What's this about?
A method for embodied visual tracking: a mobile robot is told in natural language whom to follow ("the person in the red jacket") and must keep following using only onboard cameras, decoupling identify-the-target from plan-the-pursuit (50 upvotes).
What exactly is "embodied visual tracking"?
A robot physically following a moving target through the world by sight — harder than tracking in video because the robot must also move itself, keep the target in frame, and recover when occluded.
Why does it matter?
Follow-me behavior from a plain sentence is a base capability for service robots, drones, and camera platforms — and a concrete test of whether vision-language-action models can hold an instruction over time.