Static-Scene Noise

The setup

The scene is an ordinary, unmodified office: a stable mix of surfaces at different distances, nothing staged or targeted for this test. It didn't move between any of the runs described here. Rather than issuing individual cal tof raw shell commands, both 254-frame series were read straight off the device's own live log stream — the I²C round-robin collection loop (see the project overview for the scheduler/cache architecture) already pushes a fresh distance frame roughly every 200–250 ms whenever the LiDAR's schedule is on, so 254 frames is about a minute of passive listening, not 254 separate commands.

Two independent series were captured back to back, same scene, same sensor configuration: room lights on, then room lights off. Per-zone mean and standard deviation are computed across all 254 frames in each series.

Flat region vs. the depth edge

The 8×8 array splits cleanly into two regimes. Columns 0–5 look at one far surface the whole time — mean distance is smooth and continuous across all 48 of those zones (1782–1851 mm), and every one of them holds a standard deviation between 4.7 and 8.8 mm. Columns 6 and 7 straddle a real edge in the scene: mean distance drops abruptly, from ~1.8 m in column 5 to 900–950 mm in column 6 to 600–870 mm in column 7. The standard deviation in that same region is both higher on average and far less consistent zone to zone — a few zones sit at 50–75 mm while their immediate neighbors are under 10 mm.

That pairing is the expected signature of a mixed-pixel (edge) effect: a zone whose field of view spans two surfaces at different ranges returns a distance that depends on exactly how the return energy splits between them frame to frame, which is inherently less repeatable than a zone looking at one flat surface. It is a property of where the zone sits relative to the scene, not a fault spread across the sensor — the 48 single-surface zones never show it.

Ambient light: no measurable effect

The VL53L8CX ranges by timing its own 940 nm VCSEL pulse, not by passively imaging the scene, so ordinary room lighting is not expected to change what it reports. Repeating the full 254-frame series with the office lights off, same scene, confirms that directly rather than assuming it.

Per-zone mean distance moved by 1.0 mm on average between the two conditions (11.0 mm at the single largest zone, still inside that zone's own frame-to-frame noise). Flat-region standard deviation moved by 0.33 mm on average — the median across those 48 zones was 5.71 mm with the lights on and 5.65 mm with them off. Neither difference is distinguishable from ordinary run-to-run variation.

Method & caveats

• Data source is the device's own CoAP-send log line (Sending lidar frame: ... dist=...), read passively off the live UART log stream while the commissioning schedule was running — not individual cal tof raw captures. Only distance_mm is logged this way; the richer per-target diagnostic fields used in the window-transmission study (signal_per_spad, range_sigma_mm, etc.) were disabled at the source in 2026-10-02 to cut per-frame I²C time, see that page and lidar.h's own notes — not available here.
• Single scene, single sensor, one distance range (~0.6–1.85 m). No claim here about how the flat-region noise floor or the edge-effect magnitude scale with distance, incidence angle, or target reflectance.
• "Flat region" / "depth edge" is a post-hoc split based on where the mean distance actually discontinues in this particular scene (columns 0–5 vs 6–7), not a property fixed to those column indices in general — point the sensor at a different scene and a different set of zones will straddle whatever edge is there.
• Standard deviation is computed per zone across all 254 frames of a series; frames with a non-zero status were excluded (none were, in either series — all 254/254 came back status=0 in both runs).