- 5 min read
- 7 key claims
- 1 source cited
- For Developers, AI makers, Businesses
Overview
At the Conference on Robot Learning (CoRL) in Austin, Texas, NVIDIA announced GR00T 2, an open frontier world action model, and Isaac Lab 3.0, an open-source robot-learning framework for simulation. The products target a shift in robot training from locomotion toward contact-rich manipulation: aligning gears, connecting cables and assembling small components [1].
These are NVIDIA product announcements. The company also announced new open benchmarks developed with partners, but did not name them or describe their scope or evaluation methodology in the post [1].
NVIDIA reports that GR00T 2 completed 40 of 50 evaluated tasks at least once without task-specific demonstrations. It also reports average success rates of 71% across the 50 tasks with one demonstration per task and 95% with five demonstrations [1]. These are vendor claims; the post does not provide benchmark definitions, test protocols, baselines or independent verification for the figures, so they cannot be compared with other systems from the published material [1].
Why it matters
For developers: the announcements form a more integrated path from synthetic demonstrations and generated scenes to reinforcement post-training. Teams can choose between PhysX and Newton, select renderers for perception-based training and automate parts of environment setup through agents [1]. Hypothesis: this may reduce the effort required to produce demonstrations and build training environments, but the post does not quantify that reduction.
For robotics teams: solver choice affects how cables, cloth and multi-joint hands are modeled, so comparison between physics engines becomes part of environment design rather than a separate exercise [1].
For businesses: partner evaluations and applications extend beyond research, including Nexuni testing robotic laundry handling, Foxconn applying GR00T 2 to factory automation for server manufacturing and Siemens evaluating industrial use cases. These are reported evaluations or applications, not demonstrated production outcomes [1].
Hypothesis: adopting GR00T 2, Isaac Lab and Newton together could increase a team's dependence on NVIDIA's release cadence, tooling and hardware ecosystem. The post does not provide pricing or total cost figures for assessing that trade-off [1].
Opportunities
- Build simulated training pipelines: Isaac Lab 3.0 supports agent-created environments, simulation-ready assets and reinforcement post-training for foundation robot models such as GR00T [1].
- Generate scene content: GRAIL provides a pipeline for generating 3D assets and simulator-ready scenes, while Video to Data turns human video into simulated robot demonstrations [1].
- Tune the physics environment: teams can select PhysX or Newton and choose renderers depending on whether the task involves rigid bodies, deformable objects or perception inputs [1].
- Use a third solver path: STEEL is a rigid-body physics solver built on Newton for use with Isaac Lab. This is a Skild AI announcement reported by NVIDIA [1].
- Explore industrial applications: laundry handling, server manufacturing and broader industrial evaluation are the stated areas for the named partners [1].
Risks
Simulation results may not carry over to physical robots. The principal unknown is unresolved: the post does not establish that simulation-trained skills transfer to physical robots for these tasks. Its general discussion of physical transfer does not substitute for task-level hardware results [1].
Benchmark protocols for GR00T 2 and the throughput comparison are also unspecified, which limits interpretation of the reported figures [1].
Agent-generated environment setup and calibration introduces another point of failure: an unreviewed setup can produce training data that appears valid but does not transfer to hardware [1].
The post mentions open-source components and open datasets but does not give licensing terms, data-governance details or rules for using human egocentric recordings [1].
Context
Simulation has long supported robot training for locomotion. The new tools target manipulation, where robots must control contact, friction and deformable objects such as cables and cloth [1].
Isaac Lab 3.0 adds agent-based environment setup, simulation-ready assets, a choice of PhysX or Newton physics solvers, renderer options and a reinforcement-learning post-training library for models including GR00T. NVIDIA says tighter integration with Newton allows the simulator to handle deformable objects and articulated systems such as multi-joint robotic hands [1].
GR00T 2 is trained on robot and human data, including 100,000 hours of NVIDIA's open-source human egocentric data. That corpus size is separate from the task-specific demonstration counts reported for the evaluation: one or five demonstrations per task [1].
NVIDIA's Video to Data workflow converts human video into simulated robot demonstrations, linking data collection to simulated training. GRAIL, another NVIDIA project described in the post, generates 3D assets and simulator-ready scenes, while RoboTTT is described as a robot model and training recipe for scaling visuomotor context without increasing inference latency [1].
Known, claimed, inferred
GR00T 2 and Isaac Lab 3.0 are the central product announcements. GR00T 2 supplies a policy model for dexterous manipulation, while Isaac Lab 3.0 supplies the environment in which such policies can be trained and refined [1].
The clearest reported numerical claims are GR00T 2's completion of 40 of 50 tasks at least once without task-specific demonstrations, its 71% average success with one demonstration per task and 95% with five, and up to 2.5x training throughput for Isaac Lab 3.0 with Newton over Isaac Lab 2.3 with PhysX in a vision-based dexterous manipulation benchmark [1]. All three are vendor claims without published protocols or independent verification [1].
GR00T 2 currently ranks among the top three models across vision-language-action and world-action-model categories on the named benchmarks, with multiple leaderboard wins. The post provides no leaderboard snapshot or date, so those positions cannot be checked from the page [1].
GRAIL and RoboTTT are conference-paper selections rather than product releases. Both were named as finalists among six for CoRL Best Paper Awards from thousands of submissions [1].
Hypothesis: the components appear intended to work as one workflow: generate scenes, simulate contact-rich tasks, train or post-train a policy, then move toward deployment. MicroAGI's Ethernet cable-swapping work is the clearest example, using a GR00T 2 policy post-trained with reinforcement learning and Isaac Lab 3.0 with Newton to automate environment creation and calibration before training the policy [1]. Whether the same pipeline generalizes to tasks with different geometry, materials or sensing arrangements is not shown.
Limitations & uncertainties
The source excerpt ends during the Flexion discussion, so its example and any accompanying results are unavailable. The post reports no physical-robot results for the manipulation tasks described. Benchmark definitions, test protocols, baselines and independent verification behind the GR00T 2 success rates and the Isaac Lab throughput figure are not provided, and no leaderboard snapshot or date is included. The evidence does not establish whether the reported rankings remain current beyond NVIDIA's publication, and no pricing, licensing or data-governance terms are given.
Evidence & sources
The key statements of this analysis, labeled by type, with the sources they rely on. Trendoviz does not run product tests.
- Primary-source or vendor statementNVIDIA introduced GR00T 2, an open frontier world action model, and Isaac Lab 3.0, an open-source robot-learning framework, at CoRL for work on delicate manipulation tasks.[1]
- Primary-source or vendor statementNVIDIA reports that GR00T 2 completed 40 of 50 evaluated tasks at least once without task-specific demonstrations, with 71% average success using one demonstration per task and 95% using five.[1]
- Primary-source or vendor statementNVIDIA reports up to 2.5x training throughput for Isaac Lab 3.0 with Newton compared with Isaac Lab 2.3 with PhysX in a vision-based dexterous manipulation benchmark.[1]
- Primary-source or vendor statementIsaac Lab 3.0 lets teams switch between PhysX and Newton physics solvers, choose renderers, set up environments with agents and apply reinforcement post-training to models such as GR00T.[1]
- Primary-source or vendor statementGR00T 2 is trained on robot and human data, including 100,000 hours of NVIDIA open-source human egocentric data, which is distinct from the task-specific demonstration counts used in its evaluations.[1]
- Primary-source or vendor statementNexuni is testing GR00T 2 for robotic laundry handling, Foxconn is applying it to factory automation for server manufacturing, and Siemens is evaluating it for industrial use cases.[1]
- Primary-source or vendor statementTwo NVIDIA papers, GRAIL and RoboTTT, were selected as finalists among six for CoRL Best Paper Awards from thousands of submissions.[1]
Sources
- 1Practice Makes Precise: At CoRL, NVIDIA and Partners Reveal How Robots Rehearse Delicate Work in Simulation●blogs.nvidia.comretrieved Nov 18, 2026
Updates & corrections
- First published · published automatically after passing every check
Found an error? Report it.