NVIDIA presents GR00T 2, which it describes as its most advanced “System One” model and as an “open” model for robotics research. The model learns movements from human demonstrations and reproduces them on robots that have never practised them.
What’s happening
According to NVIDIA’s research page, GR00T 2 predicts both future video frames and the robot’s actions. The company writes: “Learning to predict the world, not just act in it”. The model runs on parallel grippers, dexterous hands and humanoids.
In one example, a two-handed robot with five-fingered hands fills a laboratory funnel with filter paper, without ever having been teleoperated for that task. For the toy truck, the YAM arm learns from a handheld gripper different from its own. NVIDIA says this task, lasting up to 4 minutes, succeeds in 60% of trials.
NVIDIA also presents GR00T 2 Turbo, a lighter version built for speed. It controls the robot’s whole body while walking, moving and grasping, through the SONIC controller. According to the company, it prepares and serves popcorn in the 48 announced trials.
Key figures
- 40% success without additional training on 50 tasks never seen on the target robot.
- 95% after five demonstrations per task and a 30-minute fine-tuning on a single RTX 5090.
- 85% for placing GB300 memory modules, 65% for assembling the toy truck, with two Sharpa hands offering 22 degrees of freedom.
- 88 to 96% success on three insertion tasks requiring contact, after reinforcement learning on the real robot.
- 92% compliance with “what and where” instructions, against 40% with the task instruction alone.
- At least 97 successful assemblies out of 100, for each of three chained toys.
In context
NVIDIA’s page says human videos make up the largest share of the training data. Testing four models with 0%, 1%, 10% and 100% of these videos, the company measures a gain from 64% to 77% in video simulation, and from 50% to 70% on the real robot. It adds that 100,000 hours of egocentric videos from its Human-1M dataset are available as open access.
In November we covered NVIDIA’s five announcements in South Korea, including a section on physical AI. On November 8 we also looked at the debate around humanoid robots (article in French).
What comes next
The release gives no availability date for GR00T 2 or Turbo. NVIDIA talks about an “open” model without specifying the licence or terms of use, which still have to be published.
No price is announced. The release also does not say which hardware runs Turbo on board the robot, or at what speed. The success rates come from NVIDIA’s own trials, whose number of repetitions varies from task to task (20 trials for the chemistry tasks, 100 for the toy assemblies). Tests by outside teams are still to come.
As soon as more information is published, you will find it on iadudimanche.fr and on the AI Sunday LinkedIn page.
This news brief was written automatically by an AI system, for greater responsiveness and to bring you ever fresher news.
Sources
- NVIDIA, GR00T 2 video announcement: https://www.youtube.com/shorts/IOcODanTpN0
- NVIDIA, “GR00T 2: Scaling Robot Foundation Models for Generalization, Dexterity, and Adaptation”: https://research.nvidia.com/labs/gear/gr00t-2
- NVIDIA, news page: https://nvidianews.nvidia.com/
À découvrir aussi sur AI Sunday
Chaîne YouTubeL'IA au travail, en vidéoDernière vidéo : Notre place demain avec l'IA : on doit s'inquiéter ?Voir la chaîne Prof de langues IAParle anglais ou portugais avec un prof IABientôt : un prof vocal qui t'explique en français et te corrige.Rejoindre la liste 

Leave a Reply