sensor bias & drift
Bias and drift are two kinds of lie that a sensor tells, even when nothing is wrong with it. Bias is a steady, constant offset: a bathroom scale that always reads two kilograms heavy, or a robot's spin sensor that quietly reports a tiny turn even while the robot sits perfectly still. The error is always there, always the same size. Drift is the sneakier cousin — a slow wandering of that offset over time. The scale that reads two kilos heavy this morning might read three kilos heavy by tonight as it warms up. The reading creeps, even though the thing being measured never changed.
These small errors matter enormously when a robot adds many readings together over time, a trick called integration. An accelerometer or gyroscope inside an inertial measurement unit doesn't sense position directly; the robot must keep summing tiny motion readings to guess where it is now. But every reading carries that little bias, and summing thousands of them per second piles the error up fast. A gyroscope that is off by a hair will, after a minute of adding up, insist the robot has rotated a noticeable amount when it has not. This is why a phone's step counter or a drone's pure motion-tracking slowly loses its way — the bias has been compounding the whole time.
Engineers fight back in three ways. They measure the bias while the sensor is known to be still and subtract it off (calibration). They model how the drift creeps and predict it. And, most powerfully, they bring in a second sensor that does not drift away over time — like GPS or a camera watching fixed landmarks in the room — to gently correct the wandering one. That blending of an accurate-but-noisy sensor with a smooth-but-drifting one is exactly the job of sensor fusion and state estimation.
Leave a cheap drone resting on a table and watch its on-screen heading slowly rotate by a few degrees per minute, all by itself. Nothing is moving — that creeping number is the gyroscope's bias being integrated into a phantom turn, the textbook face of drift.
A still sensor whose reading won't sit still: bias, integrated over time, becomes drift.
Bias is constant and easy to subtract once you have measured it; drift changes over time, so a one-time calibration is never enough — it must be re-checked or corrected on the fly. Temperature is a common cause of both, which is why some sensors carry their own little thermometer to compensate.