One autonomy platform from the Banbury to the grading machine
Treadum is a perception, decision and actuation layer that runs at the factory edge, orchestrates nine process agents, simulates every change against an as-built tire twin, and writes back only inside a control envelope your engineers signed.
Deployed alongside the equipment you already own
Four layers, one loop
Each layer is replaceable. None of them is a dependency for a line decision.
Perception
Line-scan and RGB vision, thermal, X-ray, shearography, uniformity, PLC, encoder and historian streams aligned on one timeline with Holoscan so timing errors do not become localization errors.
Reasoning
Process, defect, cure-state and uniformity models plus a tire-knowledge agent grounded in your recipes, specs and standards, served by Triton and NIM at the edge.
Simulation
The as-built tire twin evaluates a proposed mix, build or cure against uniformity, quality, energy and takt objectives before the change reaches the floor.
Actuation
Bounded setpoint writes and motion plans through the control envelope, with Isaac ROS for green-tire positioning, press load and unload and AGV handoffs.
What happens between two tires
A single green tire passes through this sequence in the time the drum indexes.
- Ingest: Station sensors, drive telemetry and press traces stream into the edge runtime and are timestamp-aligned to the tire identity.
- Infer: Defect, cure-state, rheology and uniformity models run locally on Jetson-class hardware, targeting sub-100 ms hold decisions. [ASPIRATIONAL]
- Decide: The orchestrator scores proposed actions against the envelope, the twin and the current quality and takt objectives.
- Actuate: In-envelope actions are written to the line; out-of-envelope proposals are queued for engineer approval with full evidence attached.
- Record: The action, its inputs, its citations and its outcome are appended to the genealogy record and the immutable audit log.
Autonomy is earned one gate at a time
No plant hands over a press on day one, and we do not ask them to.
Watch first, prove the baseline
Treadum reads the line and predicts what it would do without writing anything. You compare its calls against your engineers and your downstream inspection results for a full production cycle across shifts, constructions and compounds.
- Zero write access
- Baseline scrap, first-pass yield, uniformity, cure energy and takt captured
- Accuracy and false-hold rates reported per station
Recommend, and let a human commit
Agents propose setpoints and holds with the evidence attached. Operators and process engineers accept, modify or reject, and every one of those decisions is training signal for the site model.
- Human commits every change
- Evidence and citations attached to each proposal
- Override reasons captured as labeled data
Act inside a signed envelope
Once accuracy and safety gates pass, low-risk actions execute automatically inside limits your process and safety engineers signed. Anything outside the envelope still stops for a human, and every envelope change is versioned and reviewable.
- Per-action risk class and hard limits
- Versioned, signed control envelope
- Instant rollback and one-key return to assist mode
{
"site": "plant-nl-02",
"signed_by": ["process.eng", "safety.ehs", "ops.dir"],
"actions": [
{ "id": "cure.dwell", "class": "low",
"delta_s": [-12, +12], "auto": true },
{ "id": "build.drum_tension", "class": "low",
"delta_pct": [-2, +2], "auto": true },
{ "id": "mix.recipe_change", "class": "high",
"auto": false, "approver": "process.eng" }
]
}Why this needs GPUs, not a CPU box in a cabinet
Treadum fuses high-rate image and sensor streams while serving multiple models per tire — inspection, cure state, uniformity, optimization and guarded actuation.
Jetson at the station
DeepStream and TensorRT process RGB, line-scan, thermal, X-ray and shearography streams for splice-open, bubble, separation, bare-spot, cord and cure-risk signals at 30–120 FPS aggregate depending on sensor mix. [ASPIRATIONAL]
Triton and NIM
Plant and fleet models — defect classifiers, segmentation, cure-state and uniformity predictors, anomaly models, control policies and the tire-process RAG service — served from one runtime with versioning and rollback.
Weekly site releases
Multi-modal models fine-tuned on linked genealogy, inspection, PLC, historian and engineer-correction data, gated by golden datasets. [ASPIRATIONAL]
Omniverse twin
Builders, presses, molds, conveyors, robots and WIP buffers simulated to predict uniformity, quality and takt before a production change.
cuOpt scheduling
Mold allocation, cure sequencing, WIP movement, AGV routing and line balancing solved under quality, capacity, due-date and energy constraints.
Everything hangs off the tire
One identity links the compound batch to the final grade, which is what makes root cause fast and audits boring.
| Entity | Linked from | Used by | Retention |
|---|---|---|---|
tire | Building drum event, RFID or laser mark | Genealogy, uniformity, grading, recall scoping | Site policy, typically 10 years |
batch | Banbury mix cycle and downstream mills | Rheology models, drift detection, escape scoping | Site policy |
component | Extrusion and calendering rolls | Gauge control, splice analysis, defect localization | Site policy |
cure_trace | Press and mold telemetry | Cure-state estimation, energy accounting | Site policy |
finding | Vision, X-ray, shearography, uniformity | Hold decisions, non-conformance, model training | Site policy |
action | Agent proposal and engineer decision | Audit log, envelope review, model training | Immutable |
Guardrails that survive an audit
Autonomy in a safety-critical plant is a governance problem as much as a modeling problem.
Grounding and citations
Every recommendation cites the recipe, spec, standard or historical trace that justified it. Ungrounded outputs are blocked, not softened.
Human-in-the-loop gates
Graduated autonomy per action class, with an explicit approver role and a one-key return to assist mode for any station, line or site.
Continuous evaluation
Golden datasets and LLM-as-judge evaluation gate every model and prompt change in CI. A regression on a rare defect family blocks the release.
Tenant isolation
Per-tenant model and retrieval isolation with strict compound and construction IP protection, plus an optional fully on-premises deployment.
Immutable audit log
Assurance-grade record of every agent action — inputs, evidence, decision, outcome and the human who approved it.
Rollback by default
Model, prompt and envelope versions are pinned per site. Any release can be rolled back without touching the line schedule.
From first walk to bounded autonomy
A typical design-partner engagement, with the gate that has to pass at each step.
Plant walk and wedge selection
We map the line, pick one wedge with a hard metric, and agree the baseline instrumentation. [ASPIRATIONAL]
Connect and baseline
Connectors land inside the OT network, historian and inspection streams are aligned, and the pre-Treadum baseline is captured across shifts and constructions.
Shadow mode
Predictions run against live production with zero writes. Gate 1: perception accuracy and false-hold rate against engineer and downstream agreement.
Assist mode
Engineers commit every proposed change. Gate 2: acceptance rate, review latency and measured effect on the wedge metric.
Bounded autonomy
Low-risk actions execute inside the signed envelope. Gate 3: safety review, rollback drill and a signed control envelope from process, safety and operations.
Operators on the first weeks
“We had three people watching X-ray images and still shipped uniformity rejects. Treadum flagged the cord shift at the building drum, not four hours later at final inspection.”
“The cure agent found 40 seconds of margin on a construction we had run the same way for eleven years. It proved it in shadow mode before it touched a press.”
“Genealogy is the part I did not expect to care about. Every tire now has a linked record from batch to grade, and audit prep went from weeks to an afternoon.”
[PLACEHOLDER] Design-partner quotes are illustrative until pilot references are published.
What we hold ourselves to
Targets are set per site during baselining and reported in the same console the operators use.
How it behaves in a real plant
No. All line decisions are made by the factory-edge runtime. Cloud is used for training, fleet management and multi-site reporting. A site can run disconnected indefinitely and reconcile when the link returns.
Perception models for common defect families transfer quickly and are usually credible within the shadow-mode window. Site-specific cure, rheology and uniformity models improve fastest where you have historian depth and linked inspection results. [ASPIRATIONAL]
Yes — that is the Cell plan. One building or curing cell or line, one wedge metric, one baseline. Expansion to the rest of the line is a configuration change, not a redeployment.
Common. We read whatever the machine exposes — API, file drop, frame grabber or camera tap — and fuse it with our own sensing where the existing coverage is thin.
Put one cell on autonomy in 90 days
Pick one wedge — splice inspection, X-ray defect detection, cure-state optimization or uniformity prediction. We baseline it, run shadow mode, then graduate to bounded autonomy under a signed control envelope.
Pilots start in shadow mode. No line changes until accuracy and safety gates pass.