NVIDIA Unveils GR00T 2, Isaac Lab 3.0, and Agentic Tools at CoRL to Advance Precision Robotics

At the Conference on Robot Learning (CoRL) in Austin, Texas, robotics research is shifting its focus from basic locomotion to complex physical manipulation, such as aligning gears, connecting cables, and assembling delicate components. To support this transition, NVIDIA announced a suite of open technologies spanning world action models, physics simulation engines, benchmarks, and agentic workflows.

Two NVIDIA papers were selected among six finalists out of thousands of submissions for the Best Paper Awards at CoRL: GRAIL (Generating Humanoid Loco-Manipulation from World Models) and RoboTTT: Context Scaling for Robot Policies, which scales visuomotor context by three orders of magnitude without increasing inference latency.

Next-Generation Autonomy with GR00T 2

To enable robots to perceive environments and execute multistep tasks, NVIDIA introduced GR00T 2, an open frontier world action model built on NVIDIA Cosmos. The model was trained on diverse human and robot datasets, including 100,000 hours of NVIDIA open-source human egocentric data.

GR00T 2 demonstrates strong generalization capabilities out of the box, completing 40 out of 50 evaluated tasks at least once without task-specific demonstrations. With a single demonstration per task, the model achieved an average success rate of 71%, which climbed to 95% with five demonstrations. It also allows skill transfer between different robot body form factors without needing target-robot demonstrations.

GR00T 2 currently ranks among the top three models across vision-language-action and world action model categories on the RoboCasa365, RoboLab, RoboDojo, and MolmoSpaces leaderboards. Early enterprise adoption includes:

• Foxconn: Applying GR00T 2 to server manufacturing automation.
• Nexuni: Testing the model for automated laundry handling.
• Siemens: Evaluating it across industrial automation scenarios.
• MicroAGI: Utilizing a reinforcement-learned post-trained GR00T 2 model for Ethernet cable swapping on a YAM robotic arm, using Isaac Lab 3.0 and Newton for environment calibration.

NVIDIA also released the open Video to Data workflow to transform human video footage into simulated robot training data, alongside the EgoHand 3D motion model—developed with data from Noitom Robotics—which surpassed 40,000 Hugging Face downloads within two weeks of release.

High-Throughput Simulation via Isaac Lab 3.0 and Newton

Simulating physical interactions at scale remains critical for safety and throughput before real-world deployment. NVIDIA released Isaac Lab 3.0, its open-source robot-learning framework, featuring tighter integration with Newton, an open-source physics engine.

The Newton integration allows Isaac Lab to accurately model deformable materials—such as cloth and flexible cables—as well as highly articulated systems like multi-jointed robotic hands. In vision-based dexterous manipulation benchmarks, Isaac Lab 3.0 paired with Newton delivered up to 2.5x the training throughput of Isaac Lab 2.3 with PhysX.

Several partners announced integrations built on top of these simulation technologies:

• Skild AI: Introduced STEEL (Scalable Throughput-optimized Engine for Embodied Learning), a rigid-body physics solver built on Newton to speed up Isaac Lab training.
• Flexion: Utilized Newton and Isaac Lab to train its Reflect whole-body manipulation model.
• AlphaNIST: Demonstrated high efficiency by training a single policy across all 20 tasks on the NIST Assembly Task Board. Using Success-Guided Sampling (SGS)—developed by NVIDIA and the University of Washington—AlphaNIST achieved over 98% training success (compared to 81% for separate task policies) while consuming approximately 39x less GPU time.
• Industry Applications: FieldAI, Dyna Robotics (training its semi-humanoid Taku robot), Walden Robotics, Dexmate (testing its Vega robot), and the University of Texas at Austin are adopting Isaac Sim and Isaac Lab across warehouse, manufacturing, and academic workflows.

Standardized Industrial Benchmarks and Agentic Workflows

Evaluating robot reliability under changing conditions is key to moving physical AI out of research labs. NVIDIA previewed version 0.4 of Isaac Lab-Arena, an open-source framework designed to run policies across thousands of parallel variations to isolate failure cases.

Built on Isaac Lab-Arena and Newton, two new open benchmarks were detailed:

• DexBench: Created by RLWRLD and NVIDIA, DexBench tests 15 dexterous tasks across retail, logistics, manufacturing, and hospitality scenes. Partners such as Genesis AI and mimic robotics are testing it via Bifrost’s Manifold platform.
• VABAR: Developed by UC Berkeley and NVIDIA, the Variational Automation Benchmark for Agentic Robotics evaluates physical performance across six standardized industrial tasks with varying difficulty tiers.

To link individual robot skills into reliable, multi-step programs, UC Berkeley and NVIDIA introduced the Graph-as-Policy (GaP) framework. GaP uses AI agents to construct interpretable robot code, test execution in Isaac simulation, and iteratively resolve errors.

"To me, agentic robotics is the most exciting technological development since the internet in 1995," said Ken Goldberg, professor of robotics at UC Berkeley and a lead researcher on VABAR and GaP. Industrial robotics companies including Ambi Robotics, Bosch Research, and Siemens are actively exploring the GaP framework for real-world deployment.