ABAB Pedia

NVIDIA GR00T

NVIDIA GR00T (Isaac GR00T) is a family of foundation models for general-purpose robots, together with a humanoid robot development platform, developed by NVIDIA; it includes the GR00T N series of vision-language-action models and the world action model GR00T 2.

Contents11 sections
Key facts

Project GR00T Announced at GTC

At its GTC conference NVIDIA announced Project GR00T, which it described as a general-purpose foundation model for humanoid robots; GR00T stands for “Generalist Robot 00 Technology.” As NVIDIA envisioned it, robots powered by GR00T would understand natural language, emulate movements by observing human actions, and quickly learn coordination, dexterity and other skills in order to navigate, adapt to and interact with the real world. In his GTC keynote Jensen HuangJensen HuangJensen Huang is a co-founder of NVIDIA who served as its president, chief executive officer, and a director from the company’s founding in 1993 until he resigned from those positions in November 2026. He has a background in electrical engineering education and the semiconductor industry, and under his leadership NVIDIA’s business expanded from graphics chips to GPU general-purpose computing, artificial intelligence, and data center computing platforms.Open the full entry demonstrated several such robots completing a variety of tasks.⁠[1]

Announced at the same time, Jetson Thor is a computer for humanoid robots based on the NVIDIA Thor system-on-a-chip, with a next-generation GPU based on the Blackwell architecture whose transformer engine delivers 800 teraflops of 8-bit floating-point AI performance to run multimodal generative AI models such as GR00T, plus an integrated functional safety processor, a high-performance CPU cluster and 100Gb of Ethernet bandwidth. NVIDIA also updated its Isaac robotics platform with Isaac Lab for reinforcement learning, the compute orchestration service OSMO, Isaac Manipulator for robotic arms and Isaac Perceptor for autonomous mobile robots.⁠[1]

NVIDIA said it was building the platform for leading humanoid robot companies including 1X Technologies, Agility Robotics, Apptronik, Boston Dynamics, Figure AIFigureFigure is a US AI robotics company founded by Brett Adcock in 2022. It develops general-purpose humanoid robots and the Helix control system for industrial logistics and household tasks. Its robot generations include Figure 01, Figure 02 and Figure 03, with production pilots at BMW.Open the full entry, Fourier Intelligence, Sanctuary AI, Unitree Robotics and XPENG Robotics. Jensen Huang said: “Building foundation models for general humanoid robots is one of the most exciting problems to solve in AI today.”⁠[1]

Six GR00T Workflows Presented at CoRL

During the Conference on Robot Learning (CoRL) in Munich, GermanyGermanyGermany, officially the Federal Republic of Germany, is a federal parliamentary republic in central Europe made up of 16 Länder, with Berlin as its capital. The Basic Law was proclaimed and entered into force on May 23, 1949, founding the Federal Republic, and on October 3, 1990, the German Democratic Republic joined the area in which the Basic Law applies, which united the country. The population statistics for 2025, based on the 2022 census, put the population at 83,467,117.Open the full entry, NVIDIA presented Project GR00T as an initiative to develop accelerated libraries, foundation models and data pipelines for the humanoid robot developer ecosystem, and introduced six workflows: GR00T-Gen for building 3D environments with generative AI and OpenUSD, GR00T-Mimic for robot motion and trajectory generation, GR00T-Dexterity for dexterous manipulation, GR00T-Control for whole-body control, GR00T-Mobility for locomotion and navigation, and GR00T-Perception for multimodal sensing; the workflows were then marked “coming soon.”⁠[2]

GR00T N1 Released as an Open Model

At GTC NVIDIA released Isaac GR00T N1, which it called the world’s first open, fully customizable foundation model for generalized humanoid reasoning and skills and the first of a family of models that it would pretrain and release to robotics developers worldwide; NVIDIA tied the work to a global labor shortage that it estimated at more than 50 million people.⁠[3]

GR00T N1 uses a dual-system design inspired by human cognition: “System 2” is a slower, deliberate decision-making model powered by a vision language model that reasons about the environment and instructions to plan actions, and “System 1” is a fast action model, comparable to human reflexes or intuition, that turns the plans into precise, continuous robot movements. The released model was the 2-billion-parameter GR00T N1 2B, whose System 2 is based on NVIDIA-Eagle with SmolLM-1.7B and whose System 1 is a diffusion transformer. Its training data formed a pyramid: internet-scale web data and human videos at the base, synthetic data generated with NVIDIA OmniverseNVIDIA OmniverseNVIDIA Omniverse is a collection of software libraries, developer tools and services for 3D collaboration, physical simulation and industrial digital twins. OpenUSD provides a common scene description so that teams can combine geometry, materials and animation produced in different applications.Open the full entry in the middle, and real robot teleoperation data at the top.⁠[3][4]

The model generalizes across common tasks such as grasping, moving objects with one or both arms and handing items from one arm to the other, and can carry out multistep tasks that need long context, for uses such as material handling, packaging and inspection; developers can post-train it with real or synthetic data for a specific robot or task. The first supported robots were the Fourier GR-1 and the 1X NEO. On simulation benchmarks with 100 demonstrations per task, GR00T N1 2B averaged a 45.0% success rate against 33.4% for Diffusion Policy and 26.4% for BC Transformer; on real Fourier GR-1 tasks it averaged 76.8% with the full data (Diffusion Policy 46.4%) and 42.6% with 10% of the data (Diffusion Policy 10.2%).⁠[3][4]

With the Isaac GR00T Blueprint for synthetic manipulation motion generation, built on Omniverse and the NVIDIA Cosmos Transfer world foundation models, NVIDIA generated 780,000 synthetic trajectories in 11 hours, the equivalent of about 6,500 hours, or nine continuous months, of human demonstration data; combining this synthetic data with real data improved GR00T N1’s performance by 40% compared with real data alone. GR00T N1 training data and task evaluation scenarios were made available for download from Hugging Face and GitHub, and the dataset was released as part of a larger open-source physical AI dataset.⁠[3]

In his keynote Jensen Huang showed a 1X humanoid robot autonomously tidying a home with a policy post-trained on GR00T N1, the result of an AI training collaboration between 1X and NVIDIA, and he also brought Disney’s BDX droids on stage. Bernt Børnich, chief executive of 1X Technologies, said that 1X had fully deployed the model on its NEO Gamma robot with minimal post-training data. Developers with early access included Agility Robotics, Boston Dynamics, Mentee Robotics and NEURA Robotics. NVIDIA, Google DeepMindGoogle DeepMindGoogle DeepMind is Google’s artificial intelligence research organization, formed in April 2023 by merging DeepMind, which Google acquired in 2014, with the Brain team from Google Research, with Demis Hassabis as CEO until August 2026, when he became chair. It developed AlphaGo, AlphaFold, and the Gemini family of models, and Hassabis and John Jumper received the 2024 Nobel Prize in Chemistry for AlphaFold’s protein structure prediction work.Open the full entry and Disney Research also announced that they were developing Newton, an open-source physics engine. Jensen Huang said: “The age of generalist robotics is here.”⁠[3]

GR00T N1.5 Announced at Computex

At Computex NVIDIA announced Isaac GR00T N1.5, the first update to GR00T N1, which adapts better to new environments and workspace layouts, recognizes objects from user instructions and has a higher success rate on material-handling and manufacturing tasks such as sorting or putting away objects; the model was to be made available for download on Hugging Face. Early adopters of the GR00T N models included AeiRobot, Foxlink, Lightwheel and NEURA Robotics.⁠[5]

The Isaac GR00T-Dreams blueprint, released at the same time, has developers post-train Cosmos Predict world models for their robot, generate from a single image videos of the robot performing new tasks in new environments, and extract action tokens from them as training data, which NVIDIA calls “neural trajectories”; NVIDIA Research used it to generate training data for GR00T N1.5 in 36 hours, compared with nearly three months of manual human data collection. The GR00T-Mimic blueprint released in March, by contrast, augments a small number of human demonstrations using Omniverse and Cosmos.⁠[5]

GR00T N1.6 Technical Report Published

NVIDIA’s GEAR team published a technical report on GR00T N1.6, with 58 named authors (besides the GEAR Team) including Linxi “Jim” Fan, Yuke Zhu, Jan Kautz and Scott Reed. N1.6 outperformed N1.5 on simulated manipulation benchmarks and on real bimanual YAM, AgiBot Genie-1 and Unitree G1 robots; its model card is named GR00T-N1.6-3B and its code is in the Isaac-GR00T repository on GitHub.⁠[6]

N1.6 deepened the diffusion transformer from N1.5’s 16 layers to 32, removed the 4-layer post-VLM transformer adapter of N1.5 and instead unfroze the top 4 layers of the vision language model during pretraining; for most embodiments it predicts state-relative action chunks rather than absolute joint angles or end-effector positions, and it uses test-time and train-time real-time chunking (RTC) for smoother asynchronous rollouts. Its vision language backbone is, in one account, an internal NVIDIA Cosmos-2B VLM variant that supports flexible resolution and encodes images at their native aspect ratio; in another, it is Eagle.⁠[6][7]

Pretraining added several thousand hours of teleoperated data to the N1.5 mixture, from bimanual YAM, AgiBot Genie-1, a simulated Galaxea R1 Pro in the BEHAVIOR suite and Unitree G1 whole-body loco-manipulation, and ran for 300,000 steps with a global batch size of 16,384; robot-specific post-training typically ran 10,000–30,000 steps with a batch size of 1,000 or less. Because the model converges faster, the team strengthened state regularization, data augmentation and co-training with pretraining data to prevent overfitting, and recommended DAgger when the model underperforms on real robots; multi-task language following and out-of-distribution generalization remained challenging.⁠[6]

GR00T N1.7 Early Access and GR00T N2 Preview

At GTC NVIDIA announced that GR00T N1.7 was available in early access with commercial licensing, bringing generalized robot skills including advanced dexterous control to production-ready robot deployments; AGIBOT, Humanoid, LG Electronics, NEURA Robotics and Noble Machines were adopting Isaac GR00T N models to accelerate industrial deployments of their humanoids.⁠[8]

In his keynote Jensen Huang previewed GR00T N2, a next-generation robot foundation model based on DreamZero research and a new world action model architecture; NVIDIA said that robots using it succeeded at new tasks in new environments more than twice as often as with leading vision language action models, that it ranked first on the MolmoSpaces and RoboArena leaderboards for generalist robot policies, and that it was expected to be available by the end of 2026. The model previewed at GTC 2026 and based on DreamZero world action model research was later released as GR00T 2.⁠[8][9]

Isaac GR00T Reference Humanoid Robot Announced

At GTC Taipei in Taipei NVIDIA announced the Isaac GR00T reference humanoid robot for academic research. Built on the Unitree H2 Plus humanoid, it is nearly 6 feet tall, weighs 150 pounds and has 31 degrees of freedom in the body, plus dual Sharpa Wave tactile five-finger hands with 22 degrees of freedom each, for 75 in total; it carries a head-mounted stereo camera, wrist cameras and an inertial measurement unit, has arm torque of up to 120 N·m and leg torque of up to 360 N·m, a rated arm payload of 7 kg and a peak payload of 15 kg, and an onboard NVIDIA Jetson AGX Thor T5000 with a Blackwell GPU (2,070 FP4 TFLOPS), a 14-core Arm CPU and 128 GB of unified memory at a configurable 40–130 W; a 15 Ah, 0.972 kWh battery gives about three hours of operation.⁠[10]

The accompanying Isaac GR00T platform comprises Isaac Teleop for capturing demonstration data, the Isaac GR00T open foundation models for reasoning and multitask behavior, Isaac Sim and Isaac Lab for simulation and training, Isaac ROS middleware for deployment and Jetson Thor for on-robot inference. The robot was expected to be available from Unitree in late 2026, and a GR00T reference workflow for the Unitree G1 was expected soon on GitHub and Hugging Face. Ai2, ETH Zurich, Stanford Robotics Center and UC San Diego’s Advanced Robotics and Controls Laboratory planned to use the design, and NVIDIA Research would use it to advance GR00T models, frameworks and hardware. Jensen Huang called it “a single, open platform to make breakthrough discoveries.”⁠[10]

GR00T 2 Unveiled at the Conference on Robot Learning

The 2026 Conference on Robot Learning was held in Austin, TexasTexasTexas is a U.S. state in the south-central United States, with its capital at Austin and its largest city Houston. It was admitted to the Union on December 29, 1845, as the 28th state, and the 2020 census (April 1, 2020) counted 29,145,505 residents.Open the full entry, from November 9 to 12, with the conference on November 9–11 and workshops on November 12. Jim Fan, NVIDIA’s director of robotics and a distinguished scientist, unveiled GR00T 2 in a keynote on the afternoon of the first day, and NVIDIA’s booth demonstrated GR00T, Isaac Lab and Newton.⁠[11][12][13]

GR00T 2 is a family of “world-action models” that jointly predict future video and robot actions, whereas earlier GR00T releases were vision-language-action models. A single diffusion transformer denoises future video and an action chunk together, and small per-robot heads convert the output into motor commands; the inputs are about 0.8 seconds of recent video, the language instruction and the robot state, with video at 320×224 from a head camera and two wrist cameras. Built on NVIDIA Cosmos, the model has 2 billion parameters and works with parallel-gripper arms, dexterous hands and humanoid robots. Jim Fan described it as a fast “System 1” skill model that a larger reasoning model can direct, saying: “Everything is going to be a skill, including System 1.”⁠[14][12]

The pretraining data comprise 402,730 hours of egocentric human video (Human-1M), 10,184 hours of handheld UMI gripper demonstrations across 2,897 tasks and 23,500 hours of robot teleoperation, with robot data from YAM and Franka parallel-gripper arms, Sharpa and Wuji dexterous hands and the Unitree G1 humanoid; 100,000 hours of the human video were made public on Hugging Face as Human-1M-100K. On 72 Dexter tasks, adding human video raised task progress from 64% to 77% in video rollouts and from 50% to 70% on the real robot.⁠[14][12]

Skills transfer across embodiments: a robot with two five-fingered hands trained only on human demonstrations placed filter paper in a laboratory funnel in 65% of trials; a YAM arm trained only on UMI data assembled a toy dump truck end to end in 60% of trials, in sequences of up to four minutes; and a Tianji robot with ROBOTERA hands, trained on DexUMI data, learned to prepare banana yogurt, scoop granola and strike a match to light a lamp.⁠[14]

On a bimanual robot with Sharpa hands, NVIDIA tested 50 tasks the target robot had not seen: zero-shot average success was 40%, with 40 of the 50 tasks succeeding at least once; with one demonstration per task it rose to 71%, and with five demonstrations per task (under three minutes of robot data) to 95%, with 37 tasks succeeding in every trial; adaptation used a 30-minute LoRA fine-tune on a single RTX 5090. Two 22-degree-of-freedom Sharpa hands seated GB300 memory modules with 85% success and assembled a toy truck with 65%; three contact-rich insertion tasks improved from 40–76% to 88–96% with real-world reinforcement learning by GRPO-induced self-imitation.⁠[14][12][15]

A reasoning agent can act as “System 2” and direct GR00T 2 through a WAM API: toy plane, tractor and Hot Wheels car assembly succeeded in 98, 97 and 97 of 100 trials respectively (292 of 300 in total), against 71–84 per toy without the agent; inserting five Ethernet cables succeeded in 19 of 20 runs and five chargers in 20 of 20; the agent recovers from errors with code-as-policy skills learned with ASPIRE. In the demonstrations the reasoning model was GPT-6 AstraGPT-6GPT-6 is an OpenAI model family introduced in 2026, comprising Astra, Sol and Luna, followed by GPT-6.1 Sol, GPT-6.1 Terra (GPT-6.1 Astra was not released), the cybersecurity research model GPT-6 Cyber, and GPT-6.2 Astra and GPT-6.2 Sol. Its applications include conversation, coding, computer use and professional work, with capabilities and availability varying across models and products.Open the full entry. Jim Fan said: “Neither Astra nor GR00T can do this alone, and you need both.”⁠[14][12]

GR00T 2 Turbo uses a 2-billion-parameter joint transformer and a higher-compression video encoder with four times fewer spatial video tokens; it runs on the Jetson AGX Thor onboard the Unitree G1 at 124 ms per prediction (38 ms on an RTX 5090), with 1.6-second action chunks planned about once per second, while the SONIC whole-body controller runs at 50 Hz on a separate Jetson Orin NX; latency is 66–86% lower than eager BF16 execution. On the G1 it scooped and served popcorn in 48 of 48 trials and cleared a table of diverse objects with 93% success; with simulated data, Turbo’s out-of-distribution success on the G1 rose from 6.7% to 63.3%, while GR00T 1.7 rose to 40.0%.⁠[14][12]

NVIDIA said GR00T 2 ranked among the top three in both the vision-language-action and world-action-model categories of the RoboCasa365, RoboLab, RoboDojo and MolmoSpaces leaderboards. Early users testing GR00T 2 included Nexuni (laundry handling), FoxconnFoxconnFoxconn is the international name of Hon Hai Technology Group, an electronics manufacturer founded by Terry Gou in Taiwan in 1974. Its listed company is Hon Hai Precision Industry Co., Ltd., traded on the Taiwan Stock Exchange as 2317. Young Liu became chairman in 2019.Open the full entry (server manufacturing automation), Siemens (evaluating industrial use cases) and MicroAGI (swapping Ethernet cables with a YAM arm). Isaac Lab 3.0, released at the same time, switches between the PhysX and Newton physics solvers and adds a reinforcement post-training library for foundation models such as GR00T.⁠[13][12]

The code, the GR00T 2 and GR00T 2 Turbo weights and the cross-embodiment adaptation training recipe were planned to be open-sourced in December 2026, and license terms for the weights had not been published; the SimFoundry 1.0 real-to-sim pipeline, the ASPIRE agent framework, the Human-1M-100K dataset and hardware designs for the UMI gripper, the DexUMI exoskeleton and a flexible robot neck were released at the same time.⁠[14][12]

Platform, Models and Licensing as of November 2026

As of November 2026 NVIDIA defined Isaac GR00T as an open reference platform for general-purpose humanoid robots that combines data and pipelines, robot foundation models, simulation frameworks, middleware, accelerated runtime libraries and Jetson Thor for real-time inference and control, comprising the reference platform (with GR00T-Ready Evaluation, the Agile ONE SN reference robot and Halos Safety for Humanoids), reference workflows and open foundation models. The family’s latest open vision-language-action model at the time was GR00T 1.7, with 3 billion parameters, a Cosmos backbone and a commercial license, available on GitHub and Hugging Face, running at up to 25 Hz on Jetson Thor with TensorRT NVFP4 quantization and usable through LeRobot.⁠[16][9]

The NVIDIA/Isaac-GR00T repository on GitHub lists N1.7 as its General Availability release, with Cosmos-Reason2-2B (Qwen3-VL architecture) as the vision language backbone, GR00T-N1.7-3B as the base checkpoint and further checkpoints such as LIBERO, DROID, SimplerEnv-Bridge and SimplerEnv-Fractal, while N1.5 and N1.6 are kept on separate branches as earlier releases; the code is under the Apache 2.0 license and the weights under the NVIDIA Open Model License. On November 21, 2026, the repository had about 8,400 stars and 1,500 forks. The GR00T N1.5-3B and GR00T N1.6-3B weights carry a non-commercial license for research and evaluation only. NVIDIA said that as of August 2026 GR00T models had been downloaded more than 274,000 times and the GR00T X-Embodiment Sim dataset more than 10 million times.⁠[7][9]

NVIDIA’s listed humanoid robotics ecosystem partners include 1X, AgiBot, Agile Robots, Agility Robotics, Apptronik, Boston Dynamics, Dexmate, Enactic, FieldAIFieldAIRobot autonomy software company.Open the full entry, Figure, Fourier, Foxconn, Galaxea, Galbot, Geely, Hexagon, Humanoid, LG, Lightwheel, Mentee Robotics, NEURA Robotics, Nexuni, Noble Machines, Sanctuary AI, Siemens, Sharpa, Skild AI, Solomon, Techman Robot, Unitree Robotics and X-Humanoid.⁠[16]

Related Organizations, Websites, and Public Accounts

Sources16 sources

  1. NVIDIA Announces Project GR00T Foundation Model for Humanoid Robots and Major Isaac Robotics Platform Update (opens in a new window)nvidianews.nvidia.com↑
  2. NVIDIA Advances Robot Learning, Humanoid Development With New AI and Simulation Tools (opens in a new window)blogs.nvidia.com↑
  3. NVIDIA Announces Isaac GR00T N1 — the World’s First Open Humanoid Robot Foundation Model — and Simulation Frameworks to Speed Robot Development (opens in a new window)nvidianews.nvidia.com↑
  4. Accelerate Generalist Humanoid Robot Development with NVIDIA Isaac GR00T N1 (opens in a new window)developer.nvidia.com↑
  5. NVIDIA Powers Humanoid Robot Industry With Cloud-to-Robot Computing Platforms for Physical AI (opens in a new window)nvidianews.nvidia.com↑
  6. GR00T N1.6: An Improved Open Foundation Model for Generalist Humanoid Robots (opens in a new window)research.nvidia.com↑
  7. NVIDIA/Isaac-GR00T: NVIDIA Isaac GR00T N1.7 - A Foundation Model for Generalist Robots. (opens in a new window)github.com↑
  8. NVIDIA and Global Robotics Leaders Take Physical AI to the Real World (opens in a new window)nvidianews.nvidia.com↑
  9. LLM Info — NVIDIA Isaac GR00T (opens in a new window)developer.nvidia.com↑
  10. NVIDIA Announces NVIDIA Isaac GR00T Reference Humanoid Robot for Academic Research (opens in a new window)investor.nvidia.com↑
  11. CoRL 2026 | Conference on Robot Learning | Austin, Texas (opens in a new window)nvidia.com↑
  12. “Robotics Is No Longer an Elitist Sport”: NVIDIA Unveils GR00T 2 and Open-Sources the Stack Behind It (opens in a new window)humanoidsdaily.com↑
  13. NVIDIA launches GR00T 2 and Isaac Lab 3.0 at CoRL (opens in a new window)therobotreport.com↑
  14. GR00T 2: Scaling Robot Foundation Models for Generalization, Dexterity, and Adaptation (opens in a new window)research.nvidia.com↑
  15. Nvidia发布GR00T 2世界动作模型,机器人可从人类视频学习技能 [Nvidia releases GR00T 2 world action model; robots can learn skills from human videos] (opens in a new window) (in Chinese)ababnews.com↑
  16. Isaac GR00T | Humanoid Robot Development Platform (opens in a new window)nvidia.com↑