One sensor. Several engineering stories.

CROSS-EPISODE FIELD NOTES · UPDATED 22 OCTOBER 2026

What changed across the episodes.

I first wanted to put colour on a point cloud. Over the next experiments, the OS0 became the geometry I kept coming back to: a stream to display, a room to reconstruct, a coordinate frame to recover, and a comparison for another depth sensor.

EP03 → EP04: getting points onto a screen was only the start.

In Episode 3, a low-cost Raspberry Pi camera met the OS0. Checkerboard and manual matching improved selected static-boundary alignment from about 13.5 to 4.2 pixels. The person ghost did not disappear. That split mattered: a better fit on the room did not solve timing or decide which surface the camera could actually see.

Episode 4 moved the data to an old Samsung S10 through a Mac bridge. It made a useful handheld viewer, but pose jumps, IMU compatibility and a disappearing Ethernet adapter interrupted the bigger mapping idea. The phone was displaying scans; it was not localising itself in the map.

EP08 → EP10: a room map becomes something to align against.

In Episode 8, the static scan supplied geometry for the office twin while phone video supplied appearance. Those captures were four days apart. Their different roles—and the work of aligning them—became part of the experiment instead of disappearing behind a polished rendering.

By Episode 10, the setup had moved. Bringing the live sensor back into the old map meant recovering its pose and fitting the iPhone view again. The first wall alignment looked convincing while putting the camera in the wrong physical place. Later iterations used scene features and my moving silhouette across 244 frames.

The saved third iteration reported about two pixels of residual on the room features and a time offset near −1.1 seconds within that recording, using estimated camera intrinsics. My LiDAR body clusters then became a rough avatar. Heading flips and missing anatomy showed why a plausible moving character is a different result from a fully observed person.

EP14 → EP15: the one-box idea inherits these questions.

The repeated integration work led to the RGB + depth module idea: a camera, depth sensor and IMU with a shared clock and known alignment. Episode 14 brought up an epc660-class iToF sample beside a Pi camera. Calibration improved from 12.7 to 4.56 pixels but missed the stated acceptance thresholds. That was a separate sensor experiment, not an OS0 upgrade.

In Episode 15, the iToF sensor bowed a wall by 12–14 cm across roughly 4.5 m while the Ouster showed it straight. Exposure changes reduced noise, but modulation frequencies still disagreed on distance. A correction fitted to that one wall improved plane RMS locally; it did not establish accuracy across ranges. The next useful tests are a range sweep and a spill test, followed by a dToF comparison.