Tokenization for continuous robot actions

Continuous actions can be discretized with piecewise-linear bins or learned vocabularies. This post compares inference latency and tracking error on a 7-DOF arm doing contact-rich insertion (filler setup).

We find that a modest codebook size keeps step time under the 20ms budget while preserving enough resolution near singularities. Full tables are placeholders for now.

As a toy model for turning a continuous field into a finite vocabulary, this lab rasterizes analytic silhouettes into a coverage map, assigns each cell a VATOM tile, then asks which counts of full / half / quarter atoms can exactly cover the pixel mask—mirroring our pygame decomposition prototype in the browser.