Factory Twin Lab / campus × line × energy

Same shift. Same failure.
Close the loop.

A simulated plant streams noisy, lossy telemetry to its digital twin. The twin mirrors the line, estimates hidden spindle wear, forecasts the billed power demand, tests what-if plans on its own model and sends commands back. Replay the shift with the loop closed, or with the same twin in shadow mode, where it only advises.

Cycles render of the procedural Blender model: a cutaway assembly hall with CNC machines, a robotic weld cell, assembly and inspection stations, cyan twin status rings linked to the operations center, and the surrounding campus.
Blender 5.2 Cycles render of the same recorded sample (10:30, CNC 2 in planned service). Procedural model; no downloaded assets.
  1. PlantSix stations, 5 buffers, 3 AMRs, hall climate, PV, EV chargers. 1 s steps.
  2. TelemetryPLC, CNC, energy, hall and AMR packets every 5 s. 2–3% lost; one sample of latency.
  3. SyncThe twin mirrors buffers, states and AMR poses from the packets it actually received.
  4. EstimateExtended Kalman filter on vibration → CNC 2 wear and wear rate.
  5. PredictEvery 5 min: 14 what-if runs of its own line model over 90 min. 15-min demand forecast every minute.
  6. Decide & actService, feed derate, setpoint and EV caps. Applied after 5 s — or only logged.

Live campus, recorded shift

Loading the recorded shift…

Good parts so far
CNC 2 spindle wear
Twin forecast
Grid exchange now
Hall climate
Latest twin decision

CNC 2 wear: plant truth vs. the twin's estimate

Amber: true wear (hidden from the twin). Cyan: EKF estimate ±2σ. Red dotted: this plant's actual failure point; grey dashed: the twin's nominal threshold. Click to jump.

Good parts: closed loop vs. shadow

Same seed, same disturbances. The only difference is whether commands reach the plant.

Grid import and 15-minute billed demand

Line: instantaneous import. Bars: completed billing intervals. Red: demand limit.

Fly through the recorded shift

24 seconds of Cycles footage from the checked closed-loop Blender project. The camera path is cinematography; every frame shows one recorded sample, from 10:00 to 13:00, with the time map in film.json.

Aerial Cycles render of the campus: assembly hall, warehouse with trucks, office tower, energy center with water tower and a solar carport.
Campus aerial · 10:30
Cutaway render of the assembly hall with the production line and twin status rings.
Hall cutaway · CNC 2 in service
Close render of CNC 2 with the twin's wear gauge: amber truth bar beside the cyan estimate and its uncertainty band.
CNC 2 wear gauge: truth vs. estimate
Render of the solar carport under a passing cloud with EV chargers glowing orange while capped at 50%.
Cloud over PV · chargers capped · 11:40

Every decision, with its evidence.

TimeTwinDecisionEvidence → predictionJump

Twelve paired shifts.

SeedGood parts
shadow → closed		CNC 2
failures		Intervals
over limit		Peak kW

Everything here is simulated. The plant model, its parameters and the twin's cost weights were chosen for this demonstration; they are not calibrated to a real factory. The comparison shows what closing this loop changes inside this model, not the value of a digital twin in general.

Offline lab ↓Blender + OpenUSD ↓

Verify offline: robot-reel factory-twin --output factory-twin --verify re-executes the plant and twin with the standard library, replays the twin from the recorded telemetry alone and checks every file. Fingerprints · Blender & USD readback · Closed trace · Shadow trace