RGCAT

RGCAT — the Roboforming Geometric Coverage Analysis Toolkit — is a Python toolkit for roboforming, the incremental sheet-forming process in which one or more robot arms push a blank into shape along a path instead of pressing it in a die. A campaign of such runs produces a table of geometries, and the question this toolkit exists to answer is which shapes the campaign has not made: where the covered region of the geometry space has holes, and which candidate geometry would fill the largest one.

It is being built by Per Starke at Ruhr-Universität Bochum as a master’s thesis, supervised by Jan Schäfer.

Why it is here before it exists

The project’s release page lists what will be published with the source code, and two of the items are the reason this tile is on the dashboard:

• Python source code
• Docker and Conda environments
• Usage and configuration guides
• Pipeline walkthrough
• arc42-based architecture documentation
• Architecture decision records
• MIT license

The documentation is being written with the system, by the person building it, and it is arriving with the first public release rather than after it. What is new here is only that the dashboard now has somewhere to say so before the chapters exist.

What it does, so far as the project says

The pipeline is config-driven — one YAML file selects each stage — and runs in nine steps: ingestion of CSV, Parquet or Feather tables; feature expansion into flat vectors and z-map features; representation by PCA, VAE or VaDE; clustering by KMeans, HDBSCAN or a combination; k-nearest-neighbour coverage and sparse-region metrics; gap detection by candidate sampling and ranking; reconstruction of a candidate back into the original feature space; an analytical before/after evaluation of adding that candidate; and a timestamped run folder holding metrics, plots, logs, a config snapshot and a static HTML report.

Reproducibility is the stated design driver rather than a quality goal added afterwards: every run writes the same folder structure so that two pipeline variants can be compared on identical outputs. The project is careful to say that reconstructed candidates are analytical suggestions, not validated process data.

When it lands

The source repository is not public yet; the software, environments, documentation and licence files are to be published together. When the arc42 chapters appear, they will be reproduced here in full under the MIT licence the project announces — the same basis on which docToolchain v4 is hosted.