Source-linked AI summary
World Models for Cross-Machine CNC Transfer under Partial Sensor Overlap
Ayoub Louaye Bouaziz, Matthieu Ostertag, Anton Demasles
TL;DR
Industrial world models must transfer across machines with different dynamics and sensor schemas, but evidence for command-conditioned latent models in that setting is limited. This study tests a schema-adaptive JEPA on an unseen CNC machine using source-only selection, finding no source RMSE gain over scratch but zero-shot transfer above persistence and below official RevIN forecasters. A post-lock ablation finds instance normalization closes that gap while harming target calibration.
Problem
It is unclear whether representations trained to predict future embeddings remain useful when a world model moves to a machine with changed dynamics and a partially different sensor schema.
Source: §1 · p. 1Method
The study trains a schema-adaptive, command-conditioned JEPA and evaluates its forecasting transfer from a 17-channel source CNC machine to a target sharing 10 channels.
Source: §1 · p. 2Results
The locked model transfers above persistence but trails official RevIN forecasters; adding input normalization closes the zero-shot RMSE gap but yields target NLL 20.6.
Source: §1 · p. 2, §9 · p. 14Takeaways & Limitations
Cross-machine transfer under partial sensor overlap is a distinct evaluation axis, and instance normalization can explain the zero-shot RMSE margin while costing calibration.
Source: §1 · p. 2, §9 · p. 14Evidence is limited to one source and one target machine, with seven independent target runs; downstream policy learning and off-policy control are not evaluated.
Source: §8 · pp. 13–14
Abstract
from arXiv · showhide
Industrial world models must move between machines whose dynamics, sensing interfaces and command conventions differ. This study asks whether a command-conditioned latent world model, trained to predict future representations of the process rather than to reconstruct future samples, keeps its value on a machine it has never seen: a source CNC machine exposes 17 sensor channels, the target sharing 10 of those. All model selection uses source data only, and the locked configuration is evaluated on the target once. Two findings follow. First, latent-predictive pretraining brings no in-domain forecasting gain over matched training from scratch, so source accuracy alone cannot show what such a representation is worth. Second, the transferred model beats persistence on the unseen machine (with $R^2\approx0.01$ against the target mean) but trails official forecasters that normalize each input window by its own statistics; a post-lock ablation, declared before it ran, shows that this input normalization alone closes the gap, and closing it costs predictive calibration. Cross-machine transfer under partial sensor overlap is therefore a distinct evaluation axis for command-conditioned world models.
1. Introduction
Industrial world models must transfer across machines with both changed dynamics and partially overlapping sensor schemas. This study tests whether command-conditioned latent prediction remains useful on an unseen CNC machine under source-only model selection.
1. Introduction
- Domain shift changes sensor signals while schema shift changes which channels exist, and target CNC machines typically exhibit both.The source has 17 channels and the target shares 10; selection uses source data before a single locked target evaluation.Source: §1 · pp. 1–2
- The study tests whether JEPA representations that predict future embeddings rather than reconstruct samples remain useful when both dynamics and sensor schema change.The paper notes that prior transfer settings discussed here keep physics and observation space unchanged, unlike this cross-machine test.Source: §1 · p. 1
- The contribution combines an audited cross-machine protocol with a schema-adaptive, command-conditioned model, while restricting its scope to forecasting rather than policy learning or closed-loop control.The model uses one canonical channel list and a presence mask; the study evaluates multi-horizon forecasts under given command sequences.Source: §1 · p. 2
- The locked model transfers above persistence but trails official RevIN forecasters; a post-lock ablation shows input normalization closes the zero-shot RMSE gap at the cost of target calibration.The normalized variant reports RMSE 0.495 ± 0.004 against 0.503 and 0.498, with target NLL 20.6.Source: §1 · p. 2
2. Related Work
Related work spans predictive representation learning, latent world models, time-series forecasting, and industrial adaptation. This paper distinguishes its setting through cross-machine transfer with partial sensor overlap and explicit command conditioning.
Joint-Embedding Predictive Learning
JEPA world models
- Prior latent world models generalize across maze layouts with unchanged physics and observation space [4], whereas this CNC setting changes dynamics and sensor schema.EB-JEPA reports that removing inverse-dynamics loss drops Two Rooms planning success from 97% to 1% [5]; LeWorldModel uses a Gaussian-embedding regularizer without inverse dynamics, EMA, or stop-gradient [6].Source: §2 · p. 3
Time-series forecasting and masked pretraining
- Forecasting baselines include masked-series reconstruction, channel-independent patch tokens, inverted multivariate tokenization, and instance normalization.These methods motivate testing transfer and schema mismatch rather than assuming JEPA will dominate source-domain forecasting.Source: §2 · p. 3
World models and industrial adaptation
3. Problem Formulation
The formulation treats CNC forecasting as predicting future sensor values from recent observations and commands, while explicitly representing which channels are present. The model must distinguish an absent channel from a measured zero.
3. Problem Formulation
- The source has 17 canonical channels, while the target shares 10; observations use the source channel order and a presence mask marks measurements actually available.Target-only channels are discarded, and the seven source channels absent on the target have mask value zero.Source: §3 · p. 3
- Commands are spindle-speed and feed setpoints, with each command acting over the interval ending at its logged step.Thus the command at t+1 is the first one acting on future rows.Source: §3 · p. 3
- The forecasting objective is the conditional distribution of future observations given the recent K rows and commands through the longest horizon.The formulation identifies this conditional distribution as the object of interest, rather than a particular training loss.Source: §3 · p. 3
- A missing channel is not treated as a zero reading: absent-channel tokens are dropped from attention and pooling, allowing the same weights to operate on both machines.A measured zero retains a present mask, while an absent channel has mask value zero.Source: §3 · p. 3
- Low numerical error alone is insufficient because highly autocorrelated streams can let a forecaster ignore future commands.The model requirement also checks whether forecasts respond meaningfully to changes in those commands.Source: §3 · p. 4
4. Method
The method encodes available sensors and past commands into a schema-adaptive context, predicts command-conditioned future latents, and decodes them into probabilistic process forecasts. It adds objectives for reduced-schema consistency and command recovery, then evaluates transfer and an instance-normalized variant.
4.1. Schema-adaptive context encoding
- The schema-adaptive encoder attends over present-channel tokens at each time step, pools them, adds the step command, then uses a temporal transformer to produce the context latent.Because the channel pass operates on sets of varying size, the same weights can process either machine’s available sensors.Source: §4.1 · p. 4
4.2. Command-conditioned latent prediction
- The predictor fills fixed horizon slots in one pass from the context latent, conditioning each on its own mean future command rather than rolling predicted latents forward.The horizons are 1, 2, 4, 8, and 16 steps; the target encoder embeds future rows jointly and is updated by EMA, with gradients stopped through that branch.Source: §4.2 · p. 4
- Training combines latent prediction with variance–covariance regularization, one-sided schema consistency, and a lightly weighted command-recovery head.The reduced view is pulled toward the detached fuller-view prediction; the command-recovery head regresses mean future commands from context and target latents. A fresh probabilistic head is attached for supervised physical forecasting, with scratch and transfer controls retained.Source: §4.2 · pp. 4–6, §4.3 · p. 6
4.3. Physical forecasting head and adaptation
- The physical head maps predicted latents to diagonal Gaussian forecasts, whose product treats horizons and channels as conditionally independent and uses future-command means.Consequently, the model represents per-horizon marginals, not their correlations; pretraining is only a surrogate for the physical forecasting objective.Source: §4.3 · p. 6
4.4. Instance-normalized variant
- Few-shot adaptation updates the predictor and physical head on leading target-run support windows, while online adaptation is not evaluated.The instance-normalized variant computes context-window statistics over present channels only and applies the inverse transform to forecasts; the paired ablation changes this transform alone.Source: §4.3 · p. 6, §4.4 · pp. 6–7
5. Experimental Protocol
The protocol pairs THWS source data with JOANNEUM target data under partial channel overlap, with model selection restricted to source validation before the target evaluation. It defines a stability-checked lock, target baselines and an evaluation metric set.
5.1. Data and audits
- THWS contributes 17 channels across 62 sessions; JOANNEUM shares 10 channels across seven runs, and source data are split by session with source-training-only normalization.Source windows do not cross group boundaries.Source: §5.1 · p. 7
5.2. Model selection and baselines
- Twenty candidates were ranked on a six-component source-validation score; after the four-seed leader changed, the schema-consistency/command-recovery model prevailed after seven valid seeds and was locked before target evaluation.The source-only ranking components include clean, schema-drop, masked and long-horizon RMSE, calibration penalty, and held-out SSL validation loss.Source: Figure 4 · p. 8, §5.2 · p. 8
- Official PatchTST and iTransformer use RevIN; the separately declared paired ablation changed only that flag, and target evaluation included six runs without target-based selection.The control reused the locked configuration, while the variant was evaluated on the target only after winning all four RMSE components on source validation.Source: §5.2 · p. 9
5.3. Metrics
- Forecasting uses RMSE, MAE, R2, and per-horizon/per-channel errors; probabilistic evaluation uses Gaussian NLL and coverage, with seven target runs as independent units.R2 is computed against the pooled mean of the evaluated channels and horizons.Source: §5.3 · p. 9
6. Results
The locked model transfers to the unseen CNC machine and beats persistence, but trails official zero-shot forecasters; adding instance normalization closes the RMSE gap at a calibration cost. Pretraining does not improve clean-source RMSE over matched scratch training, and the few-shot findings remain diagnostic.
6.1. Corrected JEPA pretraining is stable but does not improve clean-source RMSE
- Corrected pretraining raised target-encoder effective rank from approximately 5% to 58% of latent dimensionality, but scratch and pretrained RMSEs were 0.811 ± 0.022 and 0.813 ± 0.022 (p = 0.76).The corrected pipeline selects checkpoints on held-out SSL validation and applies VICReg to pooled context representations.Source: §6.1 · p. 9
6.2. Representation hyperparameters materially affect stability and forecasting
- The best tested latent-loss weight was λlat = 0.25 with RMSE 0.774; removing the latent loss yielded 0.787, while the lightest mask ratio had the best mask-ratio RMSE, 0.804.The locked candidate retained λlat = 1 and whole-channel masking because selection also weighted schema-drop and masked robustness, calibration, and command usage.Source: §6.2 · pp. 9–10
- Per-horizon VICReg improved clean-source RMSE from 0.815 to 0.798 and was required to prevent collapse; pooled horizon aggregation reached 0.787 versus 0.829 for per-horizon aggregation.The slower EMA setting, 0.9995, reached 0.799 versus 0.808 at 0.999 and approximately 0.83 for faster updates.Source: §6.2 · p. 10
6.3. Source-only model selection identifies a schema-consistent command-recovery variant as the locked configuration
- The selected schema-consistency and command-recovery variant scored 0.8216 ± 0.0089 source-validation RMSE, with schema-drop score 0.918 and command-shuffling sensitivity 1.058.The per-horizon VICReg variant led at four-seed discovery, but the selected configuration overtook it after five and then seven seeds.Source: §6.3 · p. 10
6.4. Strong forecasting baselines remain competitive on the source machine
- RSSM achieved source-test RMSE 0.759 and MLP 0.771, outperforming the pre-search JEPA model at 0.809 and the locked candidate at 0.822.All learned models outperformed linear drift and persistence.Source: §6.4 · p. 11, Figure 6 · p. 12
6.5. The locked model transfers zero-shot, but official zero-shot forecasters remain stronger in RMSE
- Target RMSE was 0.546 (R2 = 0.012), beating persistence (0.654) but trailing RevIN PatchTST (0.503) and iTransformer (0.498).The locked model was evaluated once on the target after source-only selection.Source: §6.5 · p. 11, Table 3 · p. 12
6.6. Instance normalization closes the zero-shot RMSE gap but breaks target calibration
- Target RMSE fell from 0.555 ± 0.015 to 0.495 ± 0.004 with RevIN, while target NLL rose from 0.89 to 20.6 and 90% coverage fell from 0.953 to 0.874.The paired post-lock ablation used three seeds; its results were treated as consistent wins, not significance claims.Source: Table 3 · p. 12, §6.6 · p. 12
6.7. Few-shot adaptation remains promising but is still diagnostic
- The pre-lock adaptation sweep reached RMSE 0.520 and R2 = 0.167 with 20% target support, but these results remain diagnostic rather than confirmatory.The 5% support setting lacked a valid 48-step window and therefore reduced to zero-shot transfer.Source: §6.7 · p. 13
7. Discussion
The discussion distinguishes cross-machine transfer from in-domain forecasting: JEPA pretraining does not improve clean-source forecasting over matched scratch training, while the locked model transfers above persistence but trails RevIN-equipped baselines. A post-lock ablation links the RMSE gap to input normalization, which closes it at a calibration cost.
7. Discussion
- The locked model transfers zero-shot above persistence, but official RevIN-equipped forecasters achieve stronger zero-shot RMSE; a three-seed ablation shows RevIN closes that margin while costing calibration.The same architecture with RevIN matches the baselines but loses calibration on stationary windows.Source: §7 · p. 13
- JEPA pretraining does not improve clean-source forecasting over matched scratch training, so in-domain accuracy alone does not establish the value of its representation for transfer.The discussion also reports that the plain JEPA control collapses, large scaling is unstable, and the provisional winner changes during confirmation.Source: §7 · p. 13
- Industrial predictive representations should be judged for command response, uncertainty awareness, schema-shift robustness, and adaptability under limited target data, not source interpolation alone.The study does not learn a policy, plan, or close a control loop.Source: §7 · p. 13
- Future work should measure invariance across multiple source machines, make instance normalization compatible with the probabilistic head, and evaluate decision quality and policies offline.Causal representation learning remains outside the claim set until the data provide enough intervention diversity.Source: §7 · p. 13
8. Limitations
The evidence comes from a narrow cross-machine setting and leaves important transfer properties unresolved. The study also does not yet evaluate downstream policy learning or off-policy control.
8. Limitations
- One source and one target machine limit claims of general cross-machine generalization, while seven independent JOANNEUM runs restrict statistical power.The target stream assumes a 2 ms controller cycle; resampling sensitivity remains a separate analysis, and seven source-only channels lack target counterparts.Source: §8 · p. 13
- The locked target result is a single checkpoint evaluated once, without run-level confidence intervals or command-sensitivity tests on the target machine.On the pre-lock model, zeroing or shuffling future commands changed target RMSE by less than 0.005, leaving command use under transfer unverified.Source: §8 · p. 13
- The pre-lock few-shot curve is not confirmatory, and transfer results do not isolate the latent objective because no target arm without it was evaluated.The channel-masking ratio was not varied; Table 2b ablations are single-seed, and official baselines are single-seed runs without hyperparameter search.Source: §8 · pp. 13–14
- The paper does not yet evaluate downstream policy learning or off-policy control, so its world-model evidence remains predictive rather than decision-theoretic.Source: §8 · p. 14
9. Conclusion
The study presents a schema-adaptive, command-conditioned JEPA and an audited CNC protocol for cross-machine transfer with partial sensor overlap. Its results support a controlled benchmark and methodology, not a state-of-the-art claim; code, audit materials, and datasets are made available.
9. Conclusion
- The schema-adaptive, command-conditioned JEPA transfers zero-shot above persistence, but RevIN-equipped forecasters remain stronger in raw zero-shot RMSE; input normalization closes that margin at a target-calibration cost.JEPA pretraining does not improve clean-source RMSE over matched scratch training.Source: §9 · p. 14
Code and data availability
- Code, configuration files, candidate specifications, audit scripts, schematics, and both public datasets are linked in the availability section.Source: Code and data availability · p. 14
A. Architecture Search: Twenty Candidates
The 20-candidate architecture search compared design and training variants using THWS source validation only, excluding JOANNEUM target data from model selection while assessing variability and stability. A final stability check locked the schema-consistency variant with command recovery for target evaluation.
A. Architecture Search: Twenty Candidates
- Plain JEPA collapsed in all 4 seeds, while JEPA with per-horizon VICReg was valid in all 7 seeds and ranked #1 in confirmation (0.093).The VICReg candidate ranked #1 in discovery (-0.346); the schema-adaptive latent-consistency candidate was valid in 4/4 seeds and ranked #11 in discovery (-0.272).Source: A. Architecture Search: Twenty… · p. 15
- The schema-and-command-recovery variant was locked after seven valid seeds, ranking #5 in discovery (-0.296) and #1 in confirmation (-0.830).Candidate ranking used THWS source validation only; JOANNEUM target data were excluded, and code, configuration, normalizers, and checkpoint were frozen before the confirmatory target pass.Source: A. Architecture Search: Twenty… · p. 15
Robust source normalization IQR PATHOLOGY PATHOLOGICAL ROBUST NORMALIZATION
- Reported Clean scores were 0.9813 ± 0.0360, 0.8485 ± 0.0006, and 0.8148 ± 0.0016; corresponding selections were Discovery #17 (6.25), #16 (0.462), and #15 (-0.084).Source: Robust source normalization IQR… · p. 17
- Discovery #19 scored -0.128 and had Clean 0.7937 (1 seed), while Discovery #8 scored -0.288 and had Clean 0.8359 ± 0.0036.The corresponding schema, command, and effective-rank values were 0.8857, 1.1222, and 0.2462 for #19, and 0.9309, 1.0293, and 0.2935 for #8.Source: Robust source normalization IQR… · p. 18
Full candidate IQR PATHOLOGY PATHOLOGICAL ROBUST NORMALIZATION
- Discovery #18 scored 6.3, with Clean 0.9557 ± 0.0013, schema 0.9557, command 1.0148, and effective rank 0.011.This candidate used robust normalization, a fresh head, and source validation.Source: Full candidate IQR PATHOLOGY… · p. 18
B. Complete Candidate Configuration Matrix
- Table 4 varies architectural and training factors across twenty candidates using the same source-only splits and evaluation protocol.In channel-masking mode, nominal ratios 0.35 and 0.5 both applied p = 0.15; the heavy-masking candidate differed from the locked one only in command injection and schema weight.Source: B. Complete Candidate Configuration… · p. 18
C. Source-Validation Search Outcomes
- The final closure-stage stability check selected the schema-consistency variant with command recovery after seven valid seeds for the leading pair.The plain control collapsed across all four seeds, the large model collapsed in three of four discovery seeds, and both robust-normalization candidates showed pathology associated with near-zero IQR channels.Source: C. Source-Validation Search Outcomes · p. 18
D. Architecture-Search Interpretation
The architecture search identifies anti-collapse regularization as necessary and finds that greater model capacity is not reliably better at the available source-data scale. The accompanying protocol, roadmap, and locked-configuration passages distinguish the reported evaluation from planned extensions.
D. Architecture-Search Interpretation
- The plain control collapses on all four discovery seeds, while anti-collapse regularization raises target-encoder effective rank from approximately 5% to 58%.The large model also collapses in three of four seeds, so the search does not support a “larger is better” interpretation at this data scale.Source: D. Architecture-Search Interpretation · p. 18
E. Official Forecasting Baseline Protocol
- Persistence gives RMSE 0.654 on the target windows and 1.128 on the source test split (17 sensors).The official forecasters predict command channels poorly, with target RMSE 2.01 and 2.27; commands are exogenous inputs rather than process outcomes.Source: E. Official Forecasting Baseline… · pp. 19–20
F. Concept of Study and V2 Roadmap
- The study’s V1 pipeline covers source data, a schema-adaptive JEPA model, zero-shot transfer, and pre-lock few-shot evidence; V2 extensions are planned.Planning, policy learning, and off-policy evaluation are extensions, not current contributions.Source: F. Concept of Study and V2 Roadmap · p. 20
G. Locked Configuration
- The locked candidate’s configuration was frozen with SHA-256 hashes before its single confirmatory target pass.The RevIN variant differs only by its enabled flag.Source: G. Locked Configuration · p. 20
H. Paired RevIN Ablation
The declared post-lock target read reports better RMSE with RevIN but worse predictive calibration. Source-validation results and latent-health diagnostics qualify that comparison, while the supplied roadmap identifies causal representation learning as a conditional extension.
H. Paired RevIN Ablation
- The paired ablation reruns the locked control and a variant that enables RevIN, using three seeds per arm and regression tests of the transform.The control reruns reproduce the locked model’s source-validation metrics for seeds 0–2.Source: H. Paired RevIN Ablation · p. 20
- RevIN triggers the latent-health gate only on its raw-scale sub-criterion; its scale-free sub-criterion is better and post-fine-tuning latent health is identical.The gate was not modified.Source: H. Paired RevIN Ablation · p. 20, Table 8 · p. 23
- Figure 7 marks causal representation learning as a conditional extension, not a current contribution.The roadmap separates that future direction from the V1 work.Source: Figure 7 · p. 21, H. Paired RevIN Ablation · p. 21
- On source validation, RevIN improves clean RMSE from 0.8190 ± 0.0103 to 0.7663 ± 0.0012 and improves schema-drop, masked, long-horizon, and calibration-penalty components.Table 7 reports three paired seeds and bootstrap 95% intervals; the exact sign-flip test cannot fall below p = 0.125 with three pairs.Source: H. Paired RevIN Ablation · p. 20, Table 7 · p. 23
- Target RMSE is 0.555 ± 0.015 for control and 0.495 ± 0.004 for RevIN, while nominal 90% coverage is 0.953 and 0.874, respectively.The paired RMSE differences are −0.046, −0.079, and −0.054; NLL is 0.89 for control and 20.6 for RevIN.Source: H. Paired RevIN Ablation · p. 20, Table 9 · p. 23