NVIDIA Isaac Lab 1.2 Generally Available on GitHub, New Robot Learning Tools Out

NVIDIA has unveiled a range of advanced tools and workflows to expedite the development of robotics, including humanoids, at the Conference for Robot Learning (CoRL) in Munich, Germany. 

These new offerings, designed to enhance AI simulation and robotics, include the NVIDIA Isaac Lab’s framework for robot learning, six specialised workflows for humanoid robot development under Project GR00T, and new video data processing tools, namely the NVIDIA Cosmos Tokenizer and NeMo Curator.

This release marks the general availability of NVIDIA Isaac Lab, an open-source Omniverse-based framework tailored to train robots at scale across diverse forms, including humanoids and quadruped robots. 

Prominent robotics companies such as Boston Dynamics, Agility Robotics, and Swiss-Mile, have already adopted Isaac Lab for commercial and research purposes. This framework supports increasingly complex robotics tasks, enabling robots to perform intricate movements and interact effectively within environments.

GR00T for Humanoids

Project GR00T, NVIDIA’s initiative for accelerating humanoid robot development, introduces six new workflows designed to address the core challenges of humanoid robotics. These include GR00T-Gen for creating AI-driven 3D environments, GR00T-Mimic for trajectory generation, GR00T-Dexterity for dexterous manipulation, GR00T-Control for body control, GR00T-Mobility for navigation, and GR00T-Perception for sensory processing. 

With NeMo Curator, an advanced video processing pipeline, NVIDIA plans to accelerate video data curation by up to 7x compared to conventional methods. Designed to manage extensive datasets, NeMo Curator incorporates automatic orchestration across multi-node, multi-GPU systems, handling over 100 petabytes of data.

During CoRL, NVIDIA further solidified its leadership in robotics by presenting 23 research papers and conducting nine workshops focused on robot learning advancements. 

Among several other developments in robotics, NVIDIA recently rolled out HOVER, a 1.5-million-parameter neural network and DexMimicGen, a large-scale synthetic data generator. Earlier this year, NVIDIA, in partnership with Hugging Face, also launched LeRobot, an initiative to advance open-source robotics research using NVIDIA’s Isaac Lab and Jetson platforms. 

Meta, also catching up in the open-source robotics race, recently released new research artefacts that allow robots to perceive touch. The race to open-sourcing robotic features is steadily expediting.