OpenEyes gives humanoid robots vision

Strong idea, and the scope is bigger than a demo: this is edging toward a practical edge-AI perception stack for hobbyist and research humanoids.

• –The combination of YOLO11n, MiDaS, and MediaPipe makes it immediately useful for common robotics perception tasks without forcing a custom model pipeline.
• –ROS2 support matters because it turns the project from a standalone vision script into something that can plug into navigation, control, and manipulation stacks.
• –The reported 10-15 FPS full-stack and 30-40 FPS INT8 modes suggest the author is optimizing for real deployment on constrained Jetson hardware, not just laptop-grade demos.
• –The person-following and gesture-based owner selection features are the kind of interaction layer humanoid builders actually need, even if the accuracy and edge cases will decide whether it survives outside the lab.
• –Open-source plus edge-first is the right angle here: robotics teams want fewer cloud dependencies, lower latency, and a system they can inspect and extend.