The loop

Sense, model, simulate, decide, act, prove

Six layers, closed into one loop that runs at the mill edge. Break any one of them and autonomy becomes a dashboard.

Sense

Line-scan RGB and NIR cameras, spectrophotometers, evenness meters, bath probes, PLC events, water and energy meters, and MES context fused into a roll-level state stream.

Model

Fine-tuned defect, shade, yarn-quality, tension, recipe and finishing models served from the mill edge, with time-series prognostics for equipment drift.

Simulate

An as-produced fabric-line twin that runs candidate recipes, tensions and sequences before a machine moves, and rejects the ones that miss shade or takt.

Decide

A mill orchestrator that plans, escalates on policy, and picks actions under delivery, chemistry, equipment and utility constraints.

Act

A typed, scoped tool layer that writes bounded setpoints to real machines and drives robotic handling, always reversible, always logged.

Prove

Immutable roll genealogy, conformance packets and sustainability evidence, exported for buyers and auditors on demand.

Orchestration

The orchestrator plans; the agents own their stage

A run is a plan over agents. The orchestrator decomposes the goal, routes work, holds the policy gates and keeps a single record. Agents do the physics.

RUN-4471 10 steps · 1 approval gate · 1 recovered failure · illustrative scenario COMPLETE

Scroll the graph sideways · or open the text equivalent below

SUCCEEDED APPROVAL FAILED RUNNING

Text equivalent, workflow steps, dependencies and status

Run RUN-4471 workflow graph, as a table
Step	Stage	Depends on	Status	Duration
ingest	conformance	none	SUCCEEDED	0.8s
twin	simulate	ingest	SUCCEEDED	46s
recipe	dye + colour	ingest	SUCCEEDED	2.4s
approve	human gate	twin, recipe	APPROVAL	4m 12s
dose	bath control	approve	SUCCEEDED	118m
shade	verify ΔE	dose	SUCCEEDED	9.1s
finish	stenter	dose	SUCCEEDED	64m
inspect	vision	shade, finish	FAILED	38m
rework	replan	inspect	SUCCEEDED	1.6s
grade	release	rework	SUCCEEDED	3.0s

Parallel branches2twin simulation and recipe prediction run together
Human gates1colourist approval, 15 minutes, attributed
Failures recovered1weft streak on roll 09, replanned in-run

Run timeline

RUN-4471 · Indigo poplin, 3,180 kg, due Thursday

The same scenario run as an ordered timeline: step, status, duration and the tool result behind it. Inspection failed on one roll, the run recovered, replanned the cut and still held the ship date.

RUN-4471 RUN COMPLETE Scenario mill · reactive exhaust jet dyeing · illustrative run, not a customer deployment

Goal: Dye lot DL-4471 · 3,180 kg · 40s combed cotton poplin · shade Indigo 19-4028 TCX · ΔE ≤ 0.8 · due Thu 06:00

01 ingest.lot_state SUCCEEDED 0.8s

Pulls greige genealogy for 14 rolls: yarn lots, loom, greige GSM 138, absorbency and residual size from the prep line.
Agent, input and output
```
agent   Quality-and-Conformance
result  ← 14 rolls · 3,180 kg · mean GSM 138.2 (σ 1.1) · prep OK
```
02 twin.simulate_lot SUCCEEDED 46s

Omniverse fabric-line twin simulates the exhaust curve, machine loading and takt for three candidate recipes before a drop of dye is used.
Agent, input and output
```
agent   Yield-and-Takt
result  ← 3 candidates scored · best predicted ΔE 0.52 · 118 min cycle
```
03 dye.predict_recipe SUCCEEDED 2.4s

Spectral recipe model proposes a 3-dye combination against the buyer standard, corrected for this substrate’s absorbency and the current water hardness.
Agent, input and output
```
agent   Dye-and-Color
result  ← 2.14% Navy RGB · 0.61% Blue BRF · 0.08% Black B · salt 62 g/L
```
04 human.approve_recipe APPROVAL 4m 12s

Recipe deviates >5% from the standing card, so autonomy is capped: the colourist approves, rejects or edits before any dosing. Guardrail policy MILL-02/DYE-APPROVAL.
Agent, input and output
```
agent   Dye-and-Color
result  → awaiting colourist · approved on shift 04:18
```
05 dye.control_bath SUCCEEDED 118m

Closed-loop dosing and ramp control against live bath spectrophotometry; two mid-cycle corrections applied when exhaustion ran 4% ahead of model.
Agent, input and output
```
agent   Dye-and-Color
result  ← exhaust 96.4% · 2 corrections · bath ΔE 0.41 at drain
```
06 shade.verify_lot SUCCEEDED 9.1s

Per-roll spectral read against Indigo 19-4028 TCX under D65/TL84/A; metamerism scored and rolls sorted into shade bands.
Agent, input and output
```
agent   Shade-and-Quality
result  ← mean ΔE 0.58 · 13 rolls band A · 1 roll band B
```
07 finish.control_stenter SUCCEEDED 64m

Stenter chemistry, overfeed and curing profile set to hold hand, 148 cm width and residual shrinkage under 3%.
Agent, input and output
```
agent   Finish-and-Hand
result  ← width 148.3 cm · shrink 2.1% warp / 1.4% weft
```
08 inspect.scan_rolls FAILED 38m

In-line vision flags a recurring 6 cm weft streak on roll 09 between 412 m and 445 m, traced to a stenter pin-chain slip, not a dye fault.
Agent, input and output
```
agent   Defect-and-Inspect
result  ← 1 defect cluster · roll 09 · 33 m affected · class WEFT_STREAK
```
09 plan.rework_route SUCCEEDED 1.6s

cuOpt re-plans the cut and rework route: 33 m diverted to seconds, remaining 14,890 m re-sequenced so the buyer’s Thursday ship date still holds.
Agent, input and output
```
agent   Yield-and-Takt
result  ← 33 m to seconds (0.22%) · ship date held
```
10 quality.grade_and_release SUCCEEDED 3.0s

Four-point grading, roll genealogy and the conformance packet (AATCC 61, ISO 105-C06, ZDHC MRSL declaration) are assembled and released to the buyer portal.
Agent, input and output
```
agent   Quality-and-Conformance
result  ← 13 rolls Grade A · 1 roll Grade B · packet signed
```

RUN OUTCOME

Lot released first time, on shade, on date.

Result of run RUN-4471
Measure	Value	Context
Final mean ΔE	0.58	against a 0.80 tolerance
Re-dyes avoided	1	the standing card would have missed shade
Seconds	0.22%	versus the scenario mill baseline of 1.9%
Water used	46 L/kg	versus 59 L/kg on the scenario recipe card
Human decisions	1	the recipe approval gate

Why this matters

A mill does not fail on the good lots. It fails on the lot where the substrate drifted, the recipe card was wrong, and nobody found out until the buyer opened the carton. Yarneon makes that lot legible while it is still in the machine.

Digital twin

An as-produced twin, not a CAD model

The twin is built from what the mill actually made, every metre, every fault, every bath, every setpoint, so simulation is grounded in this floor, not a generic one.

Roll-level state, metre by metre

Each roll carries its own genealogy: the bale and yarn lot it came from, the loom or machine and shift that made it, the bath it was dyed in, the stenter profile it was finished on, and the defect map recorded at inspection.

Fibre → yarn → greige → dyed → finished → roll lineage
Metre-indexed defect map with class, severity and position
Bath chemistry, exhaust curve and water profile per lot
Counterfactual simulation before any setpoint is moved
Synthetic defect generation for classes the mill rarely sees

Twin fidelity

Simulate before you dose

In the scenario run RUN-4471 the recipe change is simulated against the twin before a single litre moves.

Coverage

Metre-level

Defect map indexed to position.

Latency

< 100 ms

Design target, sensor to inference at the edge. ASPIRATIONAL

Tool-call stream

Every call the agent made, in order

Machines, spectrophotometers, the optimiser and the humans are all tools. Calls go out, results come back, and the whole exchange is real, copyable text, the same record your auditor sees.

tool-call stream · RUN-4471

04:02:11 mes.get_lot("DL-4471") → 14 rolls · 3,180 kg · poplin 40s SUCCEEDED

04:02:12 twin.simulate(recipe_candidates=3, machine="jet-07") RUNNING

04:02:58 twin.simulate → best ΔE 0.52 · cycle 118 min · water 46 L/kg SUCCEEDED

04:03:01 colour.predict_recipe(substrate="cotton", standard="19-4028 TCX") RUNNING

04:03:03 colour.predict_recipe → Navy RGB 2.14% · Blue BRF 0.61% · Black B 0.08% SUCCEEDED

04:03:04 policy.check(action="dose", deviation=7.2%) → requires human approval APPROVAL

04:18:22 human.approve(user="colourist-on-shift", decision="approve", note="ok, hold ramp") SUCCEEDED

04:18:24 plc.dose(machine="jet-07", schedule="ramp-B") RUNNING

05:41:07 spectro.read(bath) → exhaustion 96.4% · bath ΔE 0.41 SUCCEEDED

06:22:40 vision.scan(roll=09) → WEFT_STREAK 412–445 m · severity 3 FAILED

06:22:41 cuopt.replan(cut_plan, constraint="ship_thu_0600") RUNNING

06:22:43 cuopt.replan → 33 m to seconds · ship date held SUCCEEDED

06:25:02 quality.release(lot="DL-4471") → 13 Grade A · 1 Grade B · packet signed SUCCEEDED

Guardrail in action

policy.check → requires human approval

At 04:03:04 the proposed recipe deviated 7.2% from the standing card, above the 5% autonomy ceiling for dye dosing at this mill. The agent stopped and asked. Fifteen minutes later a colourist approved it, and that approval is attributed, timestamped and reversible.

APPROVAL GATEAPPROVED 04:18

Failure handled honestly

vision.scan → WEFT_STREAK, roll 09

A 33 m weft streak is a real defect, and the run shows it as failed rather than smoothing it over. The mechanical root cause, a stenter pin-chain slip, was correlated automatically and raised as a work order.

STEP FAILEDRUN RECOVERED

Accelerated computing

Physical AI needs real compute, at the machine

Line-speed vision, spectral colour science and a fabric-line twin are GPU-essential workloads. Yarneon runs them at the mill edge and trains them centrally.

The accelerated-computing stack behind the mill loop
Component	Role in the mill
Isaac + Jetson	Robotic roll, beam and doffing handling, plus mill-edge inference on inspection and control cells.
Metropolis / DeepStream + TensorRT	Line-speed fabric-defect vision pipelines targeting sub-100 ms per-frame decisions. ASPIRATIONAL
Holoscan	Sensor fusion across cameras, spectrophotometers, probes, PLC events and meters into one explainable state stream.
Omniverse + Replicator / Cosmos	The as-produced fabric twin, and synthetic generation of rare defect, barré and shade-variation cases.
cuOpt	Dye-lot sequencing, rework queues, machine assignment and water/energy allocation under hard constraints.
Triton + NIM + NeMo	Multi-model serving at the mill edge, plus grounded textile-process and colour reasoning with citations.
NVIDIA AI Enterprise	Private and on-prem deployment for groups protecting dye recipes and construction IP.

Items marked ASPIRATIONAL describe target architecture ahead of full production validation.

Reasoning trace

The reasoning, written down

Goal, observation, plan, action, decision, handoff, outcome. A colourist can read this and disagree with it, which is the point. Autonomy a mill cannot argue with is autonomy a mill will not adopt.

Goal01 / 10

Land DL-4471 on Indigo 19-4028 TCX within ΔE 0.80, first time, before the Thursday 06:00 ship cut-off.
Observation02 / 10

Greige absorbency for this lot reads 8% above the card baseline, and incoming water hardness is 142 ppm, both push the exhaust curve faster than the standing recipe assumes.
Plan03 / 10

Simulate three recipes in the fabric twin before committing dye. Prefer the candidate with the lowest predicted ΔE that also fits the 118-minute machine slot.
Action04 / 10

twin.simulate(recipe_candidates=3, machine="jet-07")
Observation05 / 10

Candidate B predicts ΔE 0.52 and finishes 6 minutes inside the slot. Candidates A and C either overshoot shade or miss takt.
Decision06 / 10

Candidate B deviates 7.2% from the standing card, above the 5% autonomy ceiling for dye dosing. Escalate to the colourist rather than act.
Handoff07 / 10

Dye-and-Color → the colourist on shift, with the simulated curve, the deviation reason and a one-click revert attached.
Action08 / 10

plc.dose(machine="jet-07", schedule="ramp-B") after approval, with two mid-cycle corrections when exhaustion led the model by 4%.
Observation09 / 10

Vision flags a 33 m weft streak on roll 09 after finishing. Correlating stenter telemetry shows a pin-chain slip at 04:51, a mechanical cause, not a dye fault.
Outcome10 / 10

Lot released at mean ΔE 0.58 with 0.22% seconds. A maintenance work order is raised against the stenter pin chain, and the streak signature is added to the mill’s defect memory.

Human in the loop

Craft is captured, not replaced

Textile mills run on scarce craft: the colourist who knows this jet runs hot, the weaving master who hears a beam going wrong. That knowledge is retiring faster than it is being replaced.

Yarneon writes it down. Every approval, override and correction becomes per-mill memory scoped to your tenant, so the mill keeps improving after the person who taught it has gone home.

Per-colourist and per-operator performance memory
Overrides captured as training signal, not noise
Versioned memory, scoped per tenant, never shared across mills

Command line

Start a run from anywhere

The same run, from a terminal. Output is real text, streamed as the mill works, and identical to what the console and the audit log record.

yarneon cli · mill-02

$ yarneon run --mill mill-02 --lot DL-4471 --goal "19-4028 TCX, dE<=0.8"
  resolved 14 rolls · 3,180 kg · 40s combed cotton poplin
  twin.simulate      ✓  46.0s   best candidate dE 0.52
  dye.predict_recipe ✓   2.4s   Navy RGB 2.14% · Blue BRF 0.61%
  human.approve      !   4m12s  approved by colourist
  dye.control_bath   ✓ 118m00s  exhaustion 96.4%
  inspect.scan_rolls ✕  38m00s  WEFT_STREAK roll 09 · 412–445 m
  plan.rework_route  ✓   1.6s   33 m to seconds · ship date held
  quality.release    ✓   3.0s   13 Grade A · 1 Grade B
  
  run RUN-4471 complete · mean dE 0.58 · seconds 0.22% · 1 human decision

Also available as

Python and TypeScript SDKs with typed tool schemas
REST and streaming event APIs for MES and ERP mirroring
Webhooks for approvals, defects and releases
The command palette (⌘K) inside the Yarneon console
An on-HMI panel at the machine, for operators who never open a laptop

Developer platform

Deployment

From first connector to bounded control, one phase at a time

No mill hands over a jet on day one. The path is deliberately slow at the start and faster once the numbers hold. Durations are quoted after the assessment, because they depend on how instrumented the mill already is.

Phase 1

Connect

Mill survey, edge appliance install, connectors to machines, MES and meters. Telemetry starts flowing; nothing is controlled.
Phase 2

Observe

Models run in shadow against live production. Predictions are scored against what the mill actually produced, on your data.
Phase 3

Assist

The agent proposes; a human commits. Approval gates, setpoint envelopes and revert paths are agreed with the process owners.
Phase 4

Control

Bounded auto-control on the setpoints you authorise, one at a time, with autonomy ceilings that you can lower at any moment.

Operating figures

The numbers a mill manager asks about second

Latency targets, connector coverage, autonomy modes and the record: the practical questions after the headline.

<100 ms

Design target for a per-frame inspection decision on Jetson-class edge hardware. ASPIRATIONAL

8

Connector families: looms, frames, dyeing, colour, inspection, MES, utilities, robotics.

3

Autonomy modes: observe, assist, bounded control. Granted per setpoint.

1

Record per mill: append-only, exportable, shared with buyers on request.

Items marked ASPIRATIONAL are design targets ahead of production validation.

Guardrails

An agent earns the right to touch a machine

Nothing here is a global "AI on" switch. Autonomy is granted per setpoint, bounded by policy, reversible by any operator and logged permanently.

Graduated autonomy: Every workflow moves observe → assist → bounded auto-control. Autonomy ceilings are per-setpoint policy, not a global switch.
Human-in-the-loop gates: Approval gates fire on deviation thresholds, new substrates, or any action outside the authorised envelope. Approvals are attributed and time-stamped.
Reversibility: Every control action ships with a revert path and a last-known-good setpoint. Operators can take manual control at any moment, from the HMI or the panel.
Immutable audit log: Assurance-grade, append-only record of every observation, decision, tool call, approval and actuation, exportable for buyer and regulatory audit.
Tenant isolation: Recipes, construction libraries and roll data are isolated per tenant. No cross-tenant training. Permission-aware retrieval with enforced citations.
Sandboxed tools: Agents call machines through a typed, scoped tool layer with rate and range limits. There is no path from a model output to an unbounded PLC write.

Read the security model

Integrations

It speaks to the machines you already own

Yarneon is not a rip-and-replace. It connects to the looms, frames, jets, stenters, cameras, spectrophotometers and MES already on your floor.

Loom & knit controllers

Picanol, Toyota, Tsudakoma, Karl Mayer and Mayer & Cie controllers over OPC UA and vendor PLC tags.

Spinning frames

Rieter, Trützschler, Saurer and Murata frames, draft, twist, speed and end-break telemetry.

Dyeing & finishing

Thies, Fongs, Brazzoli jets and Monforts/Brückner stenters via PLC, recipe systems and dosing units.

Colour & lab

Datacolor and X-Rite spectrophotometers, colour libraries, Pantone TCX standards and lab-dip workflows.

Inspection hardware

Line-scan RGB and NIR cameras, Uster and Elbit inspection frames, existing light boxes.

MES / ERP

SAP, Datatex, TIM, Coats Digital and homegrown mill MES over REST, files and database CDC.

Utilities & sustainability

Water, steam and electricity meters; ZDHC Gateway, Higg FEM and OEKO-TEX reporting.

Robotics

Vision-guided manipulators and AMRs on NVIDIA Isaac and Jetson for doffing, beams and roll transport.

Vendor names are listed for compatibility reference only and do not imply endorsement or partnership.

See the loop on your own line

We run a shadow deployment against your live production and score the predictions against what you actually made.

Book a mill assessment Read the docs

One loop, closed at the machine