NVIDIA GR00T 2: the open action model that teaches robots to work with their hands
Direct answer
On November 9, 2026, NVIDIA unveiled GR00T 2 at the Conference on Robot Learning (CoRL) in Austin, Texas. It is an open action model, built on Cosmos, NVIDIA's platform for physical AI, and it is designed for dexterous work: robots that handle real-world objects with their hands. NVIDIA also introduced simulation and evaluation tools, including Isaac Lab 3.0. The read for your business: AI robotics stops being a lab prototype and becomes an infrastructure layer that you rent and train in simulation before it touches the real world.
What an action model is and why it matters
- From text to action: a language model answers with words; an action model answers with movements. GR00T 2 takes what it sees and what it is asked to do and decides how to move the robot's hands, step by step.
- Open: NVIDIA presents it as an open model, in the Cosmos line. That means other teams can build on it, adapt it and evaluate it.
- Trained in simulation, tested in the real world: simulation tools such as Isaac Lab 3.0 let you train and evaluate robots in virtual worlds before putting them to work. Simulation is where robotics gets cheaper: failing in a simulator costs cents, failing in a warehouse costs thousands.
- Dexterous work is the bottleneck: robots already move boxes. The leap is handling varied, fragile or irregular objects: folding, assembling, sorting, serving. That is where a general action model changes the math.
What your company should watch
- Warehouses and logistics first: the first publicly traded humanoid robots already work in real warehouses. If your operation moves goods, the question is when the dexterous robot costs less than the extra shift.
- Simulation is your testing ground: before buying a robot, ask for a simulation demo with your objects and your tasks. A general model is evaluated on your real work.
- The open model lowers lock-in risk: building on an open model lets you switch vendors and hardware. A closed system ties you to one company and its calendar.
- Start with one task: pick the repetitive, dexterous job that consumes the most hours in your operation and measure whether a robot does it at the quality you need. Robotics is bought per task, just as AI agents are bought per process.
Note: this article summarizes NVIDIA's November 9, 2026 announcement at the Conference on Robot Learning (CoRL) in Austin, Texas, with coverage from the November 10 AI news digests. NVIDIA published no comparative performance figures for the model in the announcement.
Want AI that works for your business?
Message us on WhatsApp and let's talk about automation and AI agents for your business.
WhatsApp inicIARelated articles
- 5 AI trends for business in 2026 and 2027
- 7 AI myths holding your business back
- Business AI spending cooled: what Ramp's numbers say
Frequently asked questions
What is NVIDIA GR00T 2?
An open action model for robots, unveiled by NVIDIA on November 9, 2026 at the CoRL conference in Austin. It is built on Cosmos and designed for dexterous work: robots that manipulate real-world objects.
What is the difference between an action model and a language model?
A language model answers with words; an action model decides movements. GR00T 2 takes what the robot sees and what it is asked to do and generates the physical actions to do it.
What is Isaac Lab 3.0?
An NVIDIA simulation and evaluation tool for training robots in virtual worlds before putting them to work in the real world, introduced alongside GR00T 2.
Does my company need a robot now?
Probably not today, but it does need to understand the math: warehouses and logistics are the first real uses. The decision is made per task and per cost, just as with AI agents.
