NVIDIA releases Isaac Lab 3.0 and unveils Agile ONE SN reference humanoid at GTC Berlin

NVIDIA's GTC Berlin update made Isaac Lab 3.0 generally available. Isaac Lab 3.0 is open source, supports multi-GPU training, and runs on both the Newton and PhysX physics engines, with the move to Newton happening gradually. The update also added three DGX Station workflows: reinforcement-learning refinement of GR00T 1.7, dexterous manipulation training, and Cosmos post-training. Alongside these, NVIDIA announced Agile ONE SN, a 1.40 m, 40 kg humanoid with 31 degrees of freedom. It ships with WUJI hands, Jetson Thor and GR00T already installed, and is due in early 2027. [2] [1] [6]

Huang described a hybrid training setup. A conventional physics simulator handles contact and dynamics, and a world model like Cosmos re-renders the robot's camera view with different materials and lighting, which avoids the cost of rendering realistic images. He named three uses for Cosmos: generating training data, evaluating policies, and serving as the backbone of world action models for future GR00T versions. Because world models can get physics wrong, he said they still need to be grounded by physics simulators. [1]

SONIC is a 42M-parameter whole-body motion-tracking controller for humanoids. It learns from more than 100M frames (about 700 hours) of motion capture without hand-designed rewards, and accepts input from VR teleoperation, human video or vision-language-action models (VLAs). Its authors report that performance improves with more compute and more varied data. It was released openly as part of GR00T Whole-Body Control. Huang noted that it needs a lot of demonstration data and is meant to work alongside a higher-level VLA that handles long-horizon reasoning. [4] [5] [1]

Huang said the hard part of physical AI is data that captures how the world pushes back, which does not exist at web scale. Teleoperation is difficult without force feedback, so NVIDIA is turning to egocentric and third-person human video. That footage has to be retargeted to robot bodies whose joints differ from human ones. He described closing the sim-to-real gap as a loop that runs from data collection through deployment, not a single step. [1]