NVIDIA unveils GR00T 2 and Isaac Lab 3.0 for robot training

Bridging the simulation gap

NVIDIA recently showcased its latest robotics advancements at the 2026 Conference on Robot Learning (CoRL) in Austin, Texas. The company introduced two primary tools designed to help robots learn complex, dexterous tasks with more autonomy. These innovations aim to move robotics from controlled laboratory settings into unpredictable real-world environments.

The core challenge in robotics is often the difficulty of teaching a machine to handle physical objects precisely. By providing better simulation tools and world action models, NVIDIA wants to help developers bridge this gap. These tools allow robots to adapt to different bodies and environments more quickly than previous methods allowed.

As Blockchain News reported, these updates focus on precision and scalability. The goal is to create robots that can handle intricate movements, such as manipulating laundry or swapping cables, without requiring a human to guide every single motion in real-time.

How GR00T 2 learns from humans

GR00T 2 is a world action model built on NVIDIA Cosmos that uses massive amounts of data to train robots. Specifically, it utilizes 100,000 hours of open-source human egocentric data, which is footage filmed from a person's first-person perspective. This helps the model understand how humans interact with the world and how they move their hands and bodies.

One of the most significant features of GR00T 2 is its ability to generalize from very little information. In testing, the model achieved a 71% success rate across 50 different tasks using only one demonstration per task. When given five demonstrations for each task, that success rate jumped to 95%.

The model also excels at skill transfer, meaning a robot can learn a movement from a human and then apply that skill to different types of robot bodies. It also provides tools for AI agents to sequence these skills together. This allows the robot to recover from errors and plan out a series of actions to reach a specific goal.

Faster training with Isaac Lab 3.0

While GR00T 2 provides the

Validating performance and use cases

NVIDIA has verified the capabilities of GR00T 2 through several industry benchmarks. The model currently ranks among the top three performers across various testing environments, including RoboCasa365, RoboLab, RoboDojo, and MolmoSpaces.

Beyond theoretical testing, the model is already being applied to specific hardware tasks. For instance, MicroAGI used GR00T 2 to enable a YAM robotic arm to perform the specific task of swapping Ethernet cables.

Flexible tools for researchers

Isaac Lab 3.0 provides researchers with a high degree of flexibility in how they build their simulations. Developers can choose between different solvers, such as NVIDIA PhysX or the Newton physics engine, and select specific renderers for training based on perception needs.

The platform also allows for the application of reinforcement post-training libraries to foundation models like GR00T. Additionally, Flexion showcased how Isaac Lab and the Newton engine support its Reflect model, which combines reasoning with whole-body manipulation.

New workflows and data models

NVIDIA introduced several new workflows to streamline the transition from human behavior to robot action. One example is Video to Data, a process that converts standard human video into simulated demonstrations that robots can use for training.

The company also released EgoHand, a model designed to reconstruct 3D hand motions. This tool gained significant traction quickly, reaching over 40,000 downloads on Hugging Face within just two weeks of its initial launch.

Expanding industrial capabilities

NVIDIA is actively collaborating with a diverse group of partners to bring these robotics advancements to the wider community. Organizations such as Ambi Robotics, Bosch Research, Siemens, and UC Berkeley are working together to integrate these tools into various sectors. These partnerships aim to facilitate broader industrial adoption by providing the necessary infrastructure for robots to operate in complex, real-world settings.

Specific industrial applications are already being explored to test the limits of these new models. For example, MicroAGI utilized GR00T 2 to enable a YAM robotic arm to perform the precise task of swapping Ethernet cables. This demonstrates how the technology can be applied to specific, high-precision manual tasks that currently require human intervention in technical environments.

Boosting simulation and performance

The technical performance of the new software suite provides significant improvements over previous versions. Isaac Lab 3.0 achieved up to 2.5x the training speed of its predecessor, Isaac Lab 2.3, specifically in benchmarks involving vision-based dexterous manipulation. This increase in throughput allows researchers to train robots much faster than was previously possible in simulated environments.

To further accelerate these training processes, Skild AI announced the STEEL solver, which is a rigid-body physics engine built on the Newton engine. This solver is designed to work with Isaac Lab to enhance the speed and accuracy of robot learning. Additionally, GR00T 2 demonstrated strong out-of-the-box performance by successfully completing 40 of 50 evaluated tasks without requiring any task-specific demonstrations.