NVIDIA introduced GR00T 2, an open frontier world action model, and a suite of simulation and benchmarking tools at the Conference on Robot Learning in Austin, Texas, aiming to help researchers teach robots dexterous manipulation tasks. The company said the model completed 40 of 50 evaluated tasks at least once without task-specific demonstrations and achieved 71% average success across 50 tasks with a single demonstration per task, rising to 95% with five demonstrations.
Simulation has long been used to train robots to walk, roll and hop, but researchers are now applying it to tasks requiring a finer touch, such as aligning gears, connecting cables and assembling small components, according to NVIDIA. The company said it is working with partners including Siemens, Bosch Research and Ambi Robotics to bring these advances to the broader robotics community.
NVIDIA's Isaac Lab 3.0, an update to its open-source robot-learning framework, has tighter integration with the Newton physics engine to model deformable objects such as cables and cloth, and complex articulated systems like robotic hands. In a vision-based dexterous manipulation benchmark, Isaac Lab 3.0 with Newton achieved up to 2.5 times the training throughput of the previous version with NVIDIA PhysX, the company said.
Two NVIDIA papers, GRAIL and RoboTTT, were among six finalists out of thousands of submissions for the Best Paper Awards at CoRL. GRAIL generates 3D assets and simulator-ready scenes for robot learning, while RoboTTT introduces a model and training recipe that scales visuomotor context three orders of magnitude beyond current policies without increasing inference latency.
NVIDIA introduced two open benchmarks for robot evaluation. DexBench, developed with RLWRLD, tests dexterity across 15 tasks in retail, logistics, manufacturing and hospitality. VABAR, from UC Berkeley and NVIDIA, assesses robots on six industrial tasks with standardised variations and difficulty tiers. Both use Isaac Lab-Arena, an open-source framework for scaled policy evaluation.
The company highlighted agentic robotics, a paradigm where AI agents create interpretable code for physical robots. UC Berkeley’s Graph-as-Policy framework uses agents to connect individual skills into robot programs, which are tested in NVIDIA’s open simulation tools and refined based on failures. “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 on the VABAR and GaP projects. Ambi Robotics is exploring GaP for industrial automation, and Ambi Robotics, Bosch Research and Siemens are exploring GaP workflows across diverse industrial scenarios.
Several partners demonstrated new tools. Skild AI announced STEEL, a rigid-body physics solver built on Newton, while Flexion showed how Isaac Lab and Newton help train its robot-intelligence model. Foxconn is applying GR00T 2 to factory automation for server manufacturing, and Siemens is evaluating it across industrial use cases.
Analysis
Why This Matters
- Simulation-based training for precise manipulation could speed up deployment of robots in manufacturing, logistics and hospitality by reducing the need for physical trials.
- The open-source approach, including models, frameworks and benchmarks, allows a broader range of researchers and companies to contribute to and benefit from advances.
- Standardised benchmarks such as DexBench and VABAR may help the field measure and compare robot dexterity and adaptability more consistently.
Background
NVIDIA has been building simulation platforms for robotics through its Isaac framework, and the Cosmos platform for AI training. The GR00T model family is intended as a foundation for robot learning, akin to large language models for text. CoRL is a major academic conference for robot learning research, and NVIDIA frequently presents new work there.
What to Watch
- How quickly industrial partners such as Foxconn and Siemens move from evaluation to production deployment with GR00T 2.
- Whether the agentic robotics paradigm, particularly the Graph-as-Policy framework, gains adoption beyond the initial partners.
- Future updates to Newton physics engine that may enable simulation of even more complex manipulations, such as soft objects or assembly tasks.
Checked
11 claims in this story checked against 1 source: 11 supported.
Show each claim
- SupportedNVIDIA introduced GR00T 2, an open frontier world action model, at CoRL in Austin, Texas.NVIDIA Blog: “At CoRL, NVIDIA is introducing GR00T 2, an open frontier world action model built on NVIDIA Cosmos.”
- SupportedGR00T 2 completed 40 of 50 evaluated tasks at least once without task-specific demonstrations.NVIDIA Blog: “Completed 40 of 50 evaluated tasks at least once without task-specific demonstrations.”
- SupportedGR00T 2 achieved 71% average success with one demonstration per task and 95% with five.NVIDIA Blog: “Achieved 71% average success across 50 tasks with just one demonstration per task, increasing to 95% with five.”
- SupportedSimulation that has long trained robots in locomotion is now being used to teach tasks requiring finer manipulation.NVIDIA Blog: “Simulation, which has long been the proving ground for robots learning to walk, roll and hop, is now helping researchers teach robots tasks that require a finer touch: aligning gears, connecting cables and assembling small components.”
- SupportedNVIDIA is working with partners including Siemens, Bosch Research and Ambi Robotics to bring robotics advances to the broader community.NVIDIA Blog: “NVIDIA is working with partners like Ambi Robotics, Bosch Research, Dexmate, Dyna Robotics, FieldAI, Flexion, Noitom Robotics, Siemens, Skild AI, RLWRLD, UC Berkeley, Walden Robotics and the University of Washington to bring these advancements to the broader robotics community.”
- SupportedIsaac Lab 3.0 has tighter Newton integration for modelling deformable objects and complex articulated systems such as robotic hands.NVIDIA Blog: “Tighter integration with Newton, an open source physics engine, now allows Isaac Lab to model deformable objects such as cables and cloth, along with complex articulated systems such as robotic hands with many interacting joints.”
- SupportedIn a vision-based dexterous manipulation benchmark, Isaac Lab 3.0 with Newton achieved up to 2.5 times the training throughput of Isaac Lab 2.3 with PhysX.NVIDIA Blog: “In a vision-based dexterous manipulation benchmark, Isaac Lab 3.0 with Newton achieved up to 2.5x the training throughput of Isaac Lab 2.3 with PhysX.”
- SupportedGRAIL and RoboTTT were among six finalists out of thousands of submissions for CoRL Best Paper Awards.NVIDIA Blog: “Two NVIDIA papers are among six finalists out of thousands of submissions for the Best Paper Awards at CoRL: GRAIL (Generating Humanoid Loco-Manipulation from World Models)”
- SupportedDexBench, developed by RLWRLD with NVIDIA, features 15 tasks across retail, logistics, manufacturing and hospitality.NVIDIA Blog: “DexBench, developed by RLWRLD in collaboration with NVIDIA, takes dexterity evaluation off the tabletop and into real-world workflows. Built using NVIDIA Isaac Lab-Arena, it features 15 tasks across four scenes spanning retail, logistics, manufacturing and hospitality”
- SupportedVABAR from UC Berkeley and NVIDIA evaluates robots across six industrial tasks with standardized variations and difficulty tiers.NVIDIA Blog: “VABAR — the Variational Automation Benchmark for Agentic Robotics — from UC Berkeley and NVIDIA, evaluates robots across six industrial tasks. Each carries standardized variations, difficulty tiers and success metrics.”
- SupportedKen Goldberg said, “To me, agentic robotics is the most exciting technological development since the internet in 1995,” and is a UC Berkeley robotics professor and a lead on the VABAR and GaP projects.NVIDIA Blog: ““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 on the VABAR and GaP projects.”
