TF is the system robots use to keep track of where every part and sensor is relative to everything else — a live tree of coordinate frames that answers "where is this point in that frame?" at any moment.
TF keeps track of where everything on a robot is in relation to everything else — the camera to the wheels, the wheels to the room — so the robot can convert a point seen by one sensor into any other frame of reference.
A robot sees a coffee cup with its camera. But to pick it up, the arm needs that cup's position in the arm's own frame — not the camera's. Converting between those viewpoints, hundreds of times a second, is the job of TF.
The problem it solves
A robot is a bundle of coordinate frames: the map, the robot's base, each wheel, the camera, the gripper. A measurement always arrives in one frame (the camera's), but you usually need it in another (the base, or the world). TF (tf2 in ROS 2) maintains the relationships between all of them and does the conversions for you.
How it works
TF stores a tree of frames connected by transforms — each transform is a translation plus a rotation (stored as a quaternion). To express a point from frame A in frame B, TF walks the path between them and chains the transforms.
A typical mobile-robot frame tree
Each arrow is a live transform. Ask TF for camera→map and it multiplies the chain, so a point the camera sees becomes a point on the map.
The genius — and the tricky part — is that TF is time-stamped. The camera-to-base transform 200 ms ago may differ from now (the robot moved). TF buffers a short history so it can answer "where was the camera relative to the map at the instant this image was taken?" Getting this timing wrong is the classic TF bug: a valid transform at the wrong time puts objects in the wrong place.
Where you'll see it
Every ROS 2 robot that navigates or manipulates runs a TF tree. Standard conventions (REP 105) fix the important frames: map (global, drift-free but jumpy after loop closures), odom (smooth but drifts, from odometry), and base_link (the robot itself). Sensors hang off base_link via fixed transforms defined in the URDF.
Why it matters
TF is the quiet backbone of robot spatial reasoning. Without it, every node would re-implement the same fragile geometry. With it, perception, navigation, and manipulation all share one consistent, time-aware picture of where everything is.
