NVIDIA opens Isaac GR00T N1, a 2.2-billion-parameter humanoid robot model, at GTC 2025 and teams with Google DeepMind and Disney Research on the Newton physics engine

NVIDIA released Isaac GR00T N1 on 18 March 2025, an open foundation model for humanoid robots that developers can download from Hugging Face and GitHub and post-train for their own machines. Chief executive Jensen Huang unveiled it in his GTC keynote at the SAP Center in San Jose, which began at 10:00 a.m. Pacific time, and NVIDIA called it the world's first open, fully customizable foundation model for generalized humanoid reasoning and skills. In the same release the company announced Newton, an open-source physics engine it is developing with Google DeepMind and Disney Research, a blueprint for generating synthetic manipulation data and an open physical AI dataset.

The model is small by language-model standards. NVIDIA's technical report, a preprint posted to arXiv as paper 2503.14734 on 18 March and revised on 27 March, puts it at 2.2 billion parameters, of which 1.34 billion sit in the vision-language component. The Hugging Face listing, dated 17 March 2025, carries the checkpoint as GR00T-N1-2B. Pretraining used about 50,000 H100 GPU hours on up to 1,024 GPUs, according to the preprint. On an NVIDIA L40 GPU the reasoning module runs at 10 Hz, and the action module samples a chunk of 16 actions in 63.9 milliseconds, for motor commands at 120 Hz.

NVIDIA describes the design in terms borrowed from human cognition. System 2, a vision-language model built on NVIDIA's Eagle-2 with a SmolLM2 language model and a SigLIP-2 image encoder, reads camera images and a spoken or typed instruction and plans what to do. System 1, a diffusion transformer trained with flow matching, turns that plan into continuous joint commands. The two are trained together end to end. NVIDIA said the model generalises across common tasks such as grasping, moving objects with one or both arms and handing items from one arm to the other, and can handle multistep tasks that need long context. Its developer blog said post-training needs a single RTX A6000 or GeForce RTX 4090, and that the result can run on an RTX A6000 or a Jetson AGX Orin module.

The early-access list named 1X Technologies, Agility Robotics, Boston Dynamics, Mentee Robotics and NEURA Robotics. On stage, Huang showed a 1X humanoid tidying a home autonomously with a policy post-trained on GR00T N1, which NVIDIA said came out of joint work between the two companies. Bernt Børnich, 1X's chief executive, said in the release that 1X develops its own models but that GR00T N1 gave a significant boost to robot reasoning and skills with little post-training data.

The training recipe is what NVIDIA calls a data pyramid. At the base sit human videos from public collections such as Ego4D and EPIC-KITCHENS. In the middle are synthetic data, including 827 hours of generated video trajectories and 780,000 simulated trajectories produced with DexMimicGen. At the top are real robot recordings, including 88 hours of NVIDIA's own teleoperation on Fourier's GR-1 humanoid, Open X-Embodiment datasets and 140,000 trajectories from AgiBot's AgiBot-Alpha set. In the preprint's tests GR00T N1 beat Diffusion Policy, a widely used imitation-learning baseline, on three simulation benchmarks, and on a real GR-1 it reached a 76.8% average success rate against 46.4%. With only a tenth of the data, the gap widened to 42.6% against 10.2%.

The synthetic-data pieces were released alongside. The Isaac GR00T Blueprint for synthetic manipulation motion generation uses Omniverse and Cosmos Transfer world models to expand a few human demonstrations into large motion datasets. NVIDIA said it generated 780,000 trajectories, equivalent to 6,500 hours or nine continuous months of human demonstration, in 11 hours, and that mixing that data with real data lifted GR00T N1's performance by 40% over real data alone. The blueprint is offered on build.nvidia.com and GitHub. A separate open dataset on Hugging Face started at 15 terabytes with more than 320,000 robot training trajectories and up to 1,000 OpenUSD assets, and NVIDIA named UC Berkeley's DeepDrive Center, Carnegie Mellon's Safe AI Lab and UC San Diego's Contextual Robotics Institute as early users.

Newton is the longer-term bet. NVIDIA said the engine is built on its Warp framework and tuned for robot learning, and that it expected the engine to be available later in 2025. Separately, Google DeepMind and NVIDIA are developing MuJoCo-Warp, which NVIDIA expects to speed up robotics machine-learning workloads by more than 70 times. Disney Research plans to use Newton for its robotic character platforms, starting with the Star Wars-inspired BDX droids, one of which joined Huang on stage. Kyle Laughlin of Walt Disney Imagineering said the collaboration would let Disney build robotic characters that are more expressive and engaging.

Huang framed the launch as a labour question. He told the keynote audience that the world will be at least 50 million workers short by the end of the decade, and said in the release that the age of generalist robotics is here. The same keynote announced that General Motors is adopting NVIDIA AI, simulation and accelerated computing for vehicles, factories and robotics. GR00T N1 sits on top of a stack NVIDIA has built for years, from Omniverse and Isaac Lab for simulation to Jetson for on-robot compute, and it follows the first Isaac GR00T Blueprint shown at CES on 6 January 2025, which MEKAWIRE covered at the time.

Rivals moved in the same weeks. Figure AI unveiled Helix on 20 February 2025, a two-part model with a 7-billion-parameter vision-language module running at 7 to 9 Hz and an 80-million-parameter control policy at 200 Hz, trained on about 500 hours of teleoperation, with weights kept in-house. Google DeepMind announced Gemini Robotics on 12 March 2025, built on Gemini 2.0 and offered to Apptronik and a group of trusted testers. Physical Intelligence described its 3.3-billion-parameter π0 model, which controls robots at up to 50 Hz after training on about 10,000 hours of robot data, in a preprint first posted on 31 October 2024.

The open questions are real. The benchmark wins in the preprint are against Diffusion Policy and a BC-Transformer, not against π0, Helix or Gemini Robotics, and the paper has not been peer reviewed. The Hugging Face card states the model is not tested or intended for mission-critical uses that require functional safety. The weights come under NVIDIA's own licence rather than a standard open-source one, and Newton existed only as a plan with a 2025 target. NVIDIA later said, on 18 May 2025 at Computex, that it had built an update, GR00T N1.5, in 36 hours using synthetic data, against nearly three months of manual collection.