NVIDIA unveiled GR00T 2 on November 9, 2026, at the Conference on Robot Learning, known as CoRL, in Austin, Texas. GR00T 2 is a robot foundation model that learns manual manipulation skills from human first-person video.
A foundation model is a large model trained on broad data that can be adapted to many different tasks, rather than a model built for one narrow job. In robotics, a foundation model is intended to give a robot a general ability to understand and act in the physical world, which can then be adapted to specific tasks such as picking up an object, assembling a part, or handling a tool.
Manual manipulation is the part of robotics concerned with hands and grasping: how a robot picks something up, holds it, moves it, and puts it down without dropping or damaging it. It is one of the hardest problems in robotics, because it requires a robot to understand an object's shape, weight, and fragility, and to adjust its grip in real time.
Learning From Human First-Person Video
GR00T 2 trains jointly on robot-collected data and first-person human video. First-person video is video recorded from the point of view of the person performing a task, for example from a camera worn on the head or chest. It shows what the person sees as they reach for an object, grasp it, and use it.
That choice of training data is significant. Robot data is expensive to collect: a robot must be set up, operated, and supervised, and each demonstration takes time. Human video is far more abundant.
The challenge is that a human hand and a robot gripper are not the same. A model that learns from human video must translate what it sees a human do into actions a robot can actually perform. Training jointly on robot data and human video bridges that gap: the human video provides breadth, showing many different tasks, while the robot data shows what the robot itself can do. A model that can learn from human video could reduce that bottleneck and allow robots to be taught new tasks more quickly.
Why the Conference Matters
CoRL, the Conference on Robot Learning, is a research conference focused on the intersection of robotics and machine learning. An unveiling there is aimed at the research community as well as at industry.
NVIDIA is a central company in AI infrastructure. Its graphics processing units are used to train large models, and its software platforms are used to deploy them. In robotics, NVIDIA provides both the compute for training and the platforms for running models on robots. The unveiling does not by itself prove that robots powered by GR00T 2 can perform a given task reliably in a real factory, warehouse, or home. The test for any robot foundation model is deployment: whether it works consistently, safely, and at a cost that makes sense for the customer.
The Wider Context: Robots and Foundation Models
The idea of a foundation model for robots borrows from language models. A language model is trained on a vast amount of text, and can then be adapted to summarise, translate, or answer questions. A robot foundation model is trained on data about the physical world, and can then be adapted to manipulate objects, navigate spaces, or perform tasks.
The analogy has limits. Robotics involves the physical world, where a mistake can break an object, damage equipment, or injure a person. A robot that misunderstands a task does not just produce a wrong answer; it acts in the world. A robot that can pick up, handle, and place objects is far more useful than one that can only move around, and learning manipulation from human video could expand the range of tasks a robot can be taught without a specialist programming each movement by hand.
Why It Matters for Founders and Investors
For founders in robotics, GR00T 2 is a platform signal. When a major infrastructure company releases a foundation model for robots, it changes the build-versus-buy decision: a startup planning to build its own manipulation model from scratch may instead build on GR00T 2 and focus on the application, the hardware, or the customer relationship.
For investors, the unveiling is a reminder that robotics is entering a phase where software, not only hardware, is the differentiator. A robot's ability to learn new tasks quickly, from data rather than from hand-coded instructions, may matter more over time. The companies that can demonstrate reliable manipulation in real settings, not only in a laboratory, will be the ones to watch.
The key question is not whether a model can learn from human video in principle, but whether it can do so reliably enough for a customer to deploy it. That will be answered not by an unveiling, but by deployments in factories, warehouses, and other real environments in the months that follow.
Angeled.com connects founders with investors.
Read Next
OpenAI Releases Codex Security as Open-Source CLI and SDK
Databricks Unveils Always-On Genie, an AI Agent That Stops Waiting to Be Asked
DeepMind Launches AlphaGenome 2, a DNA Model That Reads Genomes Cell by Cell
ElevenLabs Unveils 13 New Products and Upgrades at Its New York Summit
Google Launches Gemini Decision 1, a Model That Answers Instead of Writing
Runway Open-Sources Praxis-1, Its Robot Model That Learned by Watching Video



