The Pulse
NVIDIA Ships Isaac ROS 5.0 With Agent Skills for Robot Developers
NVIDIA has released Isaac ROS 5.0, adding AI-agent workflows, ROS 2 Lyrical support and new tools for building physical AI applications. The open-source release runs across Jetson Orin Nano and Jetson Thor platforms and includes faster Foun
AI.info Team ·
NVIDIA released Isaac ROS 5.0 on September 22, adding AI-agent workflows to its open-source robotics software stack as the company presented the update at ROSCon in Toronto. The release targets developers building robots that combine perception, reasoning and movement, with support spanning ROS 2 Lyrical, Ubuntu 24.04 and NVIDIA Jetson hardware.
Isaac ROS is a collection of GPU-accelerated packages built on the Robot Operating System, or ROS. NVIDIA says the platform serves a community of nearly 1.3 million ROS users and is designed to bring accelerated perception, navigation, manipulation and inference into applications that can run on physical machines.
NVIDIA describes Isaac ROS 5.0 as free and open source. The accompanying release notes identify version 5.0.0 as a September 21 package release, while NVIDIA’s announcement appeared on September 22.
AI agents move into the ROS development workflow
Isaac ROS 5.0 introduces reusable “skills” that can guide developers and AI coding assistants through setup and manipulation tasks. NVIDIA has also added agent-ready documentation intended to give software agents a clearer understanding of Isaac ROS tools, packages and workflows.
One new FoundationStereo fine-tuning skill helps an agent adapt a stereo-perception model to a developer’s cameras, environment and application. The release also turns pick-and-place into a standalone skill that connects object detection, depth estimation and pose output without tying the workflow exclusively to one Isaac ROS application.
FoundationPose, NVIDIA’s model for object pose estimation and tracking, now includes an agent-ready inference library. NVIDIA says the updated implementation can process object pose and tracking workloads up to 5.5 times faster, although the announcement does not specify the hardware, test configuration or baseline used for that comparison.
ROS 2 Lyrical brings a new data path
Version 5.0 supports ROS 2 Lyrical and Ubuntu 24.04. NVIDIA worked with the Open Source Robotics Alliance on a standard data-handling interface for ROS Lyrical, allowing ROS software to exchange large data types such as tensors and point clouds more efficiently across CPUs and GPUs.
The release notes describe the change as a native rebuild of NITROS around ROS 2 Lyrical’s rosidl::Buffer and CUDA buffer backend. Standard ROS messages with array fields can now be backed by GPU cache memory when appropriate.
The transition also changes the software developers must maintain. Several older NITROS packages, including isaac_ros_nitros, isaac_ros_managed_nitros and isaac_ros_pynitros, have been removed. Code that calls NITROS APIs or types directly requires a source migration, while isaac_ros_nitros_bridge_ros2 remains available but is deprecated.
Jetson support runs from Orin Nano to Thor
Isaac ROS 5.0 supports NVIDIA’s Jetson Orin Nano through the higher-performance Jetson Thor platform. That range gives developers a route from smaller edge systems to machines capable of handling heavier sensor processing and on-device AI workloads.
NVIDIA’s customer examples show the software being used across humanoids, industrial arms, mobile robots and manufacturing systems. Mentee Robotics uses Isaac ROS as the perception and AI backbone of its MenteeBot humanoid, while Universal Robots has incorporated it into an AI Accelerator software development kit for perception and motion capabilities.
ROBOTIS is integrating Isaac ROS into its AI Worker robot for vision-guided manipulation, including picking, placing and alignment. FieldAI is using Isaac ROS on Jetson devices for robot foundation models that can run without cloud connectivity, according to NVIDIA’s announcement.
Open projects extend the release beyond NVIDIA
The new release also arrives with integrations from companies working around the ROS ecosystem. AgenticROS, an open-source project sponsored by RealSense, connects Isaac ROS with NVIDIA Nemotron models and NemoClaw blueprints so AI agents can interact with ROS-based robots.
Intrinsic’s Open Machine Tending Solution, part of its open-source Intrinsic Core suite, includes compatibility with NVIDIA FoundationPose for object registration, tracking and pose estimation. NVIDIA says the reference application is designed for computer numerical control machine tending, where flexible perception can reduce dependence on fixed fixtures.
Other integrations target the practical problems of deploying robot software. Magna combines Isaac ROS with Isaac Sim for hardware-in-the-loop testing, Prefix.dev’s Pixi helps assemble reproducible ROS and CUDA environments, and Foxglove tools appear throughout Isaac ROS tutorials for visualizing and debugging live ROS applications.
From generated code to a robot on the floor
Isaac ROS 5.0 positions AI agents as development assistants rather than as a replacement for the runtime software that controls a robot. The release supplies agents with documented workflows, migration tools and reusable skills, while the robot still depends on ROS nodes, perception models, planners and hardware-specific deployment.
NVIDIA is also adding a new isaac_ros_gpu_partitioning package that assigns a fixed portion of a GPU’s streaming multiprocessors to individual ROS 2 processes through CUDA Multi-Process Service. The package does not partition GPU memory or isolate workloads, according to the release notes, limiting what developers should assume about process separation.
Isaac ROS 5.0 is available now through NVIDIA’s Isaac ROS resources. The immediate change is not a new robot model but a new way to assemble robot software: agents can help configure, adapt and migrate parts of a ROS application, while Jetson hardware carries the resulting perception and control stack onto the machine.