NVIDIA used the Conference on Robot Learning (CoRL) on Nov. 9 to introduce GR00T 2, an open world-action model, alongside Isaac Lab 3.0, an open-source robot-learning framework. The announcement joins model, simulation and evaluation work, but its headline success rates come from company-described experiments—not evidence that robots already perform those tasks reliably in factories or homes.

What NVIDIA announced

NVIDIA’s Nov. 9 announcement at CoRL in Austin spans several related but distinct pieces of work: GR00T 2, Isaac Lab 3.0, an Isaac Lab-Arena developer preview, new benchmarks, data workflows and demonstrations from partners. It is better understood as a research and developer ecosystem update than as a single robot product launch. The company describes GR00T 2 as an open frontier world-action model, and Isaac Lab as an open-source robot-learning framework. The Robot Report independently covered the launch and the broad outline of the product claims, while the technical detail is principally supplied by NVIDIA itself.

The central proposition is that robot learning can combine human video, demonstrations on other robot bodies, simulation and relatively small quantities of task-specific robot data. NVIDIA’s model material describes predicting future video and robot actions together: rather than mapping a command directly to a movement alone, the model attempts to represent how a scene changes as actions unfold. That framing helps explain its “world-action model” label, but it should not be mistaken for a guarantee of robust physical understanding. The practical evidence remains a set of selected tasks, demonstrations and evaluations reported by the company.

Alongside the model, Isaac Lab 3.0 aims to let researchers assemble and train simulation environments through a more modular workflow. The release combines backend options, including NVIDIA PhysX and the open-source Newton physics engine, with renderer and training choices. NVIDIA’s blog also describes Isaac Lab-Arena as a framework for repeatable, large-scale policy evaluation, and says version 0.4 is arriving as a developer preview. The benchmark and arena announcements matter because performance claims are only interpretable when the task definitions, trial counts, environments and evaluation rules are visible.

GR00T 2: learning from people and robots

Three kinds of pretraining data

On its official GR00T 2 research page, NVIDIA says the pretraining mixture contains 402,730 hours of egocentric human video, 10,184 hours of handheld UMI-gripper demonstrations across 2,897 tasks and 23,500 hours of robot teleoperation. The same source says the Human-1M dataset includes 100,000 hours that are openly available. These figures are NVIDIA’s accounting of its training data; they are not an independent audit of dataset composition, quality, licensing, representativeness or the contribution of each source to model performance.

The research page describes transfer between embodiments: a skill seen in human video or demonstrated using a handheld gripper may be attempted on a robot that did not receive a task-specific demonstration on its own body. NVIDIA calls this cross-adaptation, or X-Adapt. Its examples include lab funnel preparation, a toy dump-truck assembly and two-handed food preparation. The page reports example success rates for individual demonstrations, but such selected demonstrations are illustrations of capability, not broad estimates of how often the model will succeed across unrelated tasks, users, objects or workplaces.

That distinction is important in robotics. A human video records appearance and motion, but not necessarily the forces, contact states or control signals required to reproduce a task. A gripper demonstration differs mechanically from a five-fingered hand. The model must bridge those differences while handling occlusion, object variation and errors in execution. NVIDIA says GR00T 2 learns future video and action jointly and uses robot-specific heads to translate plans into motor commands. The idea is to reuse broad visual experience while retaining embodiment-specific control; the difficult transfer is also exactly why task-level demonstrations and physical testing remain relevant.

Human demonstrations and adaptation

NVIDIA reports that on 50 tasks absent from the target robot’s own training data—but represented in human video or other robots’ data—the pretrained model succeeded at least once on 40 tasks without target-robot demonstrations. The research page further reports 40% average success at zero-shot, 71% with one demonstration per task and 95% with five. For the five-demonstration condition, the company says the demonstrations amount to under three minutes of robot data per task and that each adaptation used a 30-minute LoRA fine-tune on a single RTX 5090.

These statistics describe a controlled evaluation as NVIDIA presents it, not a deployment forecast. The 40-of-50 figure is a task-coverage measure—whether a task succeeded at least once—not the same statistic as average success across trials. A model may show one successful attempt and still fail frequently. The 71% and 95% values are reported averages across the 50 tasks; they do not imply that each individual task reaches those rates. The underlying page says 20 trials per task for several described evaluations and notes that five demonstrations brought 37 of 50 tasks to success in every trial. Trial-level outcomes, task selection and replication details are essential context when comparing these figures with other systems.

NVIDIA also reports additional capabilities beyond the headline 50-task result. Its research page describes real-robot reinforcement learning on three contact-rich insertion tasks, increasing success to between 88% and 96%, as well as a reasoning model calling GR00T 2 through an API to direct sequential tasks. For a visual-steering test, NVIDIA reports that instructions paired with marked grasp and release locations raised “what-and-where” following from 40% to 92%. These are company-reported experimental outcomes. The page’s scope—particular robots, task setups and trial sets—does not establish safety, reliability or general performance in unstructured commercial settings.

What the success rates mean

The different percentages in launch coverage should not be collapsed into one claim that the model is “95% reliable.” NVIDIA’s blog says GR00T 2 completed 40 of 50 evaluated tasks at least once without task-specific demonstrations; this is task coverage. Its 71% and 95% figures refer to average success across the 50-task suite after one and five demonstrations per task. The research page adds the evaluation details and describes zero-shot average trial success as 40%, alongside a 30-minute fine-tune for the adapted conditions. These are related measurements, but their denominators and interpretations differ.

The numbers also do not establish that the model can be installed and immediately usefully deployed without engineering. The evaluated target robot, available sensors, action representation, task instructions, setup, fine-tuning protocol and definition of success all condition results. A manipulated object can be successfully placed in a laboratory trial without demonstrating throughput, recovery from uncommon failures, safe operation near people, or tolerance of changes to lighting and hardware. There is no basis in the cited materials to generalize the quoted average to a customer’s facility without reproducing the evaluation in that setting.

NVIDIA’s research page offers more detail on selected evaluations than the launch blog, but both are company sources. The Robot Report independently confirms the announced figures and release scope, yet its account largely relays NVIDIA’s statements rather than independently reproducing experiments. Accordingly, this coverage treats the rates as attributed company findings, not externally verified benchmarks. Readers should look for full methods, task lists, data splits, uncertainty intervals, hardware configuration and independent replication before treating small differences between models as meaningful.

Isaac Lab 3.0 and Newton

Isaac Lab 3.0 is the software counterpart to the model announcement: a framework for building robot-learning environments and training or evaluating policies in simulation. NVIDIA’s Nov. 9 blog highlights simulation-ready assets, agent-assisted environment setup, renderers for perception training, a post-training library for foundation models and the ability to choose between PhysX and Newton. The official GitHub release page supplies the more concrete availability status: Isaac Lab 3.0 v3.0.0 was published Oct. 20 and is marked generally available. Thus the conference announcement is not the first availability date for the core 3.0 release, even though NVIDIA presented it as part of the CoRL news package.

The official release notes describe a multi-backend architecture and one task API across physics, rendering and visualization options. Newton is a Warp-native physics backend, with solver paths that include MuJoCo-Warp, VBD and MPM; optional OV libraries provide standalone PhysX and RTX capabilities. Isaac Sim remains available for workflows that need its full PhysX/RTX stack and associated tools, but the release notes say it is optional for many workflows. That modularity may make it easier to run experiments without adopting every component of a single simulator stack. It does not mean every task supports every backend combination: the release notes explicitly advise checking task-specific supported presets.

NVIDIA says tighter integration with Newton supports deformable materials such as cables and cloth, as well as articulated systems such as multi-joint robotic hands. Its launch blog reports that, on 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. “Up to” is a best-case comparison, not a universal speedup. The public announcement identifies a particular benchmark and software pairing but does not, in the blog itself, provide enough detail to infer the result across all tasks, hardware, solver settings or rendering choices. Throughput is also distinct from final policy quality or transfer to a physical robot.

Simulation can let researchers run many environments in parallel, explore variations and repeat experiments more cheaply than using physical hardware alone. But more simulated trials do not automatically produce more capable robots. Contact modeling, object properties, camera rendering and control timing can differ from reality. A policy that excels under one simulated configuration may fail when friction, tolerances or sensing change. Isaac Lab’s support for multiple backends and Newton’s deformable-object solvers are technical tools for experimentation, not proof that the simulation-to-reality gap has been eliminated.

Benchmarks and evaluation

NVIDIA’s release story places evaluation alongside training. The company describes Isaac Lab-Arena as an open framework for composing repeatable evaluations and running policies across thousands of variations in parallel; the latest version announced at CoRL is a developer preview, not the same as a final production guarantee. The purpose is to expose conditions associated with failure so researchers can decide what to improve. For an evaluation framework to support meaningful comparisons, users still need clear scenario definitions, repeatability and transparent reporting of how tasks and variations are selected.

Two industrial benchmarks named in the announcement are DexBench and VABAR. NVIDIA describes DexBench, developed by RLWRLD with NVIDIA, as 15 tasks across four scenes spanning retail, logistics, manufacturing and hospitality, with variations in objects, spatial arrangements and physics. NVIDIA and UC Berkeley describe VABAR—the Variational Automation Benchmark for Agentic Robotics—as covering six industrial tasks, each with standardized variations, difficulty tiers and success metrics. Those benchmarks aim to move beyond narrow tabletop demonstrations, but they are early evaluation resources; a benchmark score is not an assurance that a robot is ready for any particular industrial line.

The announcement also highlights AlphaNIST, a single policy trained in Isaac Lab for all 20 tasks on the NIST Assembly Task Board. NVIDIA reports more than 98% training success versus 81% for separate per-task policies, using roughly 39 times less GPU time, and attributes the method to Success-Guided Sampling with the University of Washington. The “training success” qualifier matters: the reported result is not a statement that independent real-world deployments achieve the same rate. It is evidence about a particular training setup and task board as described by the collaborators. Likewise, NVIDIA says Success-Guided Sampling can scale to more than a million parallel environments; this refers to simulation scale, not a million physical robots or a million independently validated tasks.

New benchmarks are useful when they make methods and failure modes inspectable, but comparison depends on implementation details. A score can change with the robot embodiment, sensors, object set, randomization, training budget or success threshold. Early access and developer previews also mean the resources may evolve. Researchers and buyers should consult the benchmark documentation, inspect the exact test conditions and distinguish simulation results from physical trials before drawing conclusions about commercial reliability.

Availability and limits

Isaac Lab 3.0 v3.0.0 is publicly available through NVIDIA’s official GitHub release, which says it is the recommended baseline for new projects and migrations from 2.x. The release describes compatibility targets and installation paths, including Newton-based workflows and optional Isaac Sim components. Developers should consult the release notes and installation guide for supported platforms, dependencies and task-specific combinations rather than assuming a feature works across every setup.

NVIDIA labels GR00T 2 an open model, links to its research page and paper, and says code availability is in December. The official page distinguishes the announcement from the code release timing; the material cited here does not establish that all model weights, training data or commercial rights are already available on Nov. 10. “Open” should therefore be read with attention to what artifacts and terms are actually published, not as shorthand for unrestricted access to the entire training corpus or a ready-to-run package. NVIDIA’s 100,000-hour Human-1M-100K release is a dataset claim about that subset, not public access to every hour of data used in pretraining.

Partners named in NVIDIA’s blog are testing, evaluating, applying or demonstrating different parts of the stack. For example, the company says Nexuni is testing GR00T 2 for laundry handling, Foxconn is applying it to server-manufacturing automation and Siemens is evaluating it across industrial use cases. These descriptions indicate activity, not proven production deployment at scale. A partner evaluation or demonstration can show a promising workflow while leaving unanswered questions about sustained operation, economic performance, safety certification and maintenance.

Why the release matters—and what it does not show

The underlying research direction is consequential: robots are difficult to program for every object and setting, and demonstrations are costly to collect on each physical platform. A model that can reuse human video or demonstrations from another embodiment, then adapt with a handful of robot examples, could reduce some data bottlenecks. A simulation framework that supports different physics engines and parallel evaluation could help teams test ideas earlier and identify brittle cases before hardware deployment. Those are plausible benefits to investigate, not outcomes guaranteed by the announcement.

The release also reflects a broader convergence: foundation models supply reusable representations, simulation supplies scale, and agents can plan sequences around a lower-level manipulation policy. Each layer adds its own failure modes. A reasoning agent may choose an inappropriate step; a world-action model may mispredict contact; simulation may omit a relevant physical effect; and the robot may encounter a condition absent from training. Recovery demonstrations and failure analysis are valuable, but claims of recovery need to be bounded by the specific faults and trials actually tested.

For now, the most defensible takeaway is narrower than a claim of general-purpose robotic autonomy. NVIDIA has released Isaac Lab 3.0 as generally available software, while unveiling GR00T 2 and reporting encouraging controlled results on a defined set of tasks. The model’s demonstrations and metrics suggest meaningful research progress in cross-embodiment transfer and few-shot adaptation. Whether those gains persist with new objects, prolonged use, safety constraints and independent testing remains an open question. For robotics teams, the practical next step is to inspect the model artifacts as they become available, reproduce published task protocols and measure performance on their own hardware and operating conditions.

FAQs

What did NVIDIA announce at CoRL?

NVIDIA introduced GR00T 2, presented Isaac Lab 3.0 in its robotics research and developer lineup, and described the Isaac Lab-Arena preview, benchmarks and data workflows. Isaac Lab 3.0 v3.0.0 had already been published as generally available on Oct. 20, according to its GitHub release page.

What does “40 of 50 tasks” mean?

NVIDIA says GR00T 2 succeeded at least once on 40 of 50 evaluated tasks without task-specific demonstrations. This is the number of tasks with at least one success, not a 80% average success rate across trials. NVIDIA separately reports average success of 71% with one demo per task and 95% with five.

Is GR00T 2 available to download now?

NVIDIA’s official research page links to the model and paper and states that code will be available in December. The cited materials do not establish full artifact availability on Nov. 10, 2026.

Is the 2.5× result a general Isaac Lab speedup?

No. NVIDIA reports up to 2.5× training throughput in a vision-based dexterous manipulation benchmark when comparing Isaac Lab 3.0 with Newton against Isaac Lab 2.3 with PhysX. It is a bounded benchmark result, not a universal speed guarantee.

References