Sensing decides precision
params.py holds every dimension and assumption; build123d generates the arm and the cell, and MuJoCo runs it with lumped spring joints. A 100-trial precision study crossed four drivetrains (hobby servo, planetary, two harmonic drives) with three sensing schemes. A hobby servo read at the motor landed a median 589 µm off and hit a test pad 19 percent of the time. The same servo with an output-side encoder landed at 15 µm and hit every pad, including QFP-pitch targets. Good harmonic drives with motor sensing still sat around 50 µm. The first mock-up deflected 20 mm under a 5 N tip load; the stiffened v2 deflected 90 µm. A wrist camera reading a dot grid cut RMS error from 140 µm to 6 µm.
The v3 cell adds a tool dock with six tools (paste, vacuum, gripper, probe, iron, laser), an upward camera, a board flip frame, a preheater and feeders, and runs a full probe, place, insert, wire and solder sequence. A separate study puts the same job on a wall-rail gantry over a 72 by 24 inch desk. Everything is modelled, nothing is measured, and the docs say which figures are assumed.
Sources and method
Written from the project's code, documentation and commit history, reviewed in November 2026. Business figures come from me. Clips and screenshots are labelled with where they came from; none show customer data.
bench-arm/README.mdbench-arm/GANTRY.mdbench-arm/params.pybench-arm/out/precision.jsonbench-arm/out/stiffness_v1.jsonbench-arm/out/stiffness_v2.jsonbench-arm/out/calib.json