axros2 is an open ROS2 Jazzy integration that connects the Axelera Metis AIPU to your robotics stack, so you can run real-time vision at the edge without burning your power budget (214 TOPS at roughly 5 to 9 watts).
What axros2 covers today:
Real time object detection on the Metis AIPU, published as standard vision_msgs, so it drops straight into any ROS2 graph. Human pose estimation and 3D wrist tracking, enough to let a robot follow your hand or take an object handed to it. A full MoveIt2 pick and place pipeline on a Niryo NED2 arm, covering detect, approach, grip, verify, and place. Helper nodes for camera input, live overlays, and annotated capture, plus ready to run launch files and full documentation. Two integration styles shown side by side, the low level AxRuntime C++ API for tight control, and the higher level AxInferenceNet pipeline for fast model swapping.
Why it matters for your application: whether you are building a mobile robot, a manipulator, or an inspection system, axros2 gives you a working, documented reference that plays nicely with the tools you already use, MoveIt2, Nav2, and Foxglove. You can reuse the inference node as is, swap in your own models, or adapt the perception to motion bridge for your own hardware. It runs on x86 and on Arm platforms like the Raspberry Pi 5, so the same stack scales from the workbench to the robot.
If you are working in ROS2 and want serious perception performance per watt, this is a great place to start. Explore the repo and get building: https://github.com/axelera-ai-hub/ax-ros2-ref
#Robotics #ROS2 #EdgeAI #ComputerVision #Manipulation #AxeleraAI #MoveIt2
