Tuning a servo without a control theory degree

Servo tuning has a reputation for being an art, and most of that reputation comes from people adjusting gains in the wrong order and then adjusting them again.

Done in the right order, it converges quickly and the reasoning at each step is straightforward.

The three loops

A servo drive runs three nested control loops. Innermost first.

Current, or torque. How much current to put through the windings to produce the demanded torque. Runs fastest, tens of kilohertz, and is set by the manufacturer from the motor data. Leave it alone.

Velocity. How much torque to demand to reach the demanded speed. Runs at a few kilohertz. This is where most of the tuning happens.

Position. How much speed to demand to reach the demanded position. Runs slowest. Usually only a proportional gain.

Tune from the inside out. A position loop tuned on top of a badly tuned velocity loop will be tuned to compensate for it, and the compensation falls apart the moment the load changes.

Inertia ratio, first and always

The single number that matters most, and the one most often left at its default of one.

It is the load inertia, reflected through the gearbox to the motor shaft, divided by the motor's rotor inertia. Reflected means divided by the gear ratio squared, which is why a modest gearbox transforms an impossible ratio into an easy one.

Every gain the drive computes is scaled by this number. Get it wrong by a factor of ten and no amount of subsequent adjustment produces a well behaved axis.

Autotune measures it. If autotune is impossible, calculate it from the mechanical drawing; even a rough figure beats leaving it at one.

Tuning the velocity loop

The procedure that converges, with the axis making small repeated moves.

• Set integral gain to zero or its minimum.
• Raise the proportional gain until the axis audibly buzzes or oscillates, then reduce it by about a third.
• Add integral gain until the steady state error goes away. Too much shows as overshoot and a slow settle.
• If a resonance stops you raising the proportional gain far enough, find its frequency and place a notch filter on it. Then continue raising.

The position loop

Usually proportional only, and it interacts with the velocity loop directly: raising it too far causes overshoot that looks like a badly tuned velocity loop.

Raise it until the axis reaches position quickly without overshooting, and stop.

Feedforward is the useful addition here. Velocity feedforward tells the velocity loop what speed will be needed before the position error demands it, which reduces following error during a move without touching stability. On a machine that must track a profile rather than reach a point, it is worth more than any amount of gain.

Mechanics beat gains

The part that saves the most time. A large share of untunable axes are mechanical.

• Backlash in a coupling or gearbox means the loop is controlling something that is not always connected.
• A compliant belt or a long shaft adds a resonance that caps the achievable gain.
• An oversized motor on a tiny load gives a very low inertia ratio and can be twitchy rather than better.

If a rigid coupling makes the problem disappear, the problem was never the gains.

Knowing when to stop

The target is not the highest gains the axis will tolerate. It is the lowest gains that meet the requirement, because everything above that is margin spent on nothing.

Set the actual requirement first, following error during a move, settling time, position tolerance at rest, and stop when it is met.