NVIDIA's Isaac ROS 5.0 Brings AI Agents to Robot Building

NVIDIA released Isaac ROS 5.0 at the ROSCon robotics conference in Toronto on Sept. 22, adding a format for AI coding agents to the software stack that roughly 1.3 million ROS developers already use to program robots.

The release is free, open source and published on GitHub. It targets NVIDIA's Jetson hardware line, from the entry-level Jetson Orin Nano up to the newer Jetson Thor module, and adds support for ROS 2 Lyrical, the current release of the Robot Operating System, alongside Ubuntu 24.04.

What ROS is, and where Isaac ROS sits on top of it

ROS is open source middleware: a set of libraries and conventions that let a robot's cameras, motors and planning software talk to each other without every team writing that plumbing from scratch. Robotics engineers have used it since 2007. Isaac ROS is NVIDIA's layer on top of that, a set of GPU-accelerated packages for perception and navigation tasks, such as object tracking, depth estimation and path planning, that run too slowly on a robot's CPU alone.

What shipped at ROSCon on Sept. 22

Isaac ROS 5.0's headline addition is a set of Agent Skills, structured in the open Agent Skills format, that let AI coding assistants follow documented robotics-development procedures instead of a developer writing setup scripts by hand. The Isaac ROS command-line tool adds isaac-ros-activate, for starting a development environment, and an early-access skill called migrate-node-to-rosidl-buffer, for moving a node from NVIDIA's older NITROS interface to a new buffer format contributed to ROS Lyrical itself.

What	Detail
Released	Sept. 22, 2026, at ROSCon in Toronto
ROS support	ROS 2 Lyrical, on Ubuntu 24.04
Hardware range	Jetson Orin Nano through Jetson Thor
FoundationPose tracking	Up to 5.5 times faster, no baseline given

The other named improvement is to FoundationPose, NVIDIA's foundation model for tracking an object's position and orientation. NVIDIA's own post says the model "enables robots to perceive and track the position and orientation of objects up to 5.5x faster" through a new agent-ready inference library, without stating what speed that multiple is measured against.

What Agent Skills adds for a developer

"There's really a lot of friction in robotics development. There's so much friction in piecing the plumbing together," Katie Washabaugh, NVIDIA's product marketing manager for robotics simulation, told The Robot Report. Washabaugh said the goal is for a developer to "release your agents and have them do all the work of getting the scripts ready, setting up the environment," rather than hand-configuring a Jetson board for each new project.

That is a real cost today. Integrating a new sensor or robot arm into a ROS project typically means writing driver code, calibration scripts and launch files by hand, work that varies by hardware and rarely transfers between projects. Whether packaging that setup as an agent skill actually saves time depends on how often a skill matches a given team's hardware, a figure neither NVIDIA nor The Robot Report's account puts a number on.

Who is actually building on Isaac ROS 5.0

NVIDIA names 13 companies integrating Isaac ROS, including Intrinsic, Magna, Universal Robots and Ouster, but gives detail on only two. Intrinsic uses FoundationPose for object registration and pose estimation, and Magna pairs Isaac ROS with NVIDIA's Isaac Sim for hardware-in-the-loop testing, according to The Robot Report. Universal Robots, which shipped its Gen 7 cobots without disclosing pricing, appears only on the adopter list, with no detail on how its arms use the release.

Manufacturing is the ecosystem's primary focus for now, though NVIDIA frames the release as open-ended across robot types. That framing puts Isaac ROS in the same race as Arm's Total Design program, which reused a chip-licensing brand for robotics software months before committing any of its own silicon to the category.

What NVIDIA didn't say

NVIDIA's announcement gives no figure for how many of the 1.3 million ROS developers have Jetson hardware to run Isaac ROS on, no pricing for anything beyond the free packages, and no baseline for the FoundationPose speed claim. The company also does not say how many of the 13 named adopters are paying customers versus unpaid open source integrations. The next test of the agent-skills pitch is whether outside developers publish their own skills to the format, something the open license permits but that, as of the Toronto announcement, nobody outside NVIDIA had done yet.