Industrial actuators — kill mid-band ring without pretending to track

Technical identity (why this is not “GSRF for robots”)

GSRF Practical is a soft thermostat in log-space: spring k_return toward a declared normal x*, tanh-bounded observation pull. Public locked strengths are mid-band ring damp and peak compression toward that normal on characterization packs, plus pathology-scoped industrial synthetic wins (stiction-class osc; provisional slow-tracking wear).

That identity maps to proprioceptive / force / OP–PV traces before the expensive model or before a carefully supervised joint path. It does not map to delayed closed-loop command smoothing, Kalman-style state fusion, servo-drive internals, or fixed false-alarm residual detection — and the site refuses those claims.

Why actuator density matters (commercial thesis)

Industrial robots already carry multiple joints. Humanoids and next-gen mobile platforms push well past 20–30 degrees of freedom; some designs head toward 50+ actuators when you count redundant joints, grippers, and secondary axes. Each loop is a potential source of mid-band oscillation, force/torque noise, and command chatter.

Density multiplies cost of noise: one slightly ringy joint is an annoyance; fifty slightly ringy loops are a fleet problem — more model input garbage, more thrash in learned policies, more mechanical chatter budget, more engineering attention per platform. The filter is intended as a per-channel pre-stage (each instance its own state and usually its own x*), not as a single black box for the whole robot.

True command+measured robot fleet (n=298 continuous)

Primary industrial ladder remains valve/SCADA-style OP/PV (DAMADICS hardness) and synthetic packs. A valve limit-cycle is still not automatically a joint oscillation — but we now have a separate, Workbench-scored robot row with native command + independent measured (not EMA-of-measured as demand).

Wear holds under true command. Proxy demand inflated oscillation — the true story is better, just different.

*Rate wins concentrate on ABB TCP; NIST and Franka joint packs are 0 on rate — do not merge into one robot rate figure. Osc is surface-dependent (weak on RMPD TCP position; stronger on joint tracking). Frozen Phase 5 MAIN vs best EMA. Sources: Zenodo RMPD 18549433 · NIST UR5 PHM · HF DROID abs-joint subsets. Full ledger: Evidence → true cmd.

EMA-proxy robot packs (labeled separately)

Earlier public rows used EMA proxy of measured as demand when native command was missing: KUKA LWR4+ external torque (Zenodo 6461868) — wear 21/21, osc 21/21 partial; collab joint pos/effort (Zenodo 7240050) — n=48, wear 48/48, osc 42/48, 0 loss. Those remain a proxy proprioception candidate — not true-command evidence. Proxy osc rates must not be quoted as true cmd.

Still open: partner native command + joint_states at plant rate; multi-rate stress on your hardware. Multi-joint density remains the commercial pressure story; claims stay metric- and source-labeled.

Competition (honest angle)

Servo drives and PLC ecosystems already embed EMA, median, and Kalman-class conditioning (Beckhoff, Siemens, Rexroth, ROS2 packages, etc.). Our angle is not “we smooth signals.” It is:

• Kill mid-band ring without pretending to track (MAD tracking is a published non-goal)
• Peak compression toward a declared normal when that is the product identity
• Publish where we fail (delay command paths, residual FA vs EMA, plant composite hardness, independent overall limits)

That is the differentiator against generic drive filtering and open-source smoothing recipes.

Scaling note (fleet of instances)

Deployment mental model: N independent GSRF instances for N channels, each with small constant state (filter memory + declared or adaptive x*). CPU cost is linear in channels and samples — a lightweight pre-stage, not a multi-body dynamics solver. Different joints may use different x* and presets; the frozen evaluator discipline still applies per series when you characterize.

We have not published a multi-robot timing benchmark on this site. Treat “cheap pre-stage” as architectural intent; measure wall-clock on your target rate (e.g. 100 Hz–1 kHz joint telemetry) during eval.

Where locked evidence supports use

How buyers estimate value (no invented $)

Three cost classes that flow from verified metrics — not vendor plant ROI calculators:

1. Actuator motion class (mechanical) — Mid-band ring and peak chase correlate with extra travel and chatter. On green pathologies, lower osc / wear proxies are the Workbench signal. You convert travel/tickets into MTBF or parts $ after a shadow run.
2. Pre-model hygiene (PdM / industrial AI) — Calmer mid-band inputs can reduce junk features into models. We do not claim “3× faster training” or GPU $ — log train time on a fixed recipe yourself. See also PdM pipelines.
3. Not residual FA man-hours — GSRF is not a fixed false-alarm residual detector (EMA residual recall wins that locked match). Do not budget “hours saved on ghost residual alarms” from our Evidence.

Red lights (actuators & plants)

• Overall independent industrial synthetic — wear ~18% [12–27%], osc ~24% [17–33%], n=100; multi-regime & mixed strata 0% at 50% bar
• Real plant DAMADICS multi-day — composite mostly loss; wear ~1/75 — not a plant wear championship
• Clean tracking / integrating — do not force GSRF as EMA replacement for pure track
• Multi-minute delayed command loops — published red light for delay misuse
• Fixed-FA residual detection — not the product; locked probes favor EMA residual
• Unscoped “~95% adversarial wear” — generator-family only, not independent law

Where it sits in the stack

Proprioception / force / OP–PV (positive or offset)
→ GSRF Practical (per channel; mid-band damp + peak pull when domain fits)
→ features / model · display · shadow before any command influence
Joint controller · drive filters · Kalman fusion · residual FA → stay separate

Scale & pricing anchors (for OEMs and fleets)

A 50-joint platform is not a hobby-tier conversation. Indicative packaging on Pricing (final terms by agreement):

Channel counts are negotiation anchors, not hard SKU meters. Quantum is a separate ladder.

Pathology-class pilot (pre-declare success)

1. Pick N valves or (when available) joint channels with a named pathology — mid-band hunt, stiction-class ring, slow-track wear class — not “all joints.”
2. Export OP/PV or demand/measured, positive or offset; freeze a normal for x* (per channel if fleet).
3. Run Try or gsrf-bench vs EMA: osc · wear proxy · MAD.
4. Write success bars before the run. Measure your own cost class (travel, tickets, train thrash) — we do not invent plant $.
5. Shadow only; kill switch = bypass. Production after safety cage.

Sources: True cmd robot fleet · Independent map · Industrial Evidence · Phase 5 batteries note · When GSRF lost · Citations · Contact info@boonmind.io