Team Green has made yet another big step in the world of robotics, as they have announced the rollout of its open-source Newton Physics Engine inside its Isaac Lab platform, alongside the new open Isaac GR00T N1.6 model and fresh AI infrastructure designed for physical AI.
The two names are essentially part of a full toolkit that researchers and developers can use to train, test and transfer robot skills more safely and efficiently from simulation into the real world. For Newton, it has the backing of Google DeepMind, Disney Research and NVIDIA itself, while management is under the Linux Foundation, currently in beta. It’s GPU-accelerated, open-source, and built on NVIDIA Warp and OpenUSD, letting developers simulate things as tricky as walking on gravel or carrying fragile objects.
Early adopters already include ETH Zurich, Technical University of Munich and Peking University, plus robotics firms like Lightwheel.
On the reasoning side, the Isaac GR00T N1.6 model gets a major upgrade thanks to Cosmos Reason, NVIDIA’s customizable reasoning VLM designed for physical AI. It’s the part that turns vague human instructions into step-by-step actions while using prior knowledge and physics to generalize across new tasks. Cosmos Reason has already been downloaded over a million times and tops Hugging Face’s Physical Reasoning Leaderboard, so it’s clearly gaining traction. Now integrated into Isaac GR00T, it even enables humanoids to move and manipulate objects simultaneously, with research partners like Franka Robotics, LG, and Neura already exploring its potential.
NVIDIA is also updating its Cosmos World Foundation Models of Predict 2.5 and Transfer 2.5 to help generate diverse training data using text, image, and video prompts, with Predict now supports up to 30-second videos with multi-camera perspectives, while Transfer is 3.5x smaller yet faster and more photorealistic.
All the while, Team Green is also adding a new workflow in Isaac Lab 2.3 built with Omniverse, now capable of training multi-fingered robots with tasks that ramp up in difficulty by tweaking physics like gravity or friction. Boston Dynamics even used it to sharpen Atlas’ manipulation skills, and companies like Agility Robotics and Figure AI are adopting the same tech.
Evaluation tasks are given to Lightwheel working closely with NVIDIA through Issac Lab – Arena, an open-source evaluation framework that lets developers run scalable, standardized tests without reinventing the wheel.
