NVIDIA Isaac ROS 5.0 Brings AI Agents Into Open-Source Robot Development
NVIDIA has released Isaac ROS 5.0, a major update to its GPU-accelerated robotics software stack that brings AI-agent workflows deeper into open-source robot development.
Announced Sept. 22 at ROSCon 2026 in Toronto, the release adds new tools designed to help developers and AI agents build, customize and deploy robotics applications while expanding support for the latest ROS 2 platform.
The update arrives as robotics development increasingly combines traditional software engineering with physical AI, where robots must perceive their surroundings, reason about what they see and take actions in the real world.
Isaac ROS 5.0 Adds Agentic Robotics Workflows
Isaac ROS is built on the open-source ROS 2 framework and provides NVIDIA-accelerated packages for robotics tasks including perception, localization, mapping, navigation and manipulation.
With version 5.0, NVIDIA is adding reusable “skills” that can be used by developers and AI agents during robotics development.
These include workflows for setup and manipulation, along with agent-ready documentation intended to help AI systems understand Isaac ROS tools and development processes.
One notable addition is a FoundationStereo fine-tuning skill. NVIDIA says an AI agent can use the workflow to help adapt a stereo perception model to a robot’s cameras, environment and application.
The release also adds an agent-ready inference library for FoundationPose, NVIDIA’s model for object pose estimation and tracking. NVIDIA says the updated implementation can allow robots to perceive and track the position and orientation of objects up to 5.5 times faster.
A standalone pick-and-place skill is also available, separating a common robotics workflow involving object detection, depth estimation and pose information into a reusable development component.
Support for ROS 2 Lyrical and Ubuntu 24.04
Isaac ROS 5.0 adds support for ROS 2 Lyrical and Ubuntu 24.04, giving developers a path toward newer versions of the underlying robotics software platform.
NVIDIA has also contributed a data-handling interface to ROS Lyrical that is designed to help robotics applications move large data between computing hardware more efficiently.
NVIDIA’s technical documentation says the rosidl::Buffer abstraction and CUDA buffer backend can allow ROS 2 nodes to exchange GPU-resident data through zero-copy transport when the required conditions are met. The goal is to reduce unnecessary movement of large workloads between GPU and CPU memory.
That matters for AI-heavy robots because perception and inference pipelines can process substantial amounts of camera, sensor and tensor data.
Open-Source Projects Are Expanding the Ecosystem
NVIDIA says several robotics companies and open-source projects are already building around Isaac ROS.
The AgenticROS project, sponsored by RealSense, connects Isaac ROS with NVIDIA Nemotron open models and NemoClaw blueprints to allow AI agents to interact with ROS-based robots.
Other organizations are using Isaac ROS for applications ranging from machine tending and manipulation to perception, navigation, localization and industrial automation.
Intrinsic, for example, is using FoundationPose in its Open Machine Tending Solution for object registration, tracking and pose estimation. Magna is combining Isaac ROS with Isaac Sim for hardware-in-the-loop testing, while Ekumen is using GPU-accelerated Isaac ROS packages for robotics applications including precision docking and motion planning.
The broader ecosystem also includes integrations involving Foxglove, Flexiv, Ouster, Seeed Studio and Prefix.dev.
NVIDIA Pushes Robotics AI Toward the Edge
The software also connects with NVIDIA’s Jetson platform, which is designed to run AI workloads directly on robots and other edge devices.
Isaac ROS 5.0 supports NVIDIA’s scalable Jetson lineup, ranging from the Jetson Orin Nano to newer Jetson Thor systems.
That development path is important because robotics applications cannot always depend on cloud connectivity. Real-time perception, motion planning and control often need to happen locally on the machine.
Companies including Mentee Robotics, Universal Robots, ROBOTIS, FieldAI and Noble Machines are using NVIDIA’s robotics software and edge computing platforms for different physical AI applications, according to NVIDIA.
Mentee Robotics, for instance, uses Isaac ROS as part of the perception and AI stack for its MenteeBot humanoid, while Universal Robots has integrated Isaac ROS into its AI Accelerator software development kit.
What Isaac ROS 5.0 Means for Robot Developers
The biggest change in Isaac ROS 5.0 is not simply another collection of accelerated robotics libraries.
NVIDIA is positioning AI agents as participants in the robotics development process itself.
Instead of using AI only to generate isolated pieces of code, developers can use agent-ready workflows to assist with tasks such as configuring robotics software, adapting perception models and optimizing ROS applications.
That approach could make increasingly complex robotics stacks easier to develop, although the resulting systems still require testing and validation before deployment on physical machines.
NVIDIA’s latest ROS 2 work also highlights the importance of optimizing data movement alongside GPU computation. Its technical documentation demonstrates an AI-agent workflow that can inspect an existing CUDA-accelerated ROS node, plan a migration to the newer buffer system and verify the resulting data path.
Isaac ROS 5.0 Is Available Now
NVIDIA says Isaac ROS 5.0 is available now as free and open-source software.
The release gives robotics developers a combination of ROS 2 compatibility, NVIDIA GPU acceleration, physical AI models and new agent-oriented development workflows.
As AI agents move beyond software development and into physical systems, Isaac ROS 5.0 represents NVIDIA’s effort to connect those two worlds: AI-assisted creation of robot applications and the computing infrastructure needed to run those applications on real machines.
