# FactoryNet Dataset Metadata Snapshot

Fetched: 2026-07-21

## Canonical Sources

- Paper: https://arxiv.org/abs/2605.09081v4
- Official project page: https://www.forgis.com/papers/factorynet.html
- Official code: https://github.com/Forgis-Labs/FactoryNet
- Official Hugging Face dataset: https://huggingface.co/datasets/Forgis/FactoryNet
- FMSD @ ICML 2026 accepted-paper listing: https://icml-structured-fm-workshop.github.io/accepted-papers/
- AI4Physics @ ICML 2026 workshop: https://ai4physics-workshop.github.io/

## Dataset Type

FactoryNet is a multi-embodiment industrial multivariate time-series corpus for dynamics pretraining, anomaly detection, predictive maintenance, and cross-embodiment transfer. The paper maps heterogeneous machine signals into a Setpoint–Effort–Feedback–Context (S-E-F-C) schema.

In the wiki's canonical terminology:

- **Setpoint** is commanded intent and is usually a control-input-like channel.
- **Effort** records applied actuation quantities such as current or torque.
- **Feedback** records measured machine response and is an observation channel.
- **Context** records task, payload, mode, environment, and other conditioning variables; anomaly labels are supervision, not control inputs.

## Paper-Reported Corpus

The arXiv v4 paper reports:

- 51 million datapoints;
- more than 23,000 task-execution episodes: 13.3k real and 9.8k synthetic;
- six embodiments: UR3, KUKA KR10, Yu-Cobot, AURSAD UR3e, UMich CNC, and simulated UR5;
- three manipulation tasks plus CNC machining;
- 27 injected anomaly types in the laboratory track;
- more than 300 heterogeneous source columns mapped into S-E-F-C roles.

The paper's composition table reports 7,141 UR3, 1,973 KUKA KR10, 2,122 voraus-AD, 2,045 AURSAD, 18 UMich CNC, and 9,799 Isaac Sim episodes.

## Current Public Artifact Snapshot

At fetch time, the official Hugging Face repository exposed:

- repository revision `386d731fab807f53ac17a69cf209a2b42b491072`;
- 13.7 GB total repository size;
- `raw` and `normalized` Parquet configurations;
- 40,592,518 rows reported by the Dataset Viewer for the visible `raw/train` configuration;
- a README that describes only the three adapted open-source sources: UMich CNC, AURSAD, and voraus-AD.

This public snapshot does not transparently reconcile with the paper's 51M-datapoint, six-embodiment corpus. Users SHOULD pin the Hugging Face revision and verify which embodiments, tasks, and configurations are actually present before treating a run as a reproduction of the paper's full corpus.

## Reported Validation

- On voraus-AD, a 24-signal S-E-F-C MLP reaches 83.2% mean AUROC. This beats 1-NN, GANF, and PCA in the paper's table, but remains below CAE, LSTM-VAE, and MVT-Flow.
- A 105k-parameter TCN-Transformer trained on voraus-AD reaches mean-centered MAE 0.339 ± 0.006 on 1,433 AURSAD episodes, versus 0.373 ± 0.005 for the kinematic baseline.
- Raw effort MAE remains worse than the zero-predictor baseline because static payload and gravity-compensation biases do not transfer; the positive result is specifically about mean-centered dynamic shape.

## Actions, Events, And Counterfactuals

FactoryNet is more useful for action-conditioned dynamics than passive anomaly datasets because Setpoint explicitly records commanded intent and Feedback records the realized response. It still lacks rewards, a standardized policy/action space, and a closed-loop candidate-control evaluation protocol. Fault injections and anomaly labels are events or experimental conditions, not automatically actions chosen by an agent.

The paper mentions paired healthy/fault or synthetic twin episodes, but these SHOULD NOT be treated as exact counterfactuals without checking the matching protocol and pre-intervention state.

## Main Limitations

- Synthetic data is restricted to Pick & Place.
- Cross-embodiment transfer is evaluated on one source-target pair.
- The public Hugging Face snapshot appears narrower or stale relative to the paper-reported full corpus.
- Licensing is not uniform across the corpus and differs across official artifacts; see `license.md`.