Integral windup happens when a controller's output is stuck at a limit but its integral term keeps growing. When the error finally changes sign, the large stored integral makes the output stay saturated too long, which gives a big overshoot. Anti-windup methods stop the integral from growing when it cannot help. It is a practical problem of Proportional, PD and PID Control.
How Windup Happens
- a motor is commanded faster than it can go, so the controller output hits its maximum (for example 100% duty)
- the error stays large, so the integral keeps adding it up
- when the target is finally reached, the integral is huge, so the output stays at maximum and the system overshoots
- the integral then needs a long time to unwind
Integral Clamping
- limit the integral itself to a range,
- simple, but the limit must be tuned
Conditional Integration
- stop adding to the integral while the output is saturated (or undo the last addition)
- simple and effective for most robots
Back-Calculation
When the unclamped output exceeds the limit , remove the excess from the integral:
- because the integral term is , this removes exactly the part that caused the excess
- the integral then sits at the value that just reaches the limit
Other Good Practice
- reset the integral when the target becomes zero or the controller is turned off
- skip the derivative on the first step after a reset, so a stale previous error does not cause a spike
- limit the controller's share of the output when a feedforward term already does most of the work
Example In Code
integral += error * dt;
integral = std::clamp(integral, -i_max, i_max); // clamping
float u = kp * error + ki * integral + kd * d_error;
if (u > u_max) {
if (ki > 1e-6f) integral -= (u - u_max) / ki; // back-calculation
u = u_max;
}