Intrinsic Core

Intrinsic's open industrial robot stack: real-time control, planning and perception

License

Apache 2.0

Code: Apache-2.0 (intrinsic-core, intrinsic-omts, intrinsic-moveit). Perception models are downloaded separately under their own terms: NVIDIA FoundationPose weights under the NVIDIA Open Model License, SAM 3 under Meta's SAM licence. 'Intrinsic' and 'Intrinsic Core' are trademarks of Intrinsic Innovation LLC. Contributions require Google's Contributor License Agreement. Checked 2026-10-23.

Overview

Intrinsic Core is the open-source part of the robotics platform built by Intrinsic, Alphabet's industrial robotics company, released under Apache-2.0 at ROSCon 2026. It is a complete local stack for programming industrial robot arms: a runtime that runs on a single Ubuntu PC, a real-time control framework called ICON with drivers for Universal Robots, KUKA and ABB arms, collision-free motion planning, 6-DoF part pose estimation with NVIDIA FoundationPose, grasp planning, camera calibration, and a digital twin simulated in Gazebo. Applications are written in code as behavior trees of reusable skills, built with Bazel and deployed with a command-line program, inctl. It works alongside ROS 2 through a ROS bridge, so RViz and other ROS tools can watch and drive it. A second open repository, the Open Machine Tending Solution, is a complete reference application, a UR5e loading and unloading a CNC machine, to start from.

Intrinsic Core is organised as a set of modules in one Bazel monorepo. The runtime packages everything as containers on k3s, a small Kubernetes, so a single command brings up the services an application needs on the local PC: an object world that tracks where everything is, the skill and asset registries, the simulator bridge and the real-time controller. Nothing requires an Intrinsic account or a cloud connection; the hosted Intrinsic Flowstate product is a separate commercial edition, and the documentation says plainly which pages apply only to it.

ICON, the Intrinsic real-time control framework, is the part most robotics stacks do not have. It runs a deterministic control loop that can switch controllers within one cycle based on live sensor readings, which is what force-guided moves such as move_to_contact need. A hardware abstraction layer separates the controller from the arm: the open hardware modules cover the UR3e, UR5e and UR10e over RTDE, five KUKA arms from the KR 6 to the KR 50 over the Robot Sensor Interface, and the ABB IRB 1300, plus a Gazebo stub for simulation. Motion planning handles Cartesian and joint-space moves in one API and blends segments with different speed and acceleration limits into one trajectory; a separate open repository plugs MoveIt 2 in for planning and grasping.

Perception standardises camera and point-cloud interfaces and runs NVIDIA FoundationPose for 6-DoF pose estimation of known parts on a local GPU, together with SAM 3 for segmentation and FoundationStereo for depth; the model weights are downloaded separately. Applications are composed in Python with the Solution Building Library as behavior trees of skills such as move_robot, move_to_contact, estimate_pose and gripper or digital I/O commands, and the same solution runs against the Gazebo digital twin (--operation_mode=sim) or the real cell.

The Open Machine Tending Solution shows the whole loop on a realistic job: a UR5e with a Robotiq Hand-E gripper takes parts from a pallet, either by vision-guided picking with FoundationPose or by fixed grid positions, opens the CNC door through digital outputs, loads the vise and unloads finished parts. Tutorials cover visualising the cell in RViz and Gazebo, jogging the robot, swapping arms and grippers, importing a new part and writing new skills.

Why it matters

Industrial robot programming has mostly lived in two places: each arm maker's own controller language, or ROS, which leaves real-time control, part localisation and production deployment for every team to assemble itself. Intrinsic Core is a production stack from a company that deploys it in factories, released whole under Apache-2.0: a real-time controller that can react to force within a cycle, drivers for common UR, KUKA and ABB arms, pose estimation that removes the need for fixtures, and a simulator-backed workflow from prototype to cell. It gives integrators, labs and students a working industrial baseline to read, change and run on their own robots without a licence fee or a cloud account, and it connects to ROS rather than replacing it.

Problem it solves

Building an industrial robot cell normally means writing glue between a vendor controller, a motion planner, a vision system and a simulator, each with its own interfaces, and then rebuilding much of it when the arm or gripper changes. Intrinsic Core supplies those pieces already integrated behind one hardware abstraction layer and one skill model, so an application written once runs in simulation and on different arms.

Key takeaways

• Apache-2.0 release of the core of Intrinsic's industrial robotics platform, announced at ROSCon 2026
• Runs entirely on a local Ubuntu PC (k3s containers); no Intrinsic account or cloud connection is needed
• ICON real-time control framework switches controllers within one cycle on sensor feedback, for force-guided moves
• Open drivers for UR3e/UR5e/UR10e, five KUKA KR arms and the ABB IRB 1300, plus a Gazebo digital twin
• Built-in 6-DoF part pose estimation with NVIDIA FoundationPose, grasp planning and camera calibration
• Applications are behavior trees of skills in Python; a full CNC machine-tending reference solution is open too
• Works with ROS 2 through a bridge (RViz, ROS tools) and offers a MoveIt 2 integration

Who it is for

• System integrators building machine tending, assembly or bin-picking cells on UR, KUKA or ABB arms
• Robotics engineers who need deterministic, force-reactive control that ROS alone does not provide
• Research labs and universities that want an industrial-grade baseline to run and modify on their own arms
• ROS developers who want ready-made pose estimation and planning for industrial manipulation

Limitations

• Heavy setup: Ubuntu 24.04 or 26.04 (the getting-started guide asks for 26.04), an x86 PC with 32 GB RAM or more, and a first build of the reference solution that takes about 50 minutes
• Real-time control of hardware needs an Intel CPU; ARM is not supported, and perception needs an NVIDIA RTX-class GPU
• Hardware support is a short list so far: three UR e-Series arms, five KUKA KR arms, the ABB IRB 1300, and the Robotiq Hand-E in the reference solution; other arms need a new hardware module
• No graphical editor: solutions are written in code and built with Bazel; the Flowstate web UI belongs to Intrinsic's commercial edition
• Pose-estimation weights come from NVIDIA under the NVIDIA Open Model License and SAM 3 under Meta's licence, not Apache-2.0
• Released in September 2026 with dated releases every few weeks, so interfaces may still change