transform tree (tf)
A transform tree is the piece of robot software that keeps track of where every named part of a robot is, relative to every other part, and how those relationships change moment to moment. A robot is full of named reference frames — one at the base, one at each joint, one on the camera, one on the gripper, one fixed to the world — and what you usually need to know is the pose (the position and orientation) of one frame as seen from another: where is the gripper relative to the camera, right now? The transform tree stores all of these relationships and answers exactly that kind of question on demand. Its best-known form, in the Robot Operating System, is literally called tf.
It is called a tree because the frames hang together like a family tree: the world frame is the root, the robot's base hangs off the world, the first arm link hangs off the base, the next link off that one, and so on out to the fingertips and the camera. Each parent-to-child link stores one rigid-body transform — the rotation and shift that takes you from the parent frame into the child frame. To find how any two frames relate, even ones far apart on different branches, the software simply walks the path of links connecting them and multiplies the transforms along the way. Arranging frames as a tree, with exactly one parent each, guarantees there is always one clear, unambiguous chain between any two of them.
Crucially, a transform tree is not a single frozen snapshot — it is time-stamped and constantly updated, because a robot's joints and wheels are always moving. Different parts report their transforms at different instants, so the system keeps a short rolling history and can interpolate to answer where was the gripper relative to the map at the exact moment that photo was taken. This is what lets perception, planning, and control all agree on one shared picture of space: a depth camera can spot an object, the tree can convert that sighting into the arm's own coordinates, and the arm can reach for it — all because every part trusts the same continuously maintained map of who is where.
A mobile robot's camera sees a doorway two meters ahead. Because the transform tree knows the chain from camera to wheels to floor, the navigation code can instantly restate that doorway in floor coordinates and drive straight through it.
The tree converts one part's view into any other part's coordinates, so the whole robot shares one map of space.
tf is the name of the well-known transform-tree library in the Robot Operating System; the underlying idea — a tree of time-stamped frame relationships — applies to robot software in general.