Nvidia GR00T 2 Achieves 95% Success Across 50 Robot Tasks

Nvidia's GR00T 2 achieved 95% average task success across 50 bimanual robot tasks when given just five demonstrations per task. That number, released at the Conference on Robot Learning (CoRL 2026) in Austin on November 9, marks the clearest evidence yet that robot foundation models have crossed the threshold from academic curiosity to commercial deployment infrastructure. The closest prior benchmark on comparable bimanual tasks was in the low 70s, and it required days of bespoke training per task rather than a 30-minute fine-tune.

From Vision-Language to World Action: What GR00T 2 Actually Is

The jump from GR00T N1 to GR00T 2 is an architectural one, not just a capability increment. GR00T N1, released in April 2025, was a vision-language-action model: it took text descriptions and camera images as input, and output joint angles. GR00T 2, as described in Nvidia's official CoRL 2026 blog post, is a world action model (WAM), a newer class that jointly predicts future video frames and robot actions in a shared latent space. The difference matters because predicting what the world should look like next forces the model to build an internal physics simulator, rather than pattern-matching text descriptions to motor commands through a language bottleneck.

Built on Nvidia Cosmos, the 2-billion-parameter model was pre-trained on a dataset that no single robotics lab could have assembled independently: 402,730 hours of human egocentric video, 10,184 hours of UMI gripper demonstrations across 2,897 distinct tasks, and 23,500 hours of robot teleoperation data. The Human-1M-100K subset, covering 100,000 hours of egocentric video, is already on Hugging Face for researchers to build on before the full GR00T 2 weights drop in December. The scale of this pre-training corpus is the structural advantage that academic labs and even well-funded robotics startups cannot replicate without Nvidia's data collection infrastructure.

At CoRL, Jim Fan, co-lead of Project GR00T, confirmed in advance of his keynote that the teaser footage showed real hardware operating autonomously at normal speed, neither sped up nor teleoperated, as reported by CryptoBriefing. Fan made that statement in writing before the keynote began, making the footage falsifiable in real time by an audience of robotics researchers. That matters because prior robot demo videos from several companies had accelerated or curated footage, and the academic community had grown skeptical of benchmark claims that only the company itself could verify.

The lighter GR00T 2 Turbo variant uses four times fewer spatial video tokens, enabling deployment on a Jetson AGX Thor at 124 milliseconds per prediction. A SONIC low-level controller runs at 50 Hz on a Jetson Orin NX, bringing the full inference stack to hardware available for under $3,000 at OEM pricing. The entire prior generation of commercial humanoid controllers required either a tethered desktop workstation or a dedicated backpack compute unit. If Turbo's latency holds across tasks beyond Nvidia's internal demos, it removes the hardware barrier that has kept robot foundation models out of mobile and unstructured factory-floor deployments where a power cable is impractical.

Nvidia also released Isaac Lab 3.0 at the same conference, delivering up to 2.5 times the training throughput of Isaac Lab 2.3 on vision-based dexterous manipulation benchmarks. Skild AI's Steel solver, built on Newton physics, demonstrated 3.9 million environment steps per second on a single RTX PRO 6000 Blackwell card, up to 13 times faster than PhysX on physics stepping alone. AlphaNIST, trained inside Isaac Lab, learned all 20 tasks on the NIST Assembly Task Board with a single policy, reaching over 98% training success versus 81% for per-task policies and using roughly 39 times less GPU compute time.

What 71% at One Demo and 95% at Five Actually Mean in Practice

Robot task success rates are only meaningful when you know what the tasks are. Nvidia's 50-task evaluation suite tested a bimanual robot using Sharpa dexterous hands across assembly, manipulation, and sequencing challenges that prior foundation models addressed only in narrow, task-specific training runs. At zero demonstrations, GR00T 2 completed 40 of the 50 tasks at least once, averaging 40% success across the full set without any task-specific conditioning, as Humanoids Daily reported from the CoRL announcement. With a single demonstration per task, success climbed to 71%. With five demonstrations, it reached 95%, with 37 of the 50 tasks succeeding in every single trial across the full evaluation run.

The standard benchmark for pre-WAM robot learning required a bespoke policy trained for three to five days of GPU compute time, delivered 70% to 80% success on a narrow task range, and failed immediately on any new hardware embodiment. GR00T 2 adaptation uses a 30-minute LoRA fine-tune on a single RTX 5090. The comparison is not just about speed: it is the difference between a one-task policy that breaks when the lighting changes and a general model that handles new embodiments without retraining from scratch. The 30-minute figure assumes five high-quality demonstrations already collected; the demo collection itself can take 30 to 60 minutes more depending on the task, so a full new-task onboarding time of roughly one hour is more accurate.

Specific tasks Nvidia published include: toy truck assembly from human data alone, achieving 60% success on action sequences up to four minutes long. Memory module reseating on a GB300 server chassis achieved 85% success. Popcorn serving on a Unitree G1 running GR00T 2 Turbo logged 48 consecutive successes in 48 trials. MicroAGI fine-tuned the model with reinforcement learning to handle Ethernet cable swapping in IT infrastructure racks. These are not toy demos: they represent the tier of tasks that factory and logistics operators have been waiting for foundation models to handle at commercially acceptable rates. The GB300 memory module result is the one that Foxconn's manufacturing engineers will read most carefully, because server assembly is the first production environment where GR00T 2 is being evaluated under real production conditions.

GR00T 2 also occupies the top three positions on four public robot benchmarks: RoboCasa365, RoboLab, RoboDojo, and MolmoSpaces, across both the vision-language-action and world action model evaluation categories. EgoHand, a companion open model for 3D hand motion reconstruction built with Noitom Robotics data, accumulated more than 40,000 Hugging Face downloads in its first two weeks, outpacing most robotics model releases from academic labs in 2024 and 2025 combined. That download rate suggests the developer community is not waiting for the December GR00T 2 release and is already building on Nvidia's tooling ecosystem.

Foxconn, Siemens, and Nexuni Are Not Waiting for December

The commercial side of GR00T 2 is in the four early adopters Nvidia named at CoRL 2026. Foxconn, the world's largest contract manufacturer with approximately 1.1 million workers on payroll, is already running early versions of GR00T 2 on server manufacturing automation lines in its own facilities. Siemens, with $80 billion in annual revenue and existing automation contracts across hundreds of industrial customers in Europe and Asia, is evaluating the model for factory use cases. Nexuni is deploying it for laundry handling in commercial laundry operations. MicroAGI is using a post-trained version with RL for Ethernet cable swapping in IT infrastructure racks.

Foxconn's choice of server manufacturing as the testing ground is strategically calculated. Server assembly involves tight tolerances, dense cable routing, and mixed-weight components that have historically required a skilled human technician. A mistake on a rack-mount server costs $3,000 to $15,000 per unit in scrapped or damaged components. If GR00T 2 passes Foxconn's production threshold, it will almost certainly be announced during an earnings call or in a customer event in early 2027. The absence of a Foxconn announcement by February 2027 would itself be a data point about how well the lab benchmark translated to a production line. Foxconn also participated in Agility Robotics' Nasdaq IPO as a lead investor in the $201 million public tranche, suggesting that its humanoid robot evaluation program is operating across multiple foundation model vendors simultaneously.

Siemens' evaluation carries a different commercial weight. Siemens sells automation hardware and software to thousands of factory operators globally. A positive internal evaluation does not mean Siemens will resell GR00T 2 directly, but it does mean Nvidia's model stack is being tested against Siemens' internal standards for industrial robot behavior, which include fault handling, repeatability, audit-logging, and safety certification requirements that academic benchmarks rarely address. If Siemens certifies GR00T 2 for use in a Siemens-branded automation product, the customer reach extends to industries Nvidia's direct sales team has never approached.

The WAM Race: Physical Intelligence, OpenAI, and the Gap Nvidia Just Defined

The world action model category is new enough that competitive comparisons are still being established, which is precisely why Nvidia moved to define the benchmark on its own terms at CoRL. Physical Intelligence (PI), the best-funded dedicated robotics AI startup at approximately $2.5 billion raised, uses a different architecture: its PI-0 and PI-0.5 models apply diffusion-based action generation and flow matching rather than joint video-action prediction. PI's proponents argue that video prediction compounds the embodiment gap, because human hand kinematics differ from robot gripper kinematics in ways that a video prediction model cannot fully reconcile without explicit embodiment modeling. Nvidia's 95% rate is on a Sharpa-hand bimanual system, not a universal benchmark, and PI has not publicly compared its models on the same 50-task evaluation, which makes direct comparison impossible at this point.

DeepMind's RT-X family and Stanford's Mobile ALOHA generated strong academic attention through 2025 but lack Nvidia's distribution leverage. Boston Dynamics, now owned by Hyundai, has not released public manipulation benchmark data for the Atlas platform despite deploying Atlas in automotive manufacturing pilots with BMW. Figure AI and Apptronik have both announced foundation model integrations in 2025 without publishing third-party-verified numbers on a comparable benchmark.

The more strategically interesting absent competitor is OpenAI. OpenAI quietly rebuilt its robotics team through 2025 and has disclosed nothing publicly about a robot foundation model. If and when OpenAI announces a robot model, the comparison with GR00T 2's 95% will be immediate and very public. Nvidia's decision to open-source GR00T 2 weights in December pre-empts that by establishing the benchmark baseline before any challenger can reset it. The WAM API Nvidia also released at CoRL raised success on a YAM arm instruction set from 40% to 92% by using a reasoning model to annotate camera images before passing them to the action model, and the ASPIRE agent framework enabled 292 successful toy assembly runs out of 300 across three toy types, compared to 71 to 84 successes per toy type without it. Fan's claim that locally hosted Flora is 125 times cheaper than GPT-6 Astra at the same performance level has not been independently verified, but the direction is consistent with the broader trend of edge-compute costs falling faster than cloud-inference costs.

The Bear Case: Lab Benchmarks and Factory Floors Are Different Universes

The 95% number is impressive, but critics argue the evaluation conditions leave important gaps that factory operators will discover before Nvidia does. The 50-task suite was designed by Nvidia's own team on hardware Nvidia selected. The Sharpa dexterous hands used in the evaluation are not widely deployed in commercial factories. High laboratory success rates on curated hardware have a documented history of falling 20 to 40 percentage points when transferred to production environments with varying illumination, thermal expansion in mechanical joints, part tolerance variation from supplier batches, and ambient vibration from adjacent equipment. No independent lab has published a GR00T 2 evaluation because the full weights are not yet publicly available, so no independent lab can run one.

The bear case is also in the training data composition. GR00T 2's pre-training is dominated by human egocentric video rather than robot data, which creates an embodiment mismatch in dexterous tasks. Humans and robot hands have different kinematics, different grip surfaces, and no tactile feedback in the video signal. The model learns from watching human hands and infers robot actions. Physical Intelligence's diffusion-based approach was specifically designed to address this embodiment gap without video prediction as an intermediary, and its proponents argue that video prediction adds an extra inference step that compounds error rather than resolving it.

Skeptics also point out that the LoRA fine-tune process, while fast, still requires five high-quality, consistent demonstrations per task and an RTX 5090. At a retail price above $1,999, an RTX 5090 is not standard equipment in mid-tier manufacturing facilities, particularly in Southeast Asia and Latin America, where the economic case for humanoid deployment is the strongest. The five-demos-to-95% result also says nothing about how performance degrades when demonstrations are inconsistent, recorded by different operators, or collected under varying conditions, which is the standard situation in real production onboarding.

Four Things to Watch Before the December Open-Source Release

The December GR00T 2 weight release will determine whether the model becomes the robot foundation model that practitioners actually deploy, or another well-benchmarked model that sits on Hugging Face. Four markers define the outcome, as covered in CryptoBriefing's CoRL coverage and the DEV Community's technical review of the Isaac Lab 3.0 release.

First and most critical: the license terms. Full Apache 2.0 or MIT licensing enables any company to build a commercial product without a revenue-sharing agreement with Nvidia. A research-only or non-commercial restriction effectively keeps GR00T 2 out of commercial robotics for the near term, regardless of how well it benchmarks. The license announcement deserves more attention than the benchmark numbers, because benchmark rankings only matter if you are allowed to use the model in a product.

Second: an independent benchmark verification from a university lab. CoRL peer review assessed the methodology, not the absolute numbers. A group at Carnegie Mellon, MIT, or ETH Zurich running GR00T 2 on the same 50 tasks with their own hardware will establish whether Nvidia's internal figures are reproducible. If verification comes back below 85% on independent hardware, the narrative will shift. Watch for a preprint on arXiv before year-end, since several CoRL attendees have the Isaac Lab 3.0 tooling needed to begin replication now that it is publicly available.

Third: Foxconn's Q4 2026 earnings call in January 2027. If Foxconn management describes GR00T 2 results in production, it will be the first third-party commercial validation in a live manufacturing environment. Foxconn has historically been quiet about internal technology evaluations until they result in a product decision, so a mention on the earnings call would indicate the pilot cleared Foxconn's production threshold.

Fourth: Physical Intelligence's response timeline. PI has consistently followed major open-source releases with capability publications of its own. Its next public update, expected before the end of Q4 2026, will likely include a direct response to the WAM benchmark claims and potentially a head-to-head evaluation on a shared task set. The Isaac Lab-Arena 0.4 framework, also released at CoRL, provides a standardized evaluation environment that any team can use to run a comparable benchmark, so the conditions for a credible response from PI or another competitor exist before the GR00T 2 weights are even public.

The question is no longer whether a robot can learn a new task in an afternoon. It's whether Nvidia's 95% survives contact with a factory floor in Guadalajara.

---