NVIDIA has released Isaac ROS 5.0, the latest version of its free, open-source robotics software, at the ROSCon developer conference in Toronto on 22 September. The headline feature is support for AI agents, alongside faster perception models and compatibility with the newest long-term release of ROS, the Robot Operating System that most robot builders use as their software foundation.
NVIDIA says Isaac ROS now reaches a community of nearly 1.3 million ROS users. The word “agentic” is doing a lot of work in the announcement, though, so it is worth being precise about what it covers.
The Agents Work For Developers, Not Robots
The agents in Isaac ROS 5.0 are AI coding assistants, the kind developers already use to write software, rather than AI running inside a robot. NVIDIA has packaged common robotics chores as reusable “skills” written in an open format that both people and coding agents can follow, and rewritten its documentation so an agent can read it and act on it.
Those skills cover jobs that usually eat days of an engineer’s time: setting up the software environment, preparing pick-and-place workflows, migrating older code to new interfaces and fine-tuning perception models to a particular camera. “You can release your agents and have them do all the work of getting the scripts ready, setting up the environment,” NVIDIA product marketing manager Katie Washabaugh told The Robot Report.
That distinction matters for anyone worried about AI taking control of machines, a question we explored in where device safety should begin. Here the agent speeds up the build; the robot still runs conventional, tested software once it ships.
Faster Eyes For Robot Arms
The release also improves how robots see. FoundationPose, NVIDIA’s AI model for working out an object’s exact position and orientation so an arm can grab it, now runs up to 5.5 times faster through a new inference library, according to NVIDIA.
A companion skill lets an agent fine-tune FoundationStereo, a model that estimates depth from a pair of cameras, to a specific camera and setting, which matters on factory floors where lighting and surfaces rarely match the lab. Motion planning, the calculation of a collision-free path for an arm, is handled by NVIDIA’s cuMotion in milliseconds, and all of it runs on Jetson boards from the small Orin Nano up to Jetson Thor, whose MLPerf results we covered earlier.
The Quieter Change Inside ROS
The most lasting part of the release may be the least flashy. NVIDIA contributed a way for ROS itself to pass data between the main processor and graphics or other accelerator chips without copying and converting it at every step, and that code now ships in ROS 2 Lyrical Luth, the long-term release supported until 2031.
Crucially, the design is vendor-neutral. AMD, Intel and other chipmakers can plug in their own versions rather than being locked out, and the ROS project’s leader, Michael Carroll, praised NVIDIA for building it with the community rather than simply handing it over.
Our read: this is a smart long-term play. By helping set the standard for how robot software moves data onto accelerators, NVIDIA makes GPU-heavy robotics the default path in ROS, while the open design lets it avoid the charge of locking the ecosystem in.
Who Is Using It, And What It Costs To Move
Early adopters lean toward manufacturing. Intrinsic, the Alphabet robotics company, has built FoundationPose into a machine-tending system; Universal Robots uses Isaac ROS in its AI Accelerator kit; and Magna, Flexiv, ROBOTIS, Seeed Studio and humanoid developer Mentee Robotics are also on NVIDIA’s list. “What we’re seeing a lot with the ecosystem is definitely a manufacturing first focus,” Washabaugh said.
Moving to 5.0 is not free of effort, though. Developer notes on the release say that projects written directly against NVIDIA’s older GPU data interface will need source code changes to adopt the new ROS standard, with a temporary bridge available in the meantime, and some camera support is limited to container-based setups for now.
The open question is whether agent skills genuinely shorten the path from prototype to production line, or mostly speed up the demo. With a manufacturing-first user base, the answer should show up quickly in how many of these arms reach real factories over the next year.
