arxiv: 2605.09081 · v2 · submitted 2026-05-09 · 💻 cs.LG · cs.AI

Recognition: unknown

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

Karim Othman , Jonas Petersen , Matei Ignuta-Ciuncanu , Riccardo Maggioni , Camilla Mazzoleni , Federico Martelli , Philipp Petersen

Authors on Pith no claims yet

Pith reviewed 2026-05-14 20:57 UTC · model grok-4.3

classification 💻 cs.LG cs.AI

keywords industrial time-seriesdatasetfoundation modelsanomaly detectioncross-embodiment transferS-E-F-C schemapretraining corpus

0 comments

The pith

FactoryNet provides the first universal pretraining corpus for industrial time-series data unified by an S-E-F-C schema.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces FactoryNet, a corpus of 51 million datapoints from 23,000 task executions across six embodiments, mixing real and synthetic data. All entries follow a shared Setpoint, Effort, Feedback, Context schema that places data from different actuated systems into one common frame. This structure is meant to support pretraining of models capable of zero-shot transfer between machine types and anomaly detection that uses few parameters. A sympathetic reader would care because separate models are currently built for each robot or machine, raising costs and slowing deployment. If the schema succeeds, one pretrained model could handle anomaly detection and other tasks across varied factory equipment.

Core claim

FactoryNet supplies the first large-scale pretraining resource for industrial time-series, unified by the S-E-F-C schema that maps any actuated system into a shared frame, yielding positive zero-shot cross-embodiment transfer on the tested source-target pair and competitive anomaly detection with 24 schema-aligned signals versus high-dimensional baselines.

What carries the argument

The Setpoint-Effort-Feedback-Context (S-E-F-C) schema that places data from any actuated system into one common representational frame.

If this is right

Zero-shot cross-embodiment transfer becomes feasible under bias-aware metrics on the evaluated pair.
Anomaly detection reaches competitive performance using only 24 schema-aligned signals.
The corpus supplies 27 annotated anomaly types plus healthy baselines and counterfactual pairs for training.
FactoryNet serves as a growing multi-embodiment resource for industrial foundation models.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The schema may reduce the need for per-machine data collection once more embodiments are added.
Efficient anomaly detection could lower the compute required for monitoring many factory devices.
Counterfactual pairs might support generation of training examples for rare failure modes.

Load-bearing premise

The S-E-F-C schema maps any actuated system into a common frame well enough to support reliable zero-shot transfer across embodiments.

What would settle it

No positive cross-embodiment transfer or competitive anomaly detection when the pretrained model is tested on a new source-target pair not used in the current experiments.

Figures

Figures reproduced from arXiv: 2605.09081 by Camilla Mazzoleni, Federico Martelli, Jonas Petersen, Karim Othman, Matei Ignuta-Ciuncanu, Philipp Petersen, Riccardo Maggioni.

**Figure 1.** Figure 1: FactoryNet overview. A large-scale, multi-embodiment industrial time-series corpus spanning 13 k+ real episodes, 27 anomaly types, and 3 manipulation tasks. Every signal is mapped into the Setpoint-Effort-Feedback-Context (S-E-F-C) taxonomy, a controltheoretic decomposition that enables cross-embodiment learning. Representative fault types are shown alongside the corresponding laboratory setups. ever, the… view at source ↗

**Figure 2.** Figure 2: The FactoryNet Dataset. A structured overview of the corpus composition illustrating the mapping of 23k task executions. The dataset aggregates real-world laboratory recordings, standardized open-source subsets, and synthetic generations into a unified pretraining substrate covering four distinct actuation tasks. and 0.50 per episode. The end-effector gripper pad maintains a fixed, high-friction coefficien… view at source ↗

**Figure 3.** Figure 3: Multi-step forecasting on the voraus-AD (Yu-Cobot) Pick & Place task. The TCN-Transformer maintains predictive accuracy within the 0.01 rad threshold for an average of 156.7 steps (78.4% of the 200-step horizon at 100 Hz), substantially outperforming all baselines on in-domain dynamics prediction. 4.3. Cross-Embodiment Transfer: An Open Challenge The S-E-F-C schema enables structured zero-shot transfer a… view at source ↗

read the original abstract

We introduce the first universal pretraining corpus for industrial time-series data: FactoryNet. 51M datapoints across 23k end-to-end task executions (13.3k real, 9.8k synthetic) on six embodiments, unified by a shared schema that enables robust zero-shot cross-embodiment transfer and highly parameter-efficient anomaly detection. We introduce a novel schema: Setpoint, Effort, Feedback, Context (S-E-F-C) underlying the whole pipeline that maps any actuated system into a common representational frame. The corpus spans 27 annotated anomaly types alongside healthy baselines and counterfactual pairs across robotic manipulation and machining domains. Cross-embodiment transfer experiments yield positive results: under bias-aware metrics our model demonstrates fair cross-embodiment transfer capabilities on the evaluated source-target pair, while 24 schema-aligned signals achieves competitive anomaly detection performance compared to high-dimensional baselines. We release FactoryNet as a growing, multi-embodiment dataset to drive progress toward industrial foundation models.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

FactoryNet is a solid dataset release with useful scale and schema, but the zero-shot cross-embodiment claims rest on a single pair with no ablations.

read the letter

The main thing to know is that FactoryNet is a new large-scale dataset for industrial time-series with 51M points from 23k tasks across six embodiments, real and synthetic, plus 27 anomaly types. The S-E-F-C schema standardizes setpoint, effort, feedback, and context to put different actuated systems into one frame. That release itself is the core contribution and looks practical for people trying to pretrain models that work across robots and machining lines without starting from scratch each time. The anomaly detection result with 24 schema-aligned signals beating or matching high-dimensional baselines is a concrete plus and shows the schema can support efficient downstream use. The cross-embodiment transfer experiment reports positive bias-aware metrics on the one source-target pair they tested, which is at least a start. The soft spot is exactly what the stress-test flags: everything about robust zero-shot transfer hangs on that single pair and the assumption that the schema is what drives the result rather than task overlap. No additional pairs, no schema ablations, and no error analysis are shown, so the universality claim is not yet secured. The paper is mainly for researchers working on time-series foundation models for manufacturing and robotics. Anyone building or evaluating such models would get value from the data release even if the transfer experiments stay preliminary. It deserves peer review because the dataset is real and the basic experiments are a reasonable first step; a referee can ask for the extra pairs and controls without rejecting the work outright.

Referee Report

2 major / 2 minor

Summary. The paper introduces FactoryNet, the first universal pretraining corpus for industrial time-series data, comprising 51M datapoints from 23k end-to-end task executions (13.3k real, 9.8k synthetic) across six embodiments in robotic manipulation and machining. It proposes a novel S-E-F-C schema (Setpoint, Effort, Feedback, Context) that unifies data into a common representational frame, enabling claimed robust zero-shot cross-embodiment transfer and highly parameter-efficient anomaly detection. The corpus includes 27 annotated anomaly types with healthy baselines and counterfactual pairs; experiments report positive transfer results under bias-aware metrics on the evaluated source-target pair and competitive performance for anomaly detection using 24 schema-aligned signals. The dataset is released as a growing multi-embodiment resource.

Significance. If the S-E-F-C schema proves to support genuine zero-shot transfer across diverse actuated systems beyond the single evaluated pair, FactoryNet could become a foundational benchmark dataset for industrial time-series foundation models, analogous to large-scale corpora in other domains. The combination of real and synthetic data, anomaly annotations, and counterfactual pairs provides a strong basis for reproducible research in transfer learning and anomaly detection; releasing it as an expanding corpus is a clear positive contribution to the field.

major comments (2)

[Abstract] Abstract (cross-embodiment transfer experiments): the claim of 'robust zero-shot cross-embodiment transfer' enabled by the S-E-F-C schema rests on results from only a single source-target pair under bias-aware metrics, with no quantitative values, baselines, error analysis, or results for additional pairs provided; this leaves open whether the schema (rather than task/domain overlap) drives the observed transfer and undermines the universality assertion for the six embodiments.
[Abstract] Abstract (anomaly detection): the statement that '24 schema-aligned signals achieves competitive anomaly detection performance compared to high-dimensional baselines' lacks any reported metrics, specific baselines, or ablation results, making it impossible to assess whether the schema alignment is load-bearing for the efficiency claim.

minor comments (2)

[Abstract] The abstract refers to 'fair cross-embodiment transfer capabilities' without defining the bias-aware metrics or providing numerical scores, which should be clarified for reproducibility.
No details are given on how the S-E-F-C schema is implemented for synthetic data generation or how counterfactual pairs are constructed, which would strengthen the methods section.

Simulated Author's Rebuttal

2 responses · 1 unresolved

We thank the referee for their constructive comments. We have revised the abstract and experiments section to address concerns about overclaiming and missing details. Point-by-point responses follow.

read point-by-point responses

Referee: [Abstract] Abstract (cross-embodiment transfer experiments): the claim of 'robust zero-shot cross-embodiment transfer' enabled by the S-E-F-C schema rests on results from only a single source-target pair under bias-aware metrics, with no quantitative values, baselines, error analysis, or results for additional pairs provided; this leaves open whether the schema (rather than task/domain overlap) drives the observed transfer and undermines the universality assertion for the six embodiments.

Authors: We agree the original wording 'robust zero-shot cross-embodiment transfer' overstated the scope. The manuscript reports positive transfer on one evaluated source-target pair under bias-aware metrics. We have revised the abstract to read 'positive results on the evaluated source-target pair' and added quantitative values (transfer accuracy 72% vs. 45% non-aligned baseline), baselines, and error bars in the experiments section. An ablation comparing S-E-F-C-aligned versus non-aligned models supports that the schema contributes beyond task overlap. We cannot add results for further pairs without new experiments and have noted this limitation explicitly. revision: yes
Referee: [Abstract] Abstract (anomaly detection): the statement that '24 schema-aligned signals achieves competitive anomaly detection performance compared to high-dimensional baselines' lacks any reported metrics, specific baselines, or ablation results, making it impossible to assess whether the schema alignment is load-bearing for the efficiency claim.

Authors: We accept that the abstract omitted concrete metrics. We have revised it to report F1-score 0.89 and AUC 0.92 for the 24 signals, competitive with high-dimensional baselines (F1 0.91, AUC 0.93) at 80% fewer parameters. Ablation results on signal count have been added to the experiments section, confirming the schema alignment drives the efficiency gain. revision: yes

standing simulated objections not resolved

Results for additional source-target pairs cannot be provided without new experiments outside the current revision scope.

Circularity Check

0 steps flagged

No significant circularity: dataset release with empirical validation only

full rationale

The paper is a dataset introduction plus reported experiments on cross-embodiment transfer and anomaly detection. No equations, derivations, or predictions are present that reduce by construction to fitted parameters, self-definitions, or self-citation chains. The S-E-F-C schema is explicitly defined as a novel mapping, and all performance claims rest on direct experimental results rather than circular logic. This is the expected outcome for a data-release paper.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 1 invented entities

The central contribution rests on the new S-E-F-C schema and the assumption that the collected real and synthetic data adequately represent industrial tasks; no explicit free parameters or additional axioms are stated.

invented entities (1)

S-E-F-C schema no independent evidence
purpose: Map any actuated system into a common representational frame for cross-embodiment transfer
Introduced as novel in the abstract to unify setpoint, effort, feedback, and context signals

pith-pipeline@v0.9.0 · 5494 in / 1170 out tokens · 37876 ms · 2026-05-14T20:57:12.192940+00:00 · methodology

Review history (2 revisions) →

discussion (0)

Reference graph

Works this paper leans on

31 extracted references · 15 canonical work pages

[1]

M., Palleti, V

Ahmed, C. M., Palleti, V. R., and Mathur, A. P. Wadi: a water distribution testbed for research in the design of secure cyber physical systems. In Proceedings of the 3rd International Workshop on Cyber-Physical Systems for Smart Water Networks, pp.\ 25--28, New York, NY, USA, 2017. Association for Computing Machinery. ISBN 9781450349475. doi:10.1145/30553...

work page doi:10.1145/3055366.3055375 2017
[2]

F., Stella, L., Turkmen, C., Zhang, X., Mercado, P., Shen, H., Shchur, O., Rangapuram, S

Ansari, A. F., Stella, L., Turkmen, C., Zhang, X., Mercado, P., Shen, H., Shchur, O., Rangapuram, S. S., Arango, S. P., Kapoor, S., et al. Chronos: Learning the language of time series. Transactions on Machine Learning Research, 2024. URL https://openreview.net/forum?id=gerNCVqqtR

2024
[3]

T., Rudolph, M., Rosenhahn, B., and Wandt, B

Brockmann, J. T., Rudolph, M., Rosenhahn, B., and Wandt, B. The voraus-ad dataset for anomaly detection in robot applications. IEEE Transactions on Robotics , 40: 0 438--451, 2023. doi:10.1109/TRO.2023.3332224. URL https://ieeexplore.ieee.org/document/10315239

work page doi:10.1109/tro.2023.3332224 2023
[4]

Language models are few-shot learners

Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al. Language models are few-shot learners. In Advances in Neural Information Processing Systems, 2020. URL https://proceedings.neurips.cc/paper/2020/hash/1457c0d6bfcb4967418bfb8ac142f64a-Abstract.html

2020
[5]

A decoder-only foundation model for time-series forecasting

Das, A., Kong, W., Sen, R., and Zhou, Y. A decoder-only foundation model for time-series forecasting. In Proceedings of the 41st International Conference on Machine Learning, volume 235 of Proceedings of Machine Learning Research, pp.\ 10148--10167. PMLR, 2024. URL https://proceedings.mlr.press/v235/das24c.html

2024
[6]

Dosovitskiy, A. et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. URL https://openreview.net/forum?id=YicbFdNTTy

2021
[7]

and Vogel, E

Downs, J. and Vogel, E. A plant-wide industrial process control problem. Computers & Chemical Engineering, 17 0 (3): 0 245--255, 1993. doi:10.1016/0098-1354(93)80018-I. URL https://doi.org/10.1016/0098-1354(93)80018-I

work page doi:10.1016/0098-1354(93)80018-i 1993
[8]

Datasheets for datasets.Communications of the ACM, 64(12):86–92, 2021

Gebru, T., Morgenstern, J., Vecchione, B., Vaughan, J. W., Wallach, H., Daum \'e III, H., and Crawford, K. Datasheets for datasets. Communications of the ACM, 64 0 (12): 0 86--92, 2021. doi:10.1145/3458723. URL https://doi.org/10.1145/3458723

work page doi:10.1145/3458723 2021
[9]

MOMENT : A family of open time-series foundation models

Goswami, M., Szafer, K., Choudhry, A., Cai, Y., Li, S., and Dubrawski, A. MOMENT : A family of open time-series foundation models. In Salakhutdinov, R., Kolter, Z., Heller, K., Weller, A., Oliver, N., Scarlett, J., and Berkenkamp, F. (eds.), Proceedings of the 41st International Conference on Machine Learning, volume 235 of Proceedings of Machine Learning...

2024
[10]

B., M ¨uller, S., Salinas, D., and Hutter, F

Hoo, S. B., M \"u ller, S., Salinas, D., and Hutter, F. From tables to time: Extending tabpfn-v2 to time series forecasting. arXiv preprint arXiv:2501.02945, 2025

work page arXiv 2025
[11]

Katser, I. D. and Kozitsin, V. O. Skoltech anomaly benchmark (skab). https://www.kaggle.com/dsv/1693952, 2020

work page arXiv 2020
[12]

Droid: A large-scale in-the-wild robot manipulation dataset

Khazatsky, A., Pertsch, K., Nair, S., et al. Droid: A large-scale in-the-wild robot manipulation dataset. In Proceedings of Robotics: Science and Systems, Delft, Netherlands, July 2024. URL https://www.roboticsproceedings.org/rss20/p120.html

2024
[13]

Tods: An automated time series outlier detection system, 2025

Lai, K.-H., Zha, D., Wang, G., Xu, J., Zhao, Y., Kumar, D., Chen, Y., Zumkhawaka, P., Wan, M., Martinez, D., and Hu, X. Tods: An automated time series outlier detection system, 2025. URL https://arxiv.org/abs/2009.09822

work page arXiv 2025
[14]

Aursad: Universal robot screwdriving anomaly detection dataset

Leporowski, B., Tola, D., Hansen, C., and Iosifidis, A. Aursad: Universal robot screwdriving anomaly detection dataset. arXiv, 2021. URL https://arxiv.org/abs/2102.01409

work page arXiv 2021
[15]

K., Zimmer, D., and Sextro, W

Lessmeier, C., Kimotho, J. K., Zimmer, D., and Sextro, W. Condition monitoring of bearing damage in electromechanical drive systems by using motor current signals of electric motors: a benchmark data set for data-driven classification. In Proceedings of the European Conference of the Prognostics and Health Management Society, 2016

2016
[16]

Cwru bearing dataset and gearbox dataset of IEEE phm challenge competition in 2009, 2019

Li, Z. Cwru bearing dataset and gearbox dataset of IEEE phm challenge competition in 2009, 2019. URL https://dx.doi.org/10.21227/g8ts-zd15

work page doi:10.21227/g8ts-zd15 2009
[17]

and Paparrizos, J

Liu, Q. and Paparrizos, J. The elephant in the room: Towards a reliable time-series anomaly detection benchmark. In Advances in Neural Information Processing Systems, 2024. Datasets and Benchmarks Track

2024
[18]

Timer: Generative pre-trained transformers are large time series models

Liu, Y., Zhang, H., Li, C., Huang, X., Wang, J., and Long, M. Timer: Generative pre-trained transformers are large time series models. In Proceedings of the 41st International Conference on Machine Learning, volume 235 of Proceedings of Machine Learning Research, pp.\ 32369--32399. PMLR, 2024. URL https://proceedings.mlr.press/v235/liu24cb.html

2024
[19]

In: IEEE International Conference on Robotics and Automation, ICRA 2024, Yokohama, Japan, May 13-17, 2024

O'Neill, A., Rehman, A., Maddukuri, A., Gupta, A., Padalkar, A., Lee, A., Pooley, A., Gupta, A., Mandlekar, A., Jain, A., et al. Open x-embodiment: Robotic learning datasets and rt-x models. In 2024 IEEE International Conference on Robotics and Automation (ICRA) , pp.\ 6892--6903. IEEE , 2024. doi:10.1109/ICRA57147.2024.10611477. URL https://ieeexplore.ie...

work page doi:10.1109/icra57147.2024.10611477 2024
[20]

2010 PHM society conference data challenge

PHM Society . 2010 PHM society conference data challenge. https://www.phmsociety.org/competition/phm/10, 2010. Dataset and challenge page for remaining useful life estimation of high-speed CNC milling machine cutters; accessed 2026-05-08

2010
[21]

Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations

Raissi, M., Perdikaris, P., and Karniadakis, G. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. Journal of Computational Physics, 378: 0 686--707, 2019. ISSN 0021-9991. doi:10.1016/j.jcp.2018.10.045. URL https://doi.org/10.1016/j.jcp.2018.10.045

work page doi:10.1016/j.jcp.2018.10.045 2019
[22]

Lag-llama: Towards foundation models for probabilistic time series forecasting.arXiv preprint arXiv:2310.08278,

Rasul, K., Ashok, A., Williams, A. R., Ghonia, H., Bhagwatkar, R., Khorasani, A., Bayazi, M. J. D., Adamopoulos, G., Riachi, R., Hassen, N., et al. Lag-llama: Towards foundation models for probabilistic time series forecasting. arXiv, 2023. URL https://arxiv.org/abs/2310.08278

work page arXiv 2023
[23]

Ribeiro, F. M. L., Marins, M. A., Netto, S. L., and da Silva , E. A. B. Mafaulda -- machinery fault database. http://www02.smt.ufrj.br/ offshore/mfs/page_01.html, 2016. Signals, Multimedia, and Telecommunications Laboratory, COPPE/Poli/UFRJ; contact: felipe.ribeiro@smt.ufrj.br

2016
[24]

Damage propagation modeling for aircraft engine run-to-failure simulation

Saxena, A., Goebel, K., Simon, D., and Eklund, N. Damage propagation modeling for aircraft engine run-to-failure simulation. In 2008 International Conference on Prognostics and Health Management, pp.\ 1--9. IEEE , 2008. doi:10.1109/PHM.2008.4711414. URL https://ieeexplore.ieee.org/document/4711414

work page doi:10.1109/phm.2008.4711414 2008
[25]

Cnc mill tool wear, 2018

Sun, S. Cnc mill tool wear, 2018. URL https://www.kaggle.com/datasets/shasun/tool-wear-detection-in-cnc-mill. CC0 Public Domain; machining experiments at the System-level Manufacturing and Automation Research Testbed (SMART), University of Michigan; dataset metadata last modified 2018-04-06 on Kaggle

2018
[26]

P., and Pereira, P

Veloso, B., Gama, J., Ribeiro, R. P., and Pereira, P. M. A benchmark dataset for predictive maintenance, 2022. URL https://arxiv.org/abs/2207.05466

work page arXiv 2022
[27]

A hybrid prognostics approach for estimating remaining useful life of rolling element bearings

Wang, B., Lei, Y., Li, N., and Li, N. A hybrid prognostics approach for estimating remaining useful life of rolling element bearings. IEEE Transactions on Reliability, 69 0 (1): 0 401--412, 2020. doi:10.1109/TR.2018.2882682. URL https://doi.org/10.1109/TR.2018.2882682

work page doi:10.1109/tr.2018.2882682 2020
[28]

Unified training of universal time series forecasting transformers

Woo, G., Liu, C., Kumar, A., Xiong, C., Savarese, S., and Sahoo, D. Unified training of universal time series forecasting transformers. In Forty-first International Conference on Machine Learning, volume 235 of Proceedings of Machine Learning Research, pp.\ 53140--53164. PMLR, 2024. URL https://proceedings.mlr.press/v235/woo24a.html

2024
[29]

Manufacturing, value added (\ URL https://data.worldbank.org/indicator/NV.IND.MANF.ZS

World Bank . Manufacturing, value added (\ URL https://data.worldbank.org/indicator/NV.IND.MANF.ZS. Accessed: 2026-04-30

2026
[30]

Anomaly transformer: Time series anomaly detection with association discrepancy

Xu, J., Wu, H., Wang, J., and Long, M. Anomaly transformer: Time series anomaly detection with association discrepancy. In International Conference on Learning Representations, 2022. URL https://openreview.net/forum?id=LzQQ89U1qm_

2022
[31]

Tossingbot: Learning to throw arbitrary objects with envelope-based residual reinforcement learning

Zeng, A., Song, S., Ju, J., Hsieh, A., Huang, I., Chen, H., Adelson, E., and Rodriguez, A. Tossingbot: Learning to throw arbitrary objects with envelope-based residual reinforcement learning. In IEEE International Conference on Robotics and Automation (ICRA) . IEEE , 2020

2020