FactoryNet
Summary
FactoryNet is a paper-reported 51M-datapoint, 23k-episode industrial multivariate time-series corpus organized around the Setpoint–Effort–Feedback–Context (S-E-F-C) schema. Its main contribution is an explicit control-loop data contract across multiple robot and machine embodiments.
Dataset Contract
- Setpoint: commanded intent / control-input-like target.
- Effort: applied actuation quantity.
- Feedback: measured machine response observation.
- Context: environment, task, mode, payload, and metadata.
- Paper-reported scope: six embodiments, real/open/synthetic tracks, 27 injected anomaly types.
- Primary tasks: dynamics pretraining, anomaly detection, predictive maintenance, and cross-embodiment transfer.
Official Artifacts
- Preprint: 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: https://huggingface.co/datasets/Forgis/FactoryNet
- Raw dataset metadata: source.md
- Raw paper Markdown: paper_factorynet-2026.md
Role In The Wiki
FactoryNet is the strongest current public industrial-data candidate for pretraining a non-vision action/control-input-conditioned foundation time-series model. It complements passive forecasting corpora by recording commanded intent separately from realized response.
Its natural local experiment is to pretrain a shared latent dynamics encoder across embodiments, specialize it with HEPA-style horizon-conditioned event prediction, and evaluate state/intervention/counterfactual/decision competence on FactoryBench.
Relation To Foundation TSFM Agenda
Use the source-level agenda mapping in FactoryNet rather than duplicating verdict rows here.
At the entity level, FactoryNet partially supplies the missing action/control-input data interface, but it is not a complete offline-RL or closed-loop control benchmark because it lacks standardized rewards, policy actions, candidate-action evaluation, and outcome-based decision scoring.
Caveats
- The visible Hugging Face snapshot does not transparently match the full 51M-datapoint, six-embodiment paper corpus.
- Cross-embodiment transfer is demonstrated on one source-target pair and primarily under mean-centered error.
- Synthetic data is restricted to pick-and-place.
- License statements differ across the paper, code repository, and adapted data subsets.