When proprioception or force/torque is mid-band ringy, models and compliance layers inherit junk. GSRF Practical is a candidate deterministic pre-stage: mid-band damp + peak compression toward a declared normal — before the expensive model, not instead of the controller.
Good fits (stack placement)
- Force/torque or current traces with mid-band chatter before feature extraction or impedance logic
- Proprioceptive series that should stay near a soft normal rather than chase every spike
- Offline dataset hygiene for imitation / residual learning (strictly positive or offset signals)
- Per-channel instances if many joints — same filter, different
x*(see scaling note on actuators guide)
Do not treat “50+ actuators” as a performance claim. Treat it as why noisy channels hurt more at fleet scale — and why pricing for multi-channel OEM is Enterprise-class (pricing).
Bad fits (published)
- Closed-loop command smoothing under multi-minute plant/network delay — locked packs show worse MAD-to-measured vs EMA
- Best-in-class tracking of a fast-moving reference (MAD 0/52 vs best EMA)
- Safety-critical residual trip logic at fixed FA rate
Eval path
Log a positive series → gsrf-bench → compare osc31/peak vs your EMA baseline → shadow in software before any command path → commercial license for production robots.
Related: Autonomous control · Delay failure note · GSRF vs EMA vs Kalman