Robots are about to get a second kind of developer: AI agents that write the code. Toronto hosted the launch, timed to the ROSCon conference. This fifth major release bundles NVIDIA’s GPU-accelerated code for ROS into one package set, and it is built for agentic workflows as much as human ones.
Support now reaches ROS Lyrical and Ubuntu 24.04. Working with the Open Source Robotics Alliance, NVIDIA contributed a standard data-handling interface to Lyrical. The aim is portability, so code written once stops caring which accelerator sits underneath; CUDA appears as the first working example. Roughly 1.3 million people use ROS.
Several new Isaac skills target setup and manipulation, and the documentation was written to be legible to agents rather than only humans. A fine-tuning skill for FoundationStereo adapts stereo perception to a specific camera and environment. FoundationPose, which estimates object position and orientation, gains an agent-ready inference library that runs up to 5.5 times faster. Pick and place is now a standalone skill usable outside Isaac ROS.
Outside projects are already extending the approach. The project is called AgenticROS and it comes from RealSense, a company that builds 3D perception hardware. Pairing Isaac ROS with NVIDIA’s Nemotron models and NemoClaw blueprints lets an agent operate a ROS-based robot on its own.
