glayguo/robot-reel-factory-twin

Robot Reel Factory Twin: close the loop

12 paired simulated shifts of a factory and campus, each run twice: once with the digital twin's commands applied (closed loop) and once with the identical twin only advising (shadow). Both modes of a pair share every disturbance and sensor-noise sample.

Over 12 pairs	Shadow twin	Closed loop
CNC 2 spindle failures	11	0
Billing intervals over the 520 kW limit	8	0
Pairs with more good parts		10 of 12 (2 fewer)

• Interactive lab · Hugging Face Space · Methods · Blender + OpenUSD projects

Configs

• pairs (default): one row per seed, closed-loop and shadow KPIs side by side.
• seeds: one row per seed and mode.
• samples: every 5-second sample of the featured seed (10) in both modes: plant truth (hidden spindle wear, buffers, power flows, hall temperature) beside the twin's estimates (wear ± σ, remaining life, demand forecast, demand level).
• decisions: every twin log entry for the featured seed with the evidence it had received and its model's prediction. actuated is True only in closed mode.

from datasets import load_dataset
pairs = load_dataset("glayguo/robot-reel-factory-twin", "pairs", split="test")
samples = load_dataset("glayguo/robot-reel-factory-twin", "samples", split="test")

Scope

This is a simulation. The plant model, its parameters and the twin's cost weights were chosen for the demonstration and are not calibrated to a real factory; nothing here is measured. It is evaluation output for inspecting a digital-twin loop, not training data or a benchmark. Every number re-executes bit for bit: pip install robot-reel then robot-reel factory-twin --output factory-twin --verify --all-seeds on the extracted offline lab.

source-manifest.json records the source commit and SHA-256 of every file; tables derive from lab.json and seeds.json with those hashes. Source commit: 7f02d86ec0ccb4a4126a0c07ceb6c3ed5fd30b94.