The robot guide dog was a three-person MSc project. The idea was a quadruped that could guide someone and tell them what was ahead. We built it on a Unitree Go1.
"We built a robot guide dog" is true and not very useful. Here is which part was mine.
My part: seeing and saying
I was responsible for vision and perception. A camera on the robot fed a YOLO detector running in ROS. The detector published what it saw, and my code decided which of those detections mattered to a person walking and turned them into spoken cues.
The decision matters more than the detection. A detector reports everything, in every frame. Someone being guided does not want "chair" announced thirty times a second, or objects that are not in their way announced at all. So detections are filtered for relevance, and an announcement is held back until the same thing has been seen consistently and has not just been announced.
My teammates' parts
Locomotion and low-level control of the robot were handled by my teammates, as was the navigation setup. I integrated with their work. I did not write it.
Where it ran
I tested the path from camera to speech in Gazebo first and then on the physical Go1. Moving to the real robot is where things like the camera's real field of view and the delay between seeing and speaking start to matter.
Why I describe it this way
Designed, built, integrated and contributed are different claims. On shared work, the precise version is easier to trust and much easier to talk about in an interview, because every sentence is something I can explain in detail.