Manuscripts based on experimental work on various maker and coding projects, sorted by
topic. Each one is a self-archived preprint on Zenodo with its own permanent DOI โ not peer-reviewed.
The informal companion to the two papers below: seven weeks of building an imitation-learning lab around a self-built SO-101 arm โ the incidents that cost the most time, what to warn a newcomer about, and the threads left unfinished.
Four quantified failure modes from an ACT pick-and-place deployment: a camera artifact that produced zero successes, three distinct kinds of background failure, a lighting threshold near 400 lux, and a confirmed baseline of 86.3% on trained and 53.6% on novel positions.
Why a published text-triggered backdoor did not reproduce on an ACT policy, a three-level test for whether a trigger can reach a policy at all, and a visual backdoor that does work.
A relational schema, composition calculator, alloy-design screening metrics and automated cross-referencing against three public materials databases for a single working alloy lab โ reported as a systems paper, including the engineering failures and how each was fixed.
Reviving a laser plotter after eight years in storage: replacement carriage plates, the firmware's non-standard acknowledgement string that defeated every modern G-code sender, and a GRBL-based engraving workflow accurate to about 0.2% over 50 mm.