Velocity feedforward (Vff) and acceleration feedforward (Aff) gains sit outside the reactive PID path and inject commands proportional to the trajectory's first and second derivatives, with Aff typically scaled by the total moving inertia (load + rotor reflected through the drive) and Vff scaled by the viscous friction coefficient in N·s/m [S1][S4].
They are specified on virtually every modern motion controller and most amplifier-integrated drives, and they deliver the largest single tuning gain on point-to-point moves where trajectory jerk is bounded and the constant-velocity segment is non-trivial [S1][S3].
What Vff and Aff actually compute, and where they sit in the loop
Inside a cascaded position/velocity/current loop, the position PID output is summed with two feedforward terms: a velocity term equal to commanded velocity × Vff, and an acceleration term equal to commanded acceleration × Aff, both added before the current loop [S1][S3]. The current loop itself is a PI regulator that forces measured motor current to track the summed command, which is why feedforward at higher levels translates into torque (current) commands proportional to inertia and friction rather than into voltage [S1].
Feedforward is computed in the controller using the reference trajectory, not the feedback, which is the structural difference versus PID: a feedback gain only reacts after the error exists, while a feedforward term pre-loads the command so the error should theoretically be zero from the instant the motion begins [S2][S4]. In practice the term is approximate because the model (inertia, friction, gravity) is approximate, so a residual error remains and PID still has to clean it up [S1].
Starting-point numerical values, and how to walk them in
Acceleration feedforward Aff is dimensionless in normalised-controller notation and physically represents the inverse of the system inertia reflected to the motor shaft; a good first estimate is Aff ≈ 1/Jt where Jt is the total inertia in kg·m² seen at the motor, after which Aff is trimmed by watching following error during accel and deceler [S1][S4]. Velocity feedforward Vff starts at the viscous-friction estimate B (N·s/m) divided by the motor torque constant Kt (N·m/A), then is walked up until the constant-velocity following error on a linear motion stage flattens to near zero [S1][S3].
A common rule of thumb: increasing Vff reduces constant-velocity error roughly linearly until you hit amplifier saturation or audible high-frequency noise, at which point you back off 10-20%; the same procedure on Aff flattens the accel and decel error lobes on a position-versus-time plot [S1][S4]. On most commercial controllers the gains are entered as unitless percentages or per-unit of full scale, and overshooting the optimum by 2-3x typically shows up as a sharp following-error overshoot at the transition from accel to constant velocity, not as servo instability [S1].
Comparison: PID-only, PID plus Vff, PID plus Vff plus Aff, and full model-based feedforward
The four configurations differ along four decision criteria a controls engineer actually uses: following error at constant velocity, following error during accel/decel, sensitivity to load inertia changes, and commissioning time [S1][S4].
PID-only: high error in all phases, robust to load variation, fastest to commission. PID plus Vff: constant-velocity error drops dramatically while accel lobes remain; tuning time grows by minutes. PID plus Vff plus Aff: both lobes flatten, the controller now requires an accurate inertia model, and load changes of more than ~20% require re-trim. Full model-based or table-driven feedforward (corection for backlash, cogging, and linkage compliance): best accuracy, but requires a characterisation pass and is typically only justified when end-point repeatability below 1 µm is the spec [S1][S5].
Failure modes and limits engineers actually hit
Feedforward assumes the model parameters are correct, so the most common field failure is Aff tuned on an empty stage that is then loaded with a payload, after which the accel lobes reappear inverted; rule of thumb is to retune whenever the load inertia changes by more than ~25-30% [S1][S4]. A second failure mode is over-tuning Vff past the point where the current loop cannot follow the demanded dI/dt, which shows up as a high-frequency buzz on the motor and an increase, not a decrease, in following error at the start of accel [S1].
Feedforward also does not fix mechanical problems: backlash, stick-slip friction, and resonant compliance above the current-loop bandwidth will still produce position errors that no amount of Vff/Aff can cancel, which is why PMD and other controller vendors pair feedforward tuning with a notch or low-pass filter on the velocity-loop output for compliance modes [S1][S6]. A practical ceiling: feedforward becomes diminishing returns below about 1 µm following error on a linear stage, because at that point encoder noise and thermal drift dominate and the model is no longer the limiting factor [S1][S5].
Profile-dependent behaviour: trapezoidal, S-curve, and jerk-limited moves
Aff directly compensates the acceleration term, so on a trapezoidal profile (constant accel, constant velocity, constant decel) you will see two clean error lobes that Aff flattens; on an S-curve profile (jerk-limited), the third derivative becomes non-zero and the error that remains after Aff is the jerk-limited rounding at the transitions, which is why some high-end controllers expose a jerk feedforward term or use a higher-order polynomial reference [S1][S6]. Reference design for high-precision motion systems is increasingly jerk- and even snap-limited rather than acceleration-limited, which pushes the burden of accuracy back onto trajectory shaping instead of into the controller gains [S6].
For a related engineering discussion of how feedforward interacts with mechanical compliance and load variability on servo-driven construction machinery and equipment drives, see this field writeup on HV-rated resistance boxes for insulation tester calibration; for a complementary look at gear and belt-driven systems where feedforward interacts with tooth-profile geometry, the timing-pulley PLD compensation note is directly relevant.
Standards, instrumentation, and sourcing signals
Feedforward is a controller-internal signal path and is not itself the subject of a single ISO or IEC standard; instead, the surrounding motion-control vocabulary is governed by general standards on servo and stepping drives (e.g. IEC 61800-7 for drive profiles), while the verification side, measuring following error with a laser interferometer or linear encoder, falls under checking and testing apparatus practice such as ISO 230-2 for positioning accuracy on machine tools [S1][S3].
Trackable signals: the migration of feedforward from a tuning-only feature to an auto-tuned output on mainstream servo drives (most vendors now expose inertia-estimation routines that write Aff directly), the widening use of S-curve and jerk-limited reference generation in mid-range controllers, and the appearance of table-driven (lookup) feedforward for compensating repeatable torque disturbances on high-end stages [S1][S5][S6].