The company unveiled the system at the Conference on Robot Learning in Austin, Texas, on Monday, alongside simulation software and evaluation tools intended to improve robotic manipulation. Nvidia says GR00T 2 can predict both how a scene will change and which movements a robot should make, rather than generating motor commands alone.
In company tests covering 50 tasks not previously demonstrated on the target robot, the model achieved an average success rate of 40% without task-specific training. One demonstration increased the figure to 71%, while five demonstrations raised it to 95%, Nvidia said. The latter involved less than three minutes of robot recordings per task and approximately 30 minutes of fine-tuning using one RTX 5090 graphics card.
Those results require qualification. Nvidia’s research materials indicate the 50-task evaluation set was selected from tasks meeting a performance threshold after adaptation. The figures therefore cannot be treated as evidence that the model will achieve comparable accuracy on every unfamiliar factory operation. Independent replication of the reported results has not been established.
GR00T 2 was pretrained using 402,730 hours of first-person video showing people performing activities, 10,184 hours of demonstrations recorded with handheld grippers, and 23,500 hours of robot teleoperation. Nvidia has made 100,000 hours of human video available through its Human-1M-100K dataset, allowing researchers to examine part of the material underpinning its approach.
The technology is intended to transfer skills between different types of robotic hardware. Nvidia demonstrated tasks involving conventional grippers, articulated hands and humanoid machines, including laboratory handling, toy assembly and component insertion. A two-handed robot successfully positioned filter paper in a laboratory funnel and wetted it in 65% of trials after learning from human demonstrations rather than task-specific robot recordings.
Other tests showed a robot inserting GB300 memory modules with 85% success and assembling a toy truck with 65% success after additional training. Nvidia said reinforcement learning performed on physical robots increased success on three precision insertion tasks to between 88% and 96%, highlighting the importance of further practice for contact-sensitive operations.
The company also combined GR00T 2 with a separate reasoning system that breaks complicated jobs into steps, checks progress and initiates recovery procedures. In repeated assembly tests, this arrangement completed 98 of 100 toy-plane assemblies, 97 toy-tractor assemblies and 97 toy-car assemblies. These figures describe a combined system, not the standalone robot model.
A smaller version, GR00T 2 Turbo, is designed to operate on a humanoid robot’s onboard computing hardware. Nvidia said the model generated predictions in 124 milliseconds on a Unitree G1 equipped with Jetson Thor hardware. The humanoid completed all 48 popcorn-serving trials and cleared a table successfully in 93% of trials under the company’s testing conditions.
Nvidia identified several organisations exploring commercial applications. Foxconn is applying GR00T 2 to server manufacturing automation, Siemens is evaluating industrial uses, and Nexuni is testing robotic laundry handling. These activities represent trials or evaluations rather than verified evidence of widespread deployment or commercial returns.
For prospective users, the central questions extend beyond headline success rates. Manufacturers would need to measure performance with their own parts, lighting, workspaces and safety requirements, including the frequency and cost of failed grasps, dropped objects and interrupted production cycles. Testing should also examine how quickly a model adapts when materials, equipment or task sequences change.
Integration with existing control systems, the availability of suitable demonstration data and the computing demands of deployment could determine whether training efficiencies translate into savings. Human supervision and reliable stopping procedures remain important where robots operate near workers or handle delicate components.
Nvidia also reported that increasing the share of human video used during pretraining improved task progress in physical robot evaluations from 50% to 70% across 72 tasks. That comparison supports the value of human footage within its experiments, but does not isolate every factor affecting performance in an actual operational setting.
Nvidia describes GR00T 2 as an open model, although its research page lists code availability for December. The company introduced the system alongside Isaac Lab 3.0 and additional robotics benchmarks at the Austin conference, where researchers are examining how simulation can support more precise manipulation tasks.
Follow Arabian Post
Select Arabian Post as your preferred source on Google and MSN News for trusted business news and Arab politics and updates.
