NVIDIA upgrades its humanoid model to GR00T N1.5 at Computex and says world-model video cut a three-month data job to 36 hours

NVIDIA used Computex 2025 in Taipei to argue that the scarcest input in humanoid robotics, demonstration data, can be manufactured on its GPUs. In a release dated May 18, 2025, the company introduced Isaac GR00T N1.5, an update to the GR00T N1 humanoid foundation model it opened at GTC in March 2025, together with Isaac GR00T-Dreams, a blueprint that turns video from its Cosmos world foundation models into robot training data.

The headline number came from NVIDIA Research. The company said its researchers used GR00T-Dreams to generate the synthetic training data for GR00T N1.5 in 36 hours, work it said would have taken nearly three months of manual human data collection. NVIDIA's research page dates the open release of the model to June 11, 2025, on GitHub and Hugging Face. The Hugging Face model card lists 3 billion parameters, against 2 billion for GR00T N1, and states that the model is ready for non-commercial use under an NVIDIA licence.

GR00T-Dreams works in three steps, according to the release. Developers first post-train a Cosmos Predict world model for their own robot. From a single image, the blueprint then generates videos of that robot carrying out new tasks in new environments. Finally it extracts action tokens from those videos, compact chunks of motion data that are used to teach the robot the new task. The method is laid out in a preprint, DreamGen, posted to arXiv on May 19, 2025 by NVIDIA researchers led by Joel Jang, Seonghyeon Ye and Zongyu Lin. The authors report that a humanoid performed 22 new behaviours in seen and unseen environments after starting from teleoperation data for a single pick-and-place task in one environment, and they published a video benchmark called DreamGen Bench along with the code.

The rest of the package filled in the pipeline around the model. Cosmos Reason, a world model that uses chain-of-thought reasoning to curate higher-quality synthetic data, went live on Hugging Face, and Cosmos Predict 2, the model used inside GR00T-Dreams, was listed as coming soon with what NVIDIA called reduced hallucination. NVIDIA said Isaac Sim 5.0 would soon be openly available on GitHub, that Isaac Lab 2.2 would add evaluation environments for testing GR00T N models, and that its open Physical AI dataset now held 24,000 humanoid motion trajectories used to develop the GR00T N models. The GR00T-Mimic blueprint, which augments existing manipulation data using Omniverse and Cosmos, completed the set.

The research page explains what changed inside the model. The vision-language model is now frozen during both pretraining and fine-tuning, and NVIDIA moved to an Eagle 2.5 model tuned for grounding, which scored 40.4 on a GR-1 grounding IoU test against 35.5 for Qwen2.5-VL 3B. A new objective called FLARE, short for future latent representation alignment, lets the model learn from human videos without generating future frames. NVIDIA trained N1.5 for 250,000 steps on 1,000 H100 GPUs with a global batch size of 16,384, on a data mix that included its own GR-1 recordings, Open X-Embodiment, simulated GR-1 data, DreamGen neural trajectories and AgiBot-Beta.

The reported gains are large but self-measured. On a real Fourier GR-1 humanoid asked to pick one of two fruits and place it on a plate, the language-following rate rose from 46.6% for N1 to 93.3% for N1.5, and overall success from 43.3% to 83.0%. On 12 new verbs taught through DreamGen data, success went from 13.1% to 38.3%. After post-training on 1,000 teleoperation episodes on a Unitree G1, N1.5 succeeded 98.8% of the time with four seen fruits and 84.2% with five unseen objects, against 44.0% for N1 on the same task, according to NVIDIA.

NVIDIA named four early users of the GR00T N models. Korea's AeiROBOT is using them so its ALICE4 robot can follow natural-language instructions in industrial pick-and-place work, Foxlink Group to make industrial manipulators more flexible and efficient, Lightwheel to validate synthetic data for faster humanoid deployment in factories, and Germany's NEURA Robotics is evaluating them for household automation. Agility Robotics, Boston Dynamics, Fourier, Mentee Robotics, NEURA Robotics and XPENG Robotics were described as simulating and training humanoids in Isaac Sim and Isaac Lab, Skild AI as using the frameworks to develop general robot intelligence, and General Robotics as integrating them into its platform. Foxconn and Foxlink are using GR00T-Mimic to speed up their robot training pipelines.

The Taiwan angle ran through the keynote. NVIDIA's keynote summary said TSMC, Foxconn, Wistron, Pegatron, Delta Electronics, Quanta and GIGABYTE are using NVIDIA Omniverse to build digital twins, and that NVIDIA and Foxconn Hon Hai Technology Group plan to work with Taiwan's government on an AI factory supercomputer built on Blackwell infrastructure. On hardware, Cisco, Dell Technologies, Hewlett Packard Enterprise, Lenovo and Supermicro announced RTX PRO 6000 Blackwell servers, and NVIDIA pointed developers to GB200 NVL72 systems on DGX Cloud for up to 18 times greater data processing performance. Jetson Thor, the on-robot computer, was still listed as coming soon at the time; it reached developers in August 2025.

The release positions NVIDIA as a supplier to companies that are also building their own robot brains. Figure unveiled its in-house Helix vision-language-action model in February 2025, and Skild AI, named in the release as an Isaac user, describes its own goal as general robot intelligence. Because the GR00T weights are open but restricted to non-commercial use, N1.5 works for most humanoid makers as a research baseline and reference design rather than a model they can ship in a product without separate terms.

In plain terms, GR00T N1.5 is a vision-language-action model. A frozen vision-language model reads camera images and a spoken or written instruction, and a diffusion-style action module turns that understanding into joint commands. GR00T-Dreams works on the training data instead. Rather than recording people steering a robot through each new task, a world model imagines the robot doing it on video, and a separate model infers the motor actions that would have produced that video.

Several questions remain open. All benchmark figures come from NVIDIA's own tests on a small set of tasks and robots, chiefly Fourier's GR-1 and Unitree's G1. The 36-hour figure measures data generation for one model update, not the time needed to put a robot to work in a factory. Generated video can contain physically impossible motion, which is why NVIDIA paired Cosmos Predict 2 with a separate reasoning model to filter data. And a non-commercial licence leaves partners to experiment, or to negotiate other terms, before anything reaches a customer.