Impedance Gain Tuning

Impedance and gain tuning

Use this skill after geometry and IK are stable enough that a control response is meaningful. It covers software outer loops, hand/arm command shaping and impedance behavior; it does not assume that the vendor SDK exposes the same loop or units.

Identify the actual loop first

Record the active checkout and configuration, control layer (position, velocity or torque), vendor inner-loop gains, command and feedback units, sample period dt, delay, filtering, saturation, rate limits and watchdog behavior. Check whether a derivative term uses measured velocity; differentiating a setpoint step can cause derivative kick. Do not tune an outer PID on top of an unknown inner PID.

For a torque impedance loop, use the model as a reference:

tau = Kp * (qd - q) + Kd * (qd_dot - q_dot) + tau_ff.

The usual units are N m/rad and N m s/rad, subject to the drive's actual interface. An estimate such as Kd ≈ 2*zeta*sqrt(M*Kp) is only an initial value; identify or bound the effective inertia M and then measure the response.

Tuning procedure

1. Establish a safe offline, replay, simulator or fake baseline with identical input, timing and limits.
2. Tune an inner loop before an outer loop. Within one loop, change one variable at a time; do not use larger Kp to hide retargeting or IK errors.
3. Sweep Kp over a bounded range, then Kd and derivative filtering. Keep position, velocity, acceleration, jerk, torque/current and workspace limits active.
4. Evaluate step and representative trajectory responses using rise time, settling time, overshoot, steady-state error, RMS tracking error, oscillation spectrum, delay, saturation rate and jerk.
5. Keep a reversible parameter record and rollback point. Stop on instability, repeated saturation, watchdog trips or limit violations.
6. Only after the offline/fake result is stable should a separately authorized HIL review consider a physical run.

Report whether each result is source, static, replay/simulation, runtime, HIL or physical-safety evidence. Never present a gain sweep or a stable simulator response as proof that a real robot is safe.