At COMPUTEX 2025, NVIDIA showcased major breakthroughs in robotics and AI development. Key updates to its Isaac robotics platform, including the enhanced GR00T N1.5 foundation model, the new GR00T-Dreams synthetic data generation pipeline, next-generation RTX PRO 6000 Blackwell systems, and the powerful NVLink Fusion platform are some of the highlights of the keynote. These technologies form the foundation for building more capable humanoid robots and scalable AI infrastructure.
GR00T N1.5 and GR00T-Dreams: Smarter Robots in Less Time
The GR00T N1.5 model is NVIDIA’s latest open-source foundation model for general-purpose humanoid robots. It improves robotic capabilities in understanding human instructions, adapting to new environments, and performing everyday tasks such as sorting objects and interacting with tools. Training time for this new model has been reduced significantly, from nearly three months down to just 36 hours, thanks to synthetic data generated by GR00T-Dreams.
GR00T-Dreams complements NVIDIA’s earlier GR00T-Mimic by generating large volumes of synthetic motion data. Developers can provide a single image, and the system produces simulation videos of a robot executing various tasks in virtual environments. From these videos, the pipeline extracts compressed action tokens, which serve as effective training inputs for models like GR00T N1.5.
The entire pipeline is integrated with NVIDIA Omniverse, Isaac Sim, and Cosmos, giving developers a seamless way to create, refine, and scale robotic behaviors using synthetic data.
RTX PRO 6000 Blackwell and NVLink Fusion: AI Compute Gets a Boost
To power these advancements, NVIDIA introduced workstations and servers built on the new RTX PRO 6000 Blackwell GPU. These systems are purpose-built to accelerate the training of robotics foundation models, high-fidelity simulations, and deployment across both edge and cloud environments. Developers can use them to simulate, validate, and fine-tune robot behaviors without the need for massive infrastructure.
In addition, NVIDIA revealed NVLink Fusion, a new platform that allows partners to build semi-custom AI infrastructure by integrating third-party CPUs and accelerators with NVIDIA GPUs. Partners like MediaTek, Marvell, Alchip Technologies, Astera Labs, Synopsys, and Cadence are already working on compatible silicon that fits into NVIDIA’s broader AI ecosystem.
Companies such as Fujitsu and Qualcomm plan to use NVLink Fusion to pair custom CPUs with NVIDIA GPUs, paving the way for high-performance AI factories. The platform supports up to 800Gb/s of throughput and incorporates components like NVIDIA ConnectX-8 SuperNICs, Spectrum-X Ethernet, and Quantum-X800 InfiniBand switches. This provides cloud providers with a modular, scalable solution to expand their AI capabilities using custom-built infrastructure alongside NVIDIA’s rack-scale systems.
Real-World Adoption Already Underway
A growing number of robotics companies are already putting these tools to work. Early adopters include Agility Robotics, Boston Dynamics, Sanctuary AI, 1X Technologies, Fourier Intelligence, and XPENG Robotics. These companies are using NVIDIA’s platforms to build more responsive, adaptable, and efficient humanoid robots.
With GR00T N1.5, GR00T-Dreams, Blackwell-powered compute systems, and NVLink Fusion, NVIDIA is offering a full-stack approach to robotics and AI development. These tools dramatically reduce training times, simplify data generation, and offer new levels of hardware flexibility for developers and enterprises alike.
