Key Notes
- Nvidia introduces GR00T 2 and Turbo as it expands its industrial robotics platform.
- Company tests showed 95% average success across 50 tasks after five demonstrations per task, with code availability listed for December.
- Simulation and evaluation tools aim to reduce development costs and improve robot reliability.
Nvidia (Nasdaq: NVDA) introduced GR00T 2 at the Conference on Robot Learning in Austin on Monday, advancing its effort to turn AI-trained robots into practical tools for factories, warehouses and other businesses.
The announcement combines a new robot foundation model with simulation and evaluation tools designed to make delicate physical work easier to automate. For Nvidia, the commercial opportunity spans the computers used to develop robot skills and the processors that eventually run them inside machines.
GR00T 2 Learns New Skills Faster
GR00T 2 predicts future video and robot actions together, learning how a movement should change the surrounding scene. The approach allows it to draw on human demonstrations and experience from different robots, reducing the amount of training needed on a particular machine.
In Nvidia’s evaluation of 50 tasks new to the target robot, average success rose from 40% without task-specific robot demonstrations to 71% with one demonstration per task and 95% with five. Those are results from the company’s research tests. They measure success on the selected tasks, rather than reliability across an entire factory shift.
Nvidia also presented GR00T 2 Turbo, a smaller version designed for onboard computing. Its research page reports operation on Jetson Thor, connecting the model work directly to Nvidia’s robotics hardware. The page lists code availability for December.
Simulation Targets Robot Training Costs
Isaac Lab 3.0 provides the environment in which developers can train and refine those skills before testing them on physical equipment. Already released in October, the open-source framework supports different physics engines, including Newton, and can model deformable materials such as cables and cloth.
At CoRL, Nvidia reported up to 2.5 times the training throughput in a vision-based manipulation benchmark using Isaac Lab 3.0 with Newton, compared with Isaac Lab 2.3 using PhysX. Faster simulation can shorten development cycles and let engineers test more variations before putting a robot to work.
New benchmarks, DexBench and VABAR, address industrial dexterity and adaptability. Nvidia is also bringing a developer preview of Isaac Lab-Arena 0.4 to the conference, giving researchers tools to evaluate policies under changing conditions and identify where they fail.
Open Tools Support a Hardware Business
Nvidia’s robotics strategy reaches across the development process. Its Isaac expansion at GTC Berlin last month paired models and simulation software with computing platforms including DGX Station and Jetson Thor.
Making development tools widely available can help more robotics teams build around that hardware. The business case depends on those teams moving from research projects to repeat deployments, creating demand for training, simulation and onboard computing along the way.
Warehouse Returns Provide a Practical Test
Ambi Robotics, a partner highlighted in Nvidia’s CoRL announcement, offers one example of how software improvements can translate into operational gains. In a separate September account, the company said an AI-agent workflow developed a fix for a package-placement problem in 10 hours. Production tests showed an improvement of 4.2 packages per hour, equivalent to more than 15,000 additional sorts a year per robot under its stated operating schedule.
Warehouse operators can earn more from equipment they have already bought when software raises throughput. Nvidia’s hardware opportunity grows if those gains persuade customers to deploy more robots and invest in the computers that train and run them.
