NVIDIA GR00T 2 Achieves 95% Success with Five Demonstrations; Foxconn and Siemens Enter Production-Line Testing, but Weights Arrive Only in December

On November 9, 2026, NVIDIA released GR00T 2 at the Conference on Robot Learning (CoRL 2026) in Austin, Texas, officially defining it as an “open frontier world-action model” based on the Cosmos architecture. According to NVIDIA’s official blog, the model achieved an average success rate of 95% on 50 evaluation tasks after five task demonstrations; under the single-demonstration condition, it reached 71%; with no demonstrations at all, its average success rate was 40%, but it was able to complete at least 40 of those tasks. Foxconn has applied it to server manufacturing lines, Siemens is evaluating industrial application scenarios, and Nexuni is using it for robotic laundry operation tests.

The technical core of GR00T 2 is the use of a single diffusion transformer to simultaneously generate future video frames and motor commands. According to mlq.ai, citing NVIDIA’s research page, the model reads recent camera frames, natural-language instructions, and the robot’s current state, and outputs a short-horizon prediction sequence in a single forward pass, integrating video prediction and action planning. The direct significance of this architecture is that skill representations are not tied to a specific robot platform—NVIDIA says it can transfer skills learned from human demonstrations across different embodiments without robot demonstrations. The scale of pretraining data is considerable: according to NVIDIA’s research page, the training set includes 402,730 hours of egocentric human video, 10,184 hours of handheld gripper demonstrations, and 23,500 hours of robot teleoperation data, covering multiple robotic arms and the Unitree G1 humanoid platform, with 100,000 hours of human video already released as an open dataset on Hugging Face. According to mlq.ai, adding human video alone to the training set increased the model’s average progress rate on 72 tasks from 50% to 70%.

From an engineering practicality perspective, the adaptation cost behind the 95% figure with five demonstrations is the key variable for judging its commercial value. According to mlq.ai, the robot operation time used in the five demonstrations was less than three minutes, after which about 30 minutes of LoRA fine-tuning on a single RTX 5090 was enough to adapt to the task. These figures do not represent a fully autonomous robot, but rather a base adaptation layer with a high learning starting point that can quickly align to new tasks at minimal cost—analogous to fine-tuning a large language model for a specific domain rather than out-of-the-box general capability. According to humanoidsdaily, the Turbo variant of GR00T 2 uses a 2-billion-parameter transformer, optimized for edge deployment, with an inference latency of 124 milliseconds on the Jetson Thor robot compute platform.

However, the selection criteria for these 50 evaluation tasks have not yet been fully disclosed, which is the most important uncertainty in the current version. According to cellcog.ai, NVIDIA has not released the task list, and external researchers cannot independently verify whether the task distribution represents the manipulation diversity of the open world. According to mlq.ai, citing NVIDIA research materials, these tasks were selected from tasks the model “is capable of learning.” GR00T 2 also ranks in the top three in both the vision-language-action and world-action model categories on four benchmark leaderboards: RoboCasa365, RoboLab, RoboDojo, and MolmoSpaces.

Isaac Lab 3.0: A More Convincing Set of Comparative Data

Released alongside GR00T 2, Isaac Lab 3.0 offers more quantifiable reference value at the engineering level. According to dev.to, in dexterous manipulation benchmarks, Isaac Lab 3.0 improves training throughput by 2.5 times compared with version 2.3; the core improvement is the introduction of the Newton physics engine, adding accurate simulation capabilities for cables, cloth, and multi-jointed robotic hands.

The AlphaNIST case provides the most convincing specific comparative data to date. According to dev.to, this system built on Isaac Lab uses a single policy across all 20 tasks on the NIST assembly task board, with a training success rate exceeding 98%, while per-task independent policies reach only 81%; the GPU time used by the single policy is about 1/39 of that used by per-task policies. The significance of these figures goes beyond accuracy improvement: they reveal an engineering economics principle—when the generalization capability of a foundation model is strong enough, a single general policy outperforms task-specific policies in both accuracy and compute cost.

The “Open” Claim and the Critical December Weight Release Milestone

NVIDIA characterizes GR00T 2 as an “open frontier model,” but there is a gap between this wording and the current reality. According to mlq.ai, as of the launch day the model weights had not been made public, the code is planned for release in December 2026, and the license terms have not been announced to date. For reference, NVIDIA’s previously released Flora series robot models used a non-commercial research license. With neither the weights nor the license in place, the teams able to actually integrate GR00T 2 at this stage are limited to official partners accessing it via API.

From an industry landscape perspective, the deeper logic revealed by this release is NVIDIA’s strategic layout of extending ecosystem control from chips to the model layer: Isaac Lab provides the simulation training environment, GR00T 2 provides the pretrained foundation, Jetson Thor provides edge inference compute, and DexBench and VABAR, two accompanying benchmark frameworks (covering 15 retail/logistics/manufacturing/hospitality tasks and 6 industrial manipulation tasks, respectively), provide evaluation standards—NVIDIA is building a complete closed loop from data and training to evaluation and deployment. For downstream robot vendors, this ecosystem both significantly lowers the development starting point and increases deep lock-in to NVIDIA’s technology stack. Foxconn’s choice of server manufacturing is itself quite telling: server assembly is a relatively structured operating environment, placing it squarely within the reasonable boundary of current robot foundation model capabilities.

From an analytical standpoint, December is the most important milestone for assessing the actual degree of openness of GR00T 2: whether the license permits commercial use, whether the weight release includes complete training configurations, and whether external teams can reproduce the benchmark results on the 50 tasks on independent hardware. These three signals will jointly determine the extent to which GR00T 2 can advance the open ecosystem for robot foundation models, or whether it will mainly serve the circle of partners within NVIDIA’s ecosystem. Siemens is currently in the evaluation stage and has not announced a specific application timetable.