 Astra scores 74.8% on RoboLab-120 but takes 18.56 seconds per inference; engineering capabilities such as precision manipulation, touch, and real-time response remain barriers for startups. Original 2026-11-16 20:03 NVIDIA's summer, winter for embodied-model companies? Yi Yuqin Quick overview
Last week, everyone was busy consuming the "big gossip" about Jensen Huang, so much so that many people missed an important NVIDIA technology update.
Just three days before Jensen Huang resigned, on November 9, NVIDIA released its new-generation robot foundation model GR00T 2 at CoRL, the robot learning conference in Austin, USA.
At this release, Jim Fan, NVIDIA's head of robot research, first joked with the phrase "winter is coming" about the shock that general reasoning models such as GPT-6 Astra have brought to robotics researchers: when general large models can directly control robots, is it still necessary to develop robot-specific brains?
However, at the end of the talk, Jim Fan changed tack and declared that "summer is here."
What made him so optimistic was precisely GR00T 2, which NVIDIA had just launched.
This robot foundation model can learn skills from human videos and operation data from other robots, and is planned to be open-sourced. Currently, NVIDIA has released 100,000 hours of first-person human video data, and plans to open the model weights, code, and adaptation solutions in December.
So, what exactly makes this robot brain so strong?
GR00T 2: robots begin to predict the consequences of actions
Compared with previous versions, GR00T 2's core advances are mainly reflected in three aspects:
Upgraded learning method: from predicting robot actions to simultaneously predicting actions and environmental changes, making massive amounts of human video usable as training material.
Cross-embodiment skill transfer: robots can learn manipulation skills from demonstrations by humans or other robots, reducing the need to repeatedly collect data for different hardware.
Faster adaptation to new tasks: quickly learn new tasks from a small number of demonstrations, and combine with an upper-layer reasoning model to complete complex operations and error recovery.
First is the change in model architecture.
Traditional vision-language-action models (VLA) mainly predict actions based on camera images and language instructions. GR00T 2 instead adopts a World Action Model (WAM) architecture, using a diffusion Transformer to simultaneously predict future frames and robot actions.
For example, when a robot inserts a part into a slot, the model must predict both how the robotic arm will move and what changes will occur in the environment after the part contacts the slot.
This design allows robots to learn manipulation experience from ordinary human videos, reducing reliance on teleoperation data that carries control information.
According to data disclosed by NVIDIA, GR00T 2 pretraining used about 403,000 hours of first-person human video, 10,000 hours of handheld gripper demonstrations, and 23,500 hours of robot teleoperation data.
In a set of controlled experiments, as the scale of human video training expanded, the model's progress metric on 72 real-robot tasks rose from 50% to 70%. This also provides further technical validation for the previously much-discussed Ego data.
Second is cross-embodiment skill transfer.
NVIDIA calls this capability X-Adapt. Simply put, a robot can use demonstrations by humans or other robots to learn tasks it has never performed itself.
For example, in a chemistry experiment, a robot that had not received teleoperation training for that task could, using only human demonstrations, complete operations such as setting up an experimental funnel and wetting filter paper with a 65% success rate.
In another experiment, a YAM robotic arm, using only handheld gripper demonstrations, could assemble a toy truck with a 60% success rate.
Of course, the target robot still needs a certain amount of general motion data so that the model can adapt to its own hardware structure.
A more intuitive advance is reflected in the efficiency of learning new tasks.
NVIDIA tested it on 50 target robot tasks that the robot had never performed before. GR00T 2 achieved an average zero-shot success rate of 40%; after adding one demonstration per task and fine-tuning, the success rate rose to 71%; after increasing to five demonstrations, the average success rate reached 95%.
The robot data required for these five demonstrations was less than 3 minutes, and lightweight fine-tuning of the model on a single RTX 5090 GPU took about 30 minutes.
In addition to fast learning, GR00T 2 also attempts to solve the problem of error recovery in complex tasks.
NVIDIA adopts a hierarchical control architecture: the upper-level reasoning model is responsible for planning tasks and checking progress, while GR00T 2 is responsible for specific actions such as grasping and insertion. Once a part is dropped or a task is interrupted, the upper-level model can call other tools to handle the exception and then continue execution.
In continuous assembly tests of three types of toys, after adding this mechanism, the robot succeeded 97 to 98 times out of 100 attempts for each type, whereas when the action model was used alone, the number of successes was 71 to 84.
To adapt to the real-time control requirements of humanoid robots, NVIDIA also launched a lightweight version, GR00T 2 Turbo. It can run directly on the Jetson Thor carried by the Unitree G1, with a single action prediction taking about 124 milliseconds, which is then executed by a high-frequency whole-body controller.
Overall, GR00T 2 is integrating human video learning, cross-robot skill transfer, few-shot adaptation, and complex task control into a single model system.
However, the above results mainly come from specific experiments published by NVIDIA, and there is still a gap before long-term stable operation in complex industrial scenarios.
If these capabilities can be verified in more real-world scenarios, robot manufacturers will have the opportunity in the future to directly use mature foundation models and then optimize them for their own hardware and tasks.
This will also bring new competitive pressure to startups that train general-purpose robot brains from scratch.
Embodied model companies are being squeezed from both sides by tech giants
Dramatically, on the same day NVIDIA released GR00T 2, Anthropic also launched a new-generation general large model, Claude Fable 5.2.
This large model, mainly aimed at tasks such as programming and reasoning, has already begun to demonstrate robot manipulation capabilities.
In a video released by Anthropic, Fable 5.2 controlled a real robot to complete tasks such as manipulating a Rubik's cube, unscrewing a bottle cap, and painting.
More noteworthy are its test results.
In the RoboDojo simulation benchmark, which includes 42 dual-arm robot tasks, Fable 5.2 achieved a score of 64.9%, higher than the 27.6% of the previous generation Fable 5.1 and the 29.8% of GPT-6 Astra.
Anthropic said this performance has surpassed most specially trained robot control models. However, the relevant results come from Anthropic's internal tests, and the real-robot demonstrations have not yet disclosed systematic success rates.
In fact, OpenAI's GPT-6 Astra had previously shown similar potential.
In October this year, a study released by NVIDIA showed that Astra, with the help of an agent framework, achieved a 74.8% task success rate in the RoboLab-120 simulation benchmark, surpassing several specialized robot models included in the comparison.
But its shortcomings are also obvious: a single action inference takes an average of 18.56 seconds, which is still a considerable distance from the requirements of real-time robot control.
English text is being prepared.
On November 13, according to Fortune, OpenAI has disbanded its independent robot research team and merged the relevant researchers into the core model team. OpenAI confirmed the adjustment and said robotics remains its long-term development direction.
A series of changes show that general large models are accelerating their extension into the robotics field.
So what is their relationship with GR00T 2?
From a technical route perspective, GR00T 2 specifically learns the relationship between video and robot actions, focusing on solving problems such as skill transfer, fine manipulation, and real-time control.
General models such as GPT-6 Astra and Fable 5.2, relying on visual understanding, spatial reasoning, and task planning capabilities, have begun to participate in robot action decision-making.
The two types of models both compete and can cooperate with each other.
For example, in the hierarchical control architecture demonstrated by NVIDIA, the upper-level general reasoning model is responsible for planning tasks and checking results, GR00T 2 is responsible for specific actions, and then the underlying controller executes them.
This means that in the future the robot's "brain" may be further divided: general models will take on more understanding and decision-making tasks, while specialized models will be responsible for physical operations that require rapid response.
For embodied model startups, pressure is coming from two directions.
On the one hand, NVIDIA continues to open up robot foundation models and development tools, reducing the need for complete-machine manufacturers to train models from scratch; on the other hand, general model vendors such as OpenAI and Anthropic are continuously expanding robot capabilities, which may further compress the technical space for specialized models.
Of course, the progress of general models does not mean that embodied model startups have lost their room to operate.
At this stage, neither GR00T 2 nor Fable 5.2 has yet proven that it can operate stably over the long term in complex industrial scenarios. Precision manipulation, tactile feedback, contact force control, real-time response, and exception recovery still require a great deal of engineering optimization.
This means that companies with proprietary scenario data, mature hardware adaptation capabilities, and real customer deployment experience still have the opportunity to build a competitive advantage.
However, as foundation models continue to advance, the technical capabilities that are scarce today may become industry standard tomorrow. Tech giants are raising the technical starting point of embodied intelligence, and also raising the threshold for startups to prove their own value.
As for who will usher in summer and who will face winter, the real market will ultimately give the answer.
Article title: NVIDIA's summer, the winter of embodied model companies?
19 articles published on Huxiu. This content may not be reproduced without permission. For authorization, please contact hezuo@huxiu.com. Everyone who is changing and wants to change the world is on the Huxiu APP[ ]