NVIDIA drops massive toolkit to speed up humanoid robots

NVIDIA just made Isaac Lab 1.2 fully open source alongside a suite of new physical AI tools aimed squarely at speeding up humanoid robotics and world model training. Built on NVIDIA Omniverse, Isaac Lab gives roboticists an open-source, OpenUSD-based framework to simulate and train robot policies at massive scale before deploying them onto actual hardware.

Alongside the simulator update, NVIDIA dropped six new humanoid workflows for Project GR00T and open-sourced Cosmos, an ultra-fast visual tokenizer tailored for video world models.

What it does

Isaac Lab is designed to train any robot embodiment, whether you are building humanoids, quadrupeds, or industrial cobots. Instead of risking costly hardware in the real world, developers spin up thousands of parallel simulation runs to teach robots locomotion, navigation, and manipulation. Teams like Boston Dynamics, Agility Robotics, 1X, Unitree, and XPENG Robotics are already building on it.

To tackle humanoid complexities, Project GR00T now includes six dedicated workflows. GR00T-Gen builds generative 3D environments, GR00T-Mimic handles trajectory generation, GR00T-Dexterity focuses on hand manipulation, GR00T-Control handles whole-body balance, GR00T-Mobility manages navigation, and GR00T-Perception covers multimodal sensing.

Under the hood: 12x faster tokenization

Training world models requires feeding neural nets thousands of hours of video so robots learn how physical objects respond to contact. That pipeline usually chokes on compute.

The new open-source Cosmos tokenizer compresses visual frames into compact tokens up to 12x faster than current visual tokenizers while preserving fine temporal and spatial details. Humanoid startup 1X has already moved its World Model Challenge dataset onto Cosmos to generate longer-horizon predictions more efficiently.

For large-scale video curation, NVIDIA updated NeMo Curator with a video pipeline that handles over 100 petabytes of footage across multi-GPU setups, running up to 7x faster than unoptimized setups.

Where to get it

Isaac Lab 1.2 is live on GitHub, complete with migration docs for teams moving over from Isaac Gym. Cosmos tokenizers are up on both GitHub and Hugging Face. The video-enabled NeMo Curator arrives later this month, while the GR00T workflows will roll out via NVIDIA developer channels. Hugging Face is also tying Isaac Lab together with its LeRobot community and Jetson hardware.

You will need an RTX-capable GPU environment to leverage the Omniverse simulation engine, and multi-GPU nodes if you plan to chew through heavy video curation.

Why it matters

If you build robot learning policies or generative world models, this release removes months of low-level plumbing. Instead of stitching together custom video tokenizers and physics engines, you get an optimized, battle-tested pipeline for free.

It is also a textbook NVIDIA playbook: open-source the highest-friction software layers so developers flock to the stack, ensuring that every serious humanoid and embodied AI project remains anchored to NVIDIA compute.