Tao Yun Huang 黃 韜 云

The problem. Thirteen boards on one CV bus is a loading problem before it is a music problem. Each board added draws on the same reference, so pitch drifts as more come online; adjacent voices coupled into each other, and the shared supply sagged under simultaneous gates.

How it was found. Board by board on the bench, not by redesigning the rack — bus loading measured against reference droop, crosstalk isolated by driving one voice and scoping its neighbours, decoupling reworked until the rails held at worst case.

The axis. Half a tonne over 20 m. A single position loop overshoots where inertia is low and lags where it is high — the inner torque loop absorbs that variation.

Absolute feedback. Position is known at power-on — no reference run down 20 m of rail with 500 kg on it. The power-off brake holds on E-stop rather than coasting.

Registration. Time-indexed playback would drift the first time the axis lagged under load, and the error would accumulate. Indexing to encoder position cannot.

Localised without markers. The building could not be modified, so the corridor was photographed densely and processed into a 3-D feature map; at runtime the device estimates full 6-DoF pose by matching live camera frames against it.

Why that corridor is the hard case. Sparse texture that repeats — every bay looks like the next. Pose locks onto the wrong bay, or drifts, and the overlay slides off the wall. The work was in deciding when to trust a match and when to discard it.

Gas side. Two supplies, regulated delivery and valve control on a repeating cycle, phase-locked to what the viewer sees. Holding a steady rhythm is a matter of valve response, line volume and settling time — the commanded profile and the delivered profile are not the same curve.

Stated plainly. This is gas handling and flow fault-finding at atmospheric pressure. It is not vacuum or plasma process experience, and is not offered as such.

Why a feedback loop is a stability problem. Any loop with real mechanics in it has dead time: solenoids and pneumatics do not respond instantly, and the structure keeps ringing after the actuator stops. Feed that into a stage reacting in milliseconds and the system runs away or hunts.

Three versions, three rebuilds. Each was rebuilt around what failed in the last. v2 ran the modified RAVE pipeline across a 49.4-channel Ambisonics system — the first at that scale in Taiwan.

The real problem is threading. Blender's Python API is not thread-safe — call bpy from the socket's worker thread and the process dies, sometimes several operations later. The server accepts on a daemon thread but never touches the scene there: scripts go onto a lock-guarded queue that a bpy.app.timers callback drains on the main thread. That one indirection is the whole design.

Why build it. The same reason I instrument anything — describe the scene once and have it constructed repeatably, rather than rebuilding it by hand every time I change my mind.