NVIDIA unveils GR00T 2 and Isaac Lab 3.0 for robot learning

Advancing robot manipulation

NVIDIA recently introduced GR00T 2, an updated open frontier world action model designed to help robots master dexterous manipulation. The company showcased these new tools at the Conference on Robot Learning (CoRL) in Austin, Texas. These models are intended to help robots learn how to handle objects with precision and grace.

To build this model, NVIDIA utilized a massive dataset consisting of 100,000 hours of open-source egocentric data. This data includes diverse information from both robots and humans to help the AI understand complex movements. The goal is to enable robots to perform tasks that require high levels of coordination and fine motor skills.

GR00T 2 shows significant progress in learning from limited human input. The model achieved a 71% average success rate across 50 tasks with only one demonstration per task. This success rate climbed to 95% when provided with five demonstrations, showing how efficiently the model can generalize new skills.

Simulating complex physics

The launch included Isaac Lab 3.0, which serves as an open-source framework for robot learning. This version provides the foundation for researchers to build training environments with simulation-ready assets. It allows users to switch between different physics solvers and renderers depending on their specific perception needs.

A major update in Isaac Lab 3.0 is the tighter integration with Newton, an open-source physics engine. This combination allows the software to model deformable objects such as cloth and cables, which are traditionally difficult to simulate. It also supports complex articulated systems like robotic hands that have many interacting joints.

Accurate physics are essential for transferring skills learned in a virtual world to a physical robot. By combining the scale of Isaac Lab with the precision of Newton, researchers can train thousands of simulated robots in parallel. This setup reportedly achieved up to 2.5x the training throughput of previous versions using different solvers.

Industrial applications and partners

Several industrial partners are already testing these technologies for real-world use cases. For example, Nexuni is exploring GR00T 2 for robotic laundry tasks, while Foxconn is applying it to factory automation for server manufacturing. Siemens is also evaluating the model across various industrial applications to see how it handles heavy-duty work.

Other companies are using the Isaac Lab framework to train specific robot types for diverse environments. Dyna Robotics is using parallel simulation to train controllers for a semi-humanoid robot intended for warehouses and hotels. Meanwhile, Walden Robotics is using photorealistic rendering to train a general-purpose robot for logistics and manufacturing.

The technology also supports specialized tasks like cable swapping and assembly. MicroAGI is using GR00T 2 to perform Ethernet cable swapping with a YAM robotic arm. Additionally, a project from the University of Washington demonstrated a policy learning 20 assembly tasks using 39x less GPU time than traditional methods.

New benchmarks for dexterity

NVIDIA also introduced new frameworks to measure how well robots perform in unpredictable environments. The Isaac Lab-Arena is an open-source framework that allows developers to run policies across thousands of variations in parallel. This helps researchers identify specific conditions that cause failures during training runs.

Two new benchmarks were released to test robot adaptability in industrial settings. DexBench, developed with RLWRLD, features 15 tasks across scenes like retail and hospitality to test spatial and physics variations. VABAR, created with UC Berkeley, evaluates robots across six industrial tasks using standardized difficulty tiers and success metrics.

These benchmarks move robot testing away from simple tabletop exercises and into real-world workflows. By providing standardized metrics, NVIDIA aims to help developers understand how robots generalize across different objects. Industry partners like Genesis AI and Mimic Robotics are already beginning to evaluate these new benchmarks.