NVIDIA Adds AI Agent Workflows to Its Open-Source Robotics Toolkit

NVIDIA has released Isaac ROS 5.0, adding reusable AI-agent workflows to its free, open-source robotics software collection. The update gives coding assistants a more direct role in configuring and adapting robot applications, while moving the toolkit onto the latest supported ROS 2 platform.

Released at ROSCon in Toronto, the package collection combines NVIDIA’s GPU-accelerated software with ROS, the open framework used to build and connect robot applications. NVIDIA says Isaac ROS reaches nearly 1.3 million ROS users; version 5.0 is available now at no cost as open-source software.

From documentation to agent actions

The new release includes agent-ready documentation and skills: reusable procedures intended for developers and AI agents completing robotics-development tasks. One skill helps fine-tune FoundationStereo, a stereo-perception model, for a developer’s cameras and operating environment. Another makes a pick-and-place workflow available as a standalone, agent-ready skill.

NVIDIA also added an agent-ready inference library for FoundationPose, its object position-and-orientation model. The company says the library can let robots perceive and track objects up to 5.5 times faster; that is NVIDIA’s performance claim, not an independent measurement.

A new foundation, with migration work

The release migrates Isaac ROS to ROS 2 Lyrical and adds packages for Ubuntu 24.04 through an Isaac ROS Buildfarm APT repository. NVIDIA and the Open Source Robotics Alliance also contributed a standard data-handling interface intended to help robotics software work across computing hardware, with CUDA as a GPU-acceleration example.

That migration is not entirely frictionless. Isaac ROS rebuilt its NITROS data path around ROS 2 Lyrical’s rosidl::Buffer and a CUDA buffer backend, and removed several NITROS packages. Developers whose code directly calls NITROS interfaces need a source-level migration; the retained bridge is deprecated and slated for removal in a future release.

Deployment is broader than the agent layer

Isaac ROS 5.0 supports deployment from entry-level Jetson Orin Nano hardware through Jetson Thor devices. NVIDIA positions that range as a path from development to running perception, navigation, AI models and application logic on a robot at the edge, rather than relying on a single class of machine.

The release notes also put boundaries around that promise. RealSense cameras are supported only in Docker mode, some Jetson Thor H.264 encoding jobs can crash during shutdown, and certain upstream robot-vendor packages have not yet been certified for ROS 2 Lyrical. Agent-assisted setup may reduce repetitive work, but it does not remove compatibility testing from robotics deployment.