A robot repeats a point to a few hundredths of a millimetre, but only in its own world. Floor calibration is the work of making that world agree with the real one — foundation, mounting, mastering and BASE frames — so that programs survive a move and a second robot can share the cell.
A KUKA robot is sold on repeatability, and rightly so: an arm that returns to a taught point within a few hundredths of a millimetre, hundreds of thousands of times, is an extraordinary machine. But repeatability describes the robot talking to itself. It says nothing about whether the point the robot believes is 500 mm from its base is actually 500 mm from the fixture on the floor. The gap between the robot’s coordinate world and the physical cell is where most commissioning time disappears, and it is what we mean when we talk about floor calibration: a chain of steps that runs from the concrete to the program, each one resting on the one below it.
Step one: the floor itself
Calibration starts before the robot is switched on, because no software can correct a base that moves. A large industrial robot braking hard at an emergency stop puts forces into its foundation far higher than those of normal motion, and the manufacturer’s specification gives those loads for exactly that reason. The assembly instructions define how the robot may be fixed: to a concrete foundation with the specified anchors, or to a steel machine frame designed for the purpose, with the concrete quality, the minimum slab thickness, the bolt type and the tightening torque all stated. A floor with a crack under the base plate, an expansion joint running across it or a thin screed laid over the structural slab will pass every static test and then allow the base to rock by fractions of a millimetre under load. Magnified by a two-metre reach, a fraction of a millimetre at the base becomes a visible error at the tool.
Levelness matters for a second reason. A robot configured for floor mounting assumes that gravity acts straight down through its base, and its dynamic model — the calculation that decides the motor torques along a path — is built on that assumption. A base plate that is noticeably out of level tilts the robot’s whole world. For work that only cares about positions relative to a fixture this may go unnoticed; for dispensing, stacking or anything else that relies on gravity it does not. We level and shim the base plate to the tolerance in the instructions, torque the anchors, record the values, and check them again at the intervals the maintenance schedule gives. Ceiling- and wall-mounted robots follow the same logic with their own configuration: the mounting orientation is part of the robot’s data, not something the controller discovers by itself.
Step two: mastering, the robot’s own zero
The next layer is mastering, and it is often confused with calibration. Mastering tells the controller where the mechanical zero of each axis is, so that an angle reading on the motor corresponds to a real angle of the arm. On KUKA robots it is done with an electronic mastering device screwed onto a gauge cartridge on each axis; as the axis moves slowly past a notch, the device detects the exact position and the controller stores it. Without valid mastering the robot cannot move in Cartesian coordinates at all, and a robot with poor mastering will reach every taught point perfectly and every calculated point slightly wrong — the kind of error that makes offline programs and offsets unreliable while hand-taught programs appear to be fine.
Mastering is done at commissioning and has to be repeated whenever the relationship between motor and arm may have changed: after replacing a motor or a gearbox, after a collision at speed, or after the data has been lost. KUKA’s procedure also allows the robot to be mastered with and without the tool load and the offset between the two to be learned, so that the effect of the payload on the arm is taken into account later. It is worth doing properly, because every layer above it inherits its errors.
Step three: WORLD and ROBROOT, where the robot stands
A KUKA controller describes the cell with a small family of coordinate systems. WORLD is the fixed reference for the cell. ROBROOT is the robot’s base, defined as an offset from WORLD. BASE is a coordinate system on a workpiece or fixture, and TOOL is the working point of the end effector. For a single floor-mounted robot, WORLD and ROBROOT are usually identical and nobody needs to think about them. The moment the cell contains a second robot, a linear track or a rotary positioner, they become essential: each robot’s ROBROOT is measured relative to a common WORLD, so that both machines agree on where a shared fixture is, and a track or positioner is set up as a kinematic system of its own, so that a BASE mounted on it moves with it.
Step four: TOOL and BASE, where the work happens
TOOL calibration tells the robot where its working point is. The usual KUKA method for position is the XYZ 4-point method: the tool tip is brought to the same fixed reference point from four clearly different orientations, and the controller calculates where the tip must be relative to the flange for all four approaches to coincide. The orientation of the tool is then set with the ABC World or ABC 2-point method. The quality of the result depends on the operator more than on the software — a reference point that is sharp and rigid, orientations that are genuinely different, and a tip that touches rather than pushes. A welding torch or a gripper that has been calibrated sloppily turns every reorientation into a small position error, and the robot will be blamed for an inaccuracy it does not have.
BASE calibration is the step that ties the program to the floor. With the 3-point method, the calibrated tool touches three points on the fixture: the origin of the new coordinate system, a point on its positive X axis, and a point in its XY plane on the positive Y side. From those three points the controller defines a frame lying on the fixture. When the origin cannot be reached directly — because it sits inside a part or beneath a clamp — the indirect method uses four points whose coordinates in the base are already known. Where the position of the fixture is known precisely, the values can simply be entered numerically.
The reason this matters so much is simple. Points taught in a BASE belong to the fixture, not to the room. If a fixture is replaced, shifted by a few millimetres during maintenance or moved to a second identical cell, re-measuring that one BASE brings every point in the program back into place. Points taught in WORLD coordinates have to be touched up one by one. As a rule we teach and program against BASE frames, name them after the physical fixture they belong to, and keep a drawing of where each origin sits. The tool and base data are stored in the robot’s configuration files, so a controller archive taken after commissioning preserves the whole calibration — and that archive is taken before the robot leaves our hands.
Load data, temperature and verification
Two further details separate a calibration that holds from one that drifts. The first is load data. The controller needs the mass, centre of gravity and inertia of the tool and payload to plan motor torques; with wrong values the path deviates at speed and the gearboxes wear faster than they should. KUKA offers software, KUKA.LoadDataDetermination, that identifies the load by moving the robot through defined motions, and it is far more reliable than an estimate taken from the CAD model of a gripper whose hoses and cables were never modelled. The second is temperature. A robot’s arm and gears grow as they warm up through a shift, and a cell near a furnace or a door to the yard sees more of it. Critical reference points are therefore checked after the robot has reached operating temperature, not only on the cold morning when they were taught.
For applications that need the robot to reach calculated positions rather than taught ones — offline programming, measurement, drilling to a CAD model — robots can be ordered with an absolute accuracy option in which the individual arm is measured and its deviations are modelled in the controller. It is a real improvement, and it still rests on the same chain: a stable floor, correct mastering, measured TOOL and BASE frames and correct load data. No model compensates for an anchor that is working loose.
Commissioning ends with verification, not with the last calibration menu. We run a short test program that approaches a set of reference points on the fixtures from different orientations, record the deviation against the agreed tolerance, and file the result with the controller archive and the anchor torque record. When a customer calls two years later because parts have started to miss, that single page answers the first question — has the robot changed, or has the cell? — in minutes rather than in a day. Floor calibration is not a menu in the controller. It is the discipline of making sure the robot’s world and the real one are the same place, and of keeping them that way.
— GANI Engineering engineering team