Key Summary
- 71% with one demo, 95% with five
- Success rate varies with number of demonstrations
- Comparing in-house tasks to DexBench's 15 tasks
- Start by listing processes where precision manipulation is the bottleneck
NVIDIA's second-generation robot AI model, 'GR00T 2,' achieved an average success rate of 71% across 50 tasks when given a single demonstration per task, rising to 95% when demonstrations were increased to five. These figures were disclosed by NVIDIA on its official blog on November 9, 2026 (local time), describing a model designed to help robots independently master intricate manual tasks such as gear meshing and cable connection.
What sets GR00T 2 apart from Isaac Lab 3.0
NVIDIA announced that it will unveil GR00T 2 and the robot learning framework 'Isaac Lab 3.0' at the Conference on Robot Learning (CoRL) held in Austin, Texas. GR00T 2 is the successor to 'GR00T N1,' released in March of last year, and is designed as a 'world action model' that allows robots to perceive their surroundings and predict how their actions will change the environment in order to carry out tasks.
- GR00T 2— Focuses on learning tasks that require precision manipulation beyond simple movement or object relocation, such as meshing gears or connecting cables.
- Isaac Lab 3.0— A training framework that simulates easily deformable objects like cables and fabric, as well as highly articulated robotic hands, in virtual space, enabling simultaneous training of thousands of robots.
- DexBench— A performance benchmarking tool for precision and dexterity developed by NVIDIA in collaboration with Korean physical AI startup RLWRLD, consisting of virtual work environments and 15 tasks across four sectors: retail, logistics, manufacturing, and hospitality.
NVIDIA emphasized that while simulation was previously used mainly to teach humanoid robots to walk or roll, it is now being applied to learning fine manual tasks as well. Global electronics manufacturer Foxconn is applying GR00T 2 to automate server manufacturing plants, while German industrial technology company Siemens is applying or evaluating its performance across various industrial sectors.
Implications for Korean Enterprise AX
Although this is an overseas announcement, there are two points Korean companies should examine. First, the success rate varies from 71% to 95% depending on the number of demonstrations. Since the 50-task average does not represent the success rate for any specific process, companies considering adoption should directly verify how many demonstrations are needed to reach a given level of performance for their own tasks. Second, the benchmarking tool DexBench was developed in partnership with Korean startup RLWRLD and includes manufacturing and logistics environments. Korean manufacturing and logistics companies can use the similarity between their own tasks and DexBench's 15 tasks as a basis for comparison.
For companies considering automation of precision manipulation, the first step is to list which manual processes—such as gear meshing or cable connection—have precision manipulation as the bottleneck.