At the Conference on Robot Learning (CoRL) in Seoul, NVIDIA unveiled a suite of groundbreaking open-source tools and infrastructure aimed at accelerating the development of physical AI and humanoid robots. The company introduced the open-source Newton physics engine, now available via NVIDIA Isaac Lab, which was co-developed with Google DeepMind and Disney Research and is managed by the Linux Foundation. Built on NVIDIA Warp and OpenUSD, Newton offers GPU-accelerated, high-fidelity simulation for complex robotic tasks such as walking on uneven terrain or manipulating delicate objects. It enables researchers and developers to train robots in realistic virtual environments before deploying them in the real world. NVIDIA also launched the latest version of its open robot foundation model, Isaac GR00T N1.6, which will soon be available on Hugging Face. This model enhances robots’ ability to understand ambiguous human instructions and execute complex, multi-step tasks—like pushing open a heavy door—by integrating NVIDIA Cosmos Reason, a vision-language reasoning model that uses common sense and physics knowledge to break down tasks into actionable plans. Cosmos Reason, already downloaded over 1 million times, is now available as a microservice via NVIDIA NIM and supports data labeling and training for physical AI models. A major update to the Cosmos world foundation model (WFM), now called Cosmos Predict 2.5, combines three models into one, reducing size by 3.5x while enabling the generation of up to 30-second video sequences with multi-camera views. This allows developers to create rich, high-quality synthetic data for training, significantly speeding up model development. To improve robot skill acquisition, NVIDIA introduced a new open-source workflow in Isaac Lab 2.3 for training dexterous manipulation skills, such as grasping objects. The system uses automated curriculum learning, gradually increasing task difficulty and adjusting physics parameters like gravity and friction. Boston Dynamics’ Atlas robot has already used this workflow to improve its grasping capabilities. NVIDIA also unveiled Isaac Lab Arena, a new open framework for large-scale, standardized evaluation of robot skills in simulation, helping developers assess performance across diverse, complex scenarios without building test environments from scratch. To support these tools, NVIDIA launched new AI infrastructure, including the GB200 NVL72 rack system with 36 Grace CPUs and 72 Blackwell GPUs, RTX PRO servers for unified robot development workloads, and Jetson Thor—powered by Blackwell GPUs—to enable real-time AI inference on robots. Companies like Meta, Google DeepMind, Figure AI, and Yuque Robotics are already adopting these platforms. The impact of NVIDIA’s ecosystem is evident: nearly half of all CoRL papers reference NVIDIA technologies. Institutions like Stanford, ETH Zurich, and Peking University are leveraging NVIDIA’s tools for research, while companies including Agility Robotics, Universal Robots, and Volkswagen use Isaac Sim and Omniverse to build digital twins and streamline robot deployment. With open frameworks, powerful models, and scalable infrastructure, NVIDIA is driving a “sim-first” revolution in robotics, empowering developers to build smarter, safer, and more adaptable robots for real-world applications.