Alex Open Research Wiki

FactoryNet: A Large-Scale Dataset toward Industrial Time-Series Foundation Models

6 min read

  • Important
  • Skimmed

FactoryNet: A Large-Scale Dataset toward Industrial Time-Series Foundation Models

Source

Status And Credibility

FactoryNet is an arXiv preprint first submitted on 2026-05-09 and revised to v4 on 2026-06-03. The arXiv record states that it was accepted at the FMSD and AI4Physics workshops at ICML 2026. The official FMSD accepted-paper page independently lists it as a poster. This is workshop acceptance, not ICML main-conference acceptance.

The paper is current and highly relevant to the wiki because it exposes real commanded-intent and measured-response channels in an industrial multivariate time-series corpus. Public code and a Hugging Face dataset exist. Credibility is limited by one source-target transfer experiment, a release snapshot that does not transparently match the full paper corpus, and no independent reproduction.

Core Claim

FactoryNet argues that industrial time-series transfer requires more than data volume. A reusable corpus should separate what the controller requested from what the machine did. Its Setpoint–Effort–Feedback–Context (S-E-F-C) schema maps heterogeneous machinery into four roles:

RoleCanonical interpretationExamples
Setpointcommanded intent / control-input-like targettarget joint position, velocity, acceleration, torque
Effortapplied actuation quantitymotor current, torque, electrical power
Feedbackmeasured observation / realized responseactual position, velocity, tool-center-point force
Contextconditioning variables and metadatapayload, temperature, mode, task phase, labels
flowchart LR
  C[context and boundary conditions] --> D[dynamics]
  S[setpoint / commanded intent] --> D
  D --> E[effort / actuation]
  D --> F[feedback / measured response]
  E --> R[structured residuals and transfer]
  F --> R

This schema is valuable because it makes the control loop explicit rather than flattening all channels into anonymous sensor values.

Corpus Composition

The paper reports 51M datapoints and more than 23k task executions:

SourceEmbodimentEpisodesApprox. datapoints
Forgis LabUR37,14118M
Forgis LabKUKA KR101,9734M
voraus-ADYu-Cobot2,12216M
AURSADUR3e2,0453M
UMich CNCCNC gantry1818K
Isaac SimUR59,79910M

The real laboratory track covers pick-and-place, screwdriving, and peg-in-hole with 27 injected fault types. The synthetic track uses Isaac Sim domain randomization over payload mass, friction, controller gains, sensor noise, and task geometry.

Evidence And Results

S-E-F-C anomaly detection

A 24-signal MLP predicts six effort channels from 18 setpoint channels on voraus-AD and uses per-episode effort MAE as the anomaly score. It reaches 83.2% mean AUROC. This is a useful parameter-efficiency result, not a new anomaly-detection SOTA: it beats 1-NN, GANF, and PCA in the reported table but remains below CAE, LSTM-VAE, and MVT-Flow.

Zero-shot cross-embodiment transfer

A 105k-parameter TCN-Transformer trained only on voraus-AD Yu-Cobot pick-and-place is evaluated on 1,433 AURSAD UR3e screwdriving episodes. Under mean-centered MAE, it reaches 0.339 ± 0.006, versus 0.373 ± 0.005 for a kinematic baseline and worse values for unstructured learned baselines.

The metric boundary is critical: raw effort MAE remains high because static payload and gravity-compensation biases differ between embodiments. The positive result concerns the shape of dynamic effort after per-episode mean removal, not absolute force calibration.

Sim-to-real

The paper reports paired sim-to-real gap analysis under aligned S-E-F-C fields. This supplies a useful forward-dynamics audit, but the synthetic track is limited to pick-and-place and does not prove broad sim-to-real policy or control transfer.

Public Release Audit

The official Hugging Face repository currently reports 13.7 GB and 40,592,518 visible rows for raw/train. Its README describes only UMich CNC, AURSAD, and voraus-AD, whereas the paper reports 51M datapoints across six embodiments including Forgis lab data and Isaac Sim.

The public snapshot therefore does not transparently reconcile with the paper’s full corpus. Reproduction work MUST pin the Hugging Face revision and enumerate present embodiments, tasks, episode IDs, and configs before claiming to use the paper’s full FactoryNet v1.0.

There is also a licensing mismatch: the paper says novel lab/synthetic data are MIT and adapted subsets retain upstream licenses, while the official GitHub repository states CC BY-NC-SA 4.0. The raw dataset license note records this boundary.

Actions, Events, And Counterfactuals

FactoryNet materially improves on passive telemetry datasets because Setpoint records commanded intent and Feedback records the realized response. This makes it a plausible substrate for action-conditioned short-horizon dynamics.

It is not yet a complete action-conditioned world-model benchmark:

  • setpoints are logged controller targets, not necessarily high-level actions chosen by a learned policy;
  • faults are events or experimental conditions, not automatically actions;
  • no reward, task-cost, candidate-action ranking, or closed-loop policy protocol is standardized;
  • matched healthy/fault or synthetic twins are useful but should not be treated as exact counterfactuals without state-matching audits.

Limitations

  • Cross-embodiment transfer is tested on one source-target pair.
  • Synthetic data is restricted to pick-and-place.
  • Absolute effort transfer remains poor; the positive transfer result uses mean-centered dynamics.
  • The current public Hugging Face snapshot does not transparently match the paper’s six-embodiment scale.
  • The schema standardizes roles but not necessarily units, missing channels, sampling rates, controller semantics, or action granularity across machines.
  • Faults are mostly injected and may not represent gradual, compound production failures.
  • License terms need per-subset resolution.

Foundation TSFM Relevance

Agenda slotVerdictEvidenceMissing pieces
Context and control-input interfacepartially closesS-E-F-C explicitly separates commanded intent, actuation effort, measured feedback, and context.Need typed action semantics, units, failed-action receipts, rewards/costs, and intervention timing.
Native multivariate encodingpartially closesMore than 300 heterogeneous source columns are mapped into machine-agnostic roles across six paper-reported embodiments.Need public full-corpus verification, missing-channel masks, topology/kinematic context, and higher-channel model evidence.
Data diversity and transferpartially closesMixes real lab, adapted open, CNC, and simulated robotic telemetry; one zero-shot source-target transfer is positive under MC-MAE.One pair and one bias-aware metric are too narrow for general cross-embodiment transfer.
Control and counterfactualspartially closesLogged setpoints and realized feedback support action/control-input-conditioned transition learning more directly than passive anomaly corpora.No reward, candidate-action rollout, exact counterfactual protocol, or closed-loop decision benchmark.
Benchmark levelwarningMean-centering isolates dynamic shape and exposes an absolute-calibration failure that aggregate transfer claims could hide.Report raw and centered metrics, per-phase errors, calibration, downstream action ranking, and safety constraints.
Public reproducibilitywarningCode and Hugging Face artifacts are public.Resolve paper-vs-release scale and license mismatches; pin revisions and inventory embodiments.

Why It Matters For This Wiki

FactoryNet is probably the closest public dataset in the current corpus to the desired non-vision foundation time-series model interface. Unlike ordinary forecasting data, it carries a control-loop decomposition. Unlike generic offline RL, it carries dense industrial telemetry and machine context. Unlike passive anomaly benchmarks, it distinguishes commanded intent from realized response.

Its strongest role is as the data substrate beneath HEPA and FactoryBench: pretrain a latent dynamics model on S-E-F-C trajectories, specialize it for horizon-indexed events, and test whether the resulting state supports intervention and decision questions. The papers do not yet demonstrate this integrated stack.

Open Questions

  • Can one HEPA/LeNEPA-style encoder be pretrained across all FactoryNet embodiments without hiding machine identity in context leakage?
  • Which S-E-F-C fields are true control inputs, which are low-level controller traces, and which are consequences rather than causes?
  • Does transfer survive without per-episode mean-centering and with absolute-unit calibration?
  • How should topology, kinematics, units, axis count, sampling rate, and missing channels enter the context interface?
  • Can FactoryNet support calibrated candidate-control rollouts, not only next-effort prediction and anomaly scoring?
  • When will the public Hugging Face release expose and document the full six-embodiment paper corpus under unambiguous per-subset licenses?

Graph View

Backlinks