Source machine with 17 sensors (teal) and 4 commands (amber) → latent state → target machine. The same weights read both; the 7 sensors the target lacks (grey) simply drop out.
arXiv preprint, 2026
project page
arXiv
code
demo
bibtex
@misc{bouaziz2026worldmodelscrossmachinecnc,
title = {World Models for Cross-Machine CNC Transfer
under Partial Sensor Overlap},
author = {Ayoub Louaye Bouaziz and Matthieu Ostertag
and Anton Demasles},
year = {2026},
eprint = {2609.16071},
archivePrefix = {arXiv},
primaryClass = {cs.LG},
doi = {10.48550/arXiv.2609.16071}
}
One set of weights reads any subset of a known sensor vocabulary. Locked on a CNC machine with 17 sensors and evaluated once, zero-shot, on a machine sharing only 10, the command-conditioned JEPA beats persistence but trails RevIN forecasters; a declared ablation shows that input normalization alone closes the gap, at a cost in calibration.
Next: verifying that the transferred model uses the commands, training on several source machines, and planning in the learned latent space.