RCI 017 · Visual-system data audit

11 records. 16 error or precision values. Zero complete public reproduction packages.

“Depth accuracy” is not one portable metric. RCI separated absolute Z-error, spatial precision, plane-fit RMS, systematic error and method-undefined accuracy—then checked whether the public source disclosed the hardware, target, light, thermal state, aggregation and software needed to reproduce each statement.

Published 2026-08-1014 min readResearch dataset v0.1.0 · Catalog v0.4.5
11catalog records audited
25range observations preserved
16selected error / precision values
0complete public reproduction packages

The direct answer

Can these eleven cameras be ranked by the smallest published percentage? No. The 16 selected values span five metric boundaries, three camera architectures, different distances and lens options, and radically different levels of method disclosure. A smaller numeral can describe a different quantity rather than better real-world depth.

Is any public value fully reproducible from the cited material? No under RCI’s 13-field test. Every record names a hardware and distance boundary, but none publishes both the tested-unit sample and the exact software or firmware state. Only two records—Basler blaze-101 and blaze-102—publish ambient light, test temperature, warm-up and aggregation beside the value.

Audit result

All 11 records publish at least one numeric error or precision statement. Eight publish an ROI, five publish a target, two publish quantitative light and thermal conditions, and zero identify both the tested-unit count and exact software or firmware used to produce the headline result.

Five metric boundaries sit behind the 16 values

RCI boundaryValuesWhat it answersWhat it does not answer
Absolute depth error3How far reported depth is from a reference distance.Spatial or temporal noise.
Spatial precision1How noisy valid points are around a fitted plane.Absolute distance bias.
Depth accuracy; method undefined8A manufacturer percentage whose public table does not define the calculation.A harmonized cross-brand score.
Systematic depth accuracy2Mean systematic distance error under a fixed setup.Performance under different light, target or thermal states.
Plane-fit RMS labeled accuracy2A plane-fit RMS calculation published under a broader accuracy label.A harmonized cross-brand score.

Three RealSense records use a family Z-accuracy or absolute-error method. Gemini 335’s selected value is explicitly spatial precision. Gemini 2 and Gemini 2 XL call their rows “Depth Accuracy,” but the accompanying text calculates RMS distance to a best-fit plane without a ground-truth reference. Stereolabs publishes eight lens-and-distance percentages without defining the metric beside the table. Basler defines its ±5 mm result as systematic error.

The eleven record boundaries

RecordSelected published value(s)Metric boundaryFully published fieldsEditorial boundary
RealSense D455
D455 camera using the D450 depth module; D455f filter variant is not substituted
<2% at 4 mAbsolute depth error4 / 13An ideal range and a percentage at one distance answer different questions. D455 is not ranked against plane-fit precision or method-undefined accuracy values.
Orbbec Gemini 335
Gemini 335 / G40155-170, not the longer-baseline 335L or filtered 336 variants
≤1.5% RMSE at 2 mSpatial precision6 / 13The 0.26 m versus 0.3 m current-source difference remains explicit, and the theoretical 65 m value is not merged into the ideal range.
Stereolabs ZED X
ZED X with separate 2.2 mm and 4 mm lens columns; ZED X Mini is excluded
<0.8% at 2 m · <4% at 12 m · <0.4% to 2 m · <7% at 20 mDepth accuracy; method undefined2 / 13The two lens options are not collapsed into one range or one accuracy value and are not ranked against cameras that publish a different metric.
Basler blaze-101
blaze-101, Sony DepthSense IMX556 ToF camera
±5 mm from 0.5–5.5 mSystematic depth accuracy9 / 13The no-ambient-light ±5 mm test is separate from Basler’s 6 m ambient-light robustness setup and must not inherit that setup’s illumination condition.
RealSense D435i
D435i camera using the D430 depth module and integrated IMU
<2% at 2 mAbsolute depth error4 / 13D435i’s 2 m value and D455/D457’s 4 m values are not a one-number cross-model ranking.
Orbbec Gemini 2
Gemini 2 / G20155-15, 50 mm baseline
≤2% at 2 mPlane-fit RMS labeled accuracy5 / 13The 10 m working range, 5 m optimal range and 2 m plane-fit value are separate boundaries.
Stereolabs ZED 2i 2.1 mm
ZED 2i with the 2.1 mm lens; 4 mm lens values are excluded
<0.8% at 2 m · <4% at 12 mDepth accuracy; method undefined2 / 13The 4 mm option is a different hardware boundary and does not populate this exact catalog record.
RealSense D457
D457 camera using the D450 optical module with GMSL/FAKRA transport
<2% at 4 mAbsolute depth error4 / 13Transport and enclosure differences do not make the D457 accuracy row independent of its D450 depth-module and method boundary.
Orbbec Gemini 2 XL
Gemini 2 XL camera plus OBox processing boundary, G30056-66 / G30056-86
≤2% at 4 mPlane-fit RMS labeled accuracy5 / 13The 20 m working interval, 10 m optimal interval, 5 m binned interval and 4 m error statement remain separate.
Stereolabs ZED X Mini 2.2 mm
ZED X Mini with the 2.2 mm lens; 4 mm lens and full-size ZED X are excluded
<1.6% at 2 m · <3% at 4 mDepth accuracy; method undefined2 / 13The old product-page U state is resolved only through the current versioned V1.2 datasheet and remains lens-specific.
Basler blaze-102
blaze-102 industrial ToF camera
±5 mm from 0.5–5.5 mSystematic depth accuracy9 / 13The fixed-condition accuracy test and the separate ambient-light robustness test are not merged.

The row count is a catalog count, not a claim of eleven independent test programs. D455 and D457 share a D450 optical-module boundary; blaze-101 and blaze-102 share Basler’s published test conditions; ZED X contributes four values because the two lenses and two distances remain explicit.

The 13-field disclosure matrix

“Published” means the selected source bundle states a usable, claim-specific value. “Partial” means a general method or qualitative condition exists but does not identify the setup that produced the headline number. An operating limit is not counted as a test temperature, and a documentation revision is not counted as the camera firmware used in the test.

Reproduction fieldPublishedPartialNot publishedN/A
Exact hardware / lens boundary11 / 11000
Metric definition and calculation6 / 11230
Distance or distance interval11 / 11000
Resolution / acquisition mode3 / 11530
Region of interest8 / 11030
Target geometry / reflectivity5 / 11330
Ambient illumination2 / 11450
Test temperature2 / 11090
Warm-up / thermal state2 / 11090
Aggregation / frame count2 / 11540
Software / firmware version0 / 110110
Tested-unit sample count0 / 111100
Ground-truth reference0 / 11551

Basler is the most transparent public boundary in this sample. Its current documentation states default settings, 20 minutes of warm-up, a flat white 90% target, no ambient light, 22°C, a central 40 × 40-pixel region and a 25-image average. It still does not identify the exact ground-truth instrument, the number of cameras behind the published result or the firmware and processing state, so RCI scores nine fields as published—not thirteen.

RealSense publishes a useful method, but not the complete claim-generation record. The D400 method defines Z-accuracy, recommends a wall or textured target and describes independent ground-truth measurement and report capture. The product specification supplies distance, HD resolution and 80% ROI. It does not tie a particular target, illumination, temperature, frame count, firmware or unit sample to each headline percentage.

The audit found current-source drift and a metric-label mismatch

Gemini 335 lower recommended boundary: the current product page states a 0.26–3 m optimal range. The current V1.6 datasheet states a 0.3–3 m ideal range. Both state 0.1–20 m+ working reach and a theoretical maximum up to 65 m. Catalog v0.4.5 preserves the 0.26/0.3 difference as a C-state claim because neither source names the hardware, firmware or test revision behind it.

Gemini 2 family metric wording: the Gemini 2 v1.1 datasheet and Gemini 2 XL v1.0 datasheet label their rows “Depth Accuracy,” then describe RMS distance from valid points to a best-fit plane. That calculation is closer to spatial precision than ground-truth absolute error. RCI publishes the label and method together instead of silently choosing one interpretation.

ZED X wording: the ZED X V1.2 datasheet prints “<0.4% to 2 m” in the 4 mm column while parallel cells use “at.” The machine-readable dataset preserves the displayed phrase and normalizes only the distance field to 2 m.

Twenty-five range observations, not eleven maximums

RecordPreserved range observationsWhy one maximum is insufficient
RealSense D455Ideal range: 0.6–6 mThe single ideal interval still does not equal an accuracy-qualified maximum.
Orbbec Gemini 335Working range: 0.1–20 m+ · Product-page optimal range: 0.26–3 m · V1.6 datasheet ideal range: 0.3–3 m · Theoretical maximum: Up to 65 mLens, mode, source wording or ideal/maximum boundary changes the interval.
Stereolabs ZED XMaximum depth range: 0.3–20 m (ZED X 2.2 mm) · Ideal depth range: 0.3–12 m (ZED X 2.2 mm) · Maximum depth range: 1–35 m (ZED X 4 mm) · Ideal depth range: 1–20 m (ZED X 4 mm)Lens, mode, source wording or ideal/maximum boundary changes the interval.
Basler blaze-101Short-range mode: 0.3–1.5 m (Short) · Long-range mode: 0.3–10 m (Long)Lens, mode, source wording or ideal/maximum boundary changes the interval.
RealSense D435iIdeal range: 0.3–3 mThe single ideal interval still does not equal an accuracy-qualified maximum.
Orbbec Gemini 2Unbinned working range: 0.15–10 m (Unbinned Dense or Sparse Default) · Unbinned optimal range: 0.2–5 m (Unbinned Dense or Sparse Default) · Binned optimal range: 0.15–2.5 m (Binned Sparse Default, 640 × 400)Lens, mode, source wording or ideal/maximum boundary changes the interval.
Stereolabs ZED 2i 2.1 mmMaximum depth range: 0.3–20 m (2.1 mm lens) · Ideal depth range: 0.3–12 m (2.1 mm lens)Lens, mode, source wording or ideal/maximum boundary changes the interval.
RealSense D457Ideal range: 0.6–6 mThe single ideal interval still does not equal an accuracy-qualified maximum.
Orbbec Gemini 2 XLUnbinned working range: 0.4–20 m (Unbinned Dense or Sparse Default) · Unbinned optimal range: 0.4–10 m (Unbinned Dense or Sparse Default) · Binned optimal range: 0.4–5 m (Binned Sparse Default, 640 × 400)Lens, mode, source wording or ideal/maximum boundary changes the interval.
Stereolabs ZED X Mini 2.2 mmMaximum depth range: 0.1–8 m (ZED X Mini 2.2 mm) · Ideal depth range: 0.1–4 m (ZED X Mini 2.2 mm)Lens, mode, source wording or ideal/maximum boundary changes the interval.
Basler blaze-102Short-range mode: 0.3–1.5 m (Short) · Long-range mode: 0.3–10 m (Long)Lens, mode, source wording or ideal/maximum boundary changes the interval.

Working, ideal, optimal, maximum and theoretical-maximum ranges are separate source claims. The Gemini 335 theoretical 65 m boundary is not a 65 m accuracy result. ZED X’s 35 m maximum belongs to the 4 mm lens, while its 20 m ideal upper bound carries a separate percentage. Gemini 2 XL’s 20 m working interval narrows to 10 m optimal in unbinned modes and 5 m optimal in the binned mode.

What to request before selecting a robot depth camera

  1. Identity: exact camera, lens, filter, hardware revision, processing box and interface configuration.
  2. Metric: signed or absolute Z-error, spatial RMSE, temporal noise, systematic bias, fill rate or another calculation—with formula and outlier policy.
  3. Scene: target material, reflectivity, texture, size, orientation, distance and whether it fills the stated ROI.
  4. Environment: illuminance or spectral irradiance, temperature, humidity, warm-up, vibration and multi-camera interference.
  5. Acquisition: resolution, frame rate, exposure, projector power, preset, filters, ROI and aggregation count.
  6. Traceability: ground-truth instrument and uncertainty, tested-unit count, raw results, camera firmware, SDK and algorithm version.
  7. Robot test: repeat the test with the actual mounting, motion, surfaces, sunlight and perception stack before making a safety or procurement decision.

Method and limitations

RCI audited the eleven records already in the robot depth-camera comparison group. The selected error observation is the catalog-facing accuracy or precision field, not an exhaustive extraction of every adjacent metric. Each distinct lens-and-distance value becomes one observation; each meaningful working, ideal, optimal, maximum or mode range becomes a separate range observation. Current manufacturer pages, versioned datasheets and linked manufacturer methods were checked on 2026-08-10.

A reproduction package is complete only when every applicable field in the 13-field matrix is published. A generic method can earn “partial” but does not prove that the headline specification used that exact target or environment. “Not applicable” is used only where the chosen metric genuinely does not require a ground-truth distance. This is a public-evidence audit, not independent camera testing, a market census or a safety certification.

Download the observation-level data

Download the immutable research dataset as JSON or CSV. Stable current aliases are /data/research/depth-camera-range-error-claims.json and /data/research/depth-camera-range-error-claims.csv. Updated component evidence is in catalog v0.4.5 JSON and its claim CSV.

Suggested citation: Robot Component Index, “Depth-Camera Error Claims: 11 Records, 16 Selected Values, 0 Complete Reproduction Packages,” RCI 017, research dataset v0.1.0 and component catalog v0.4.5, 2026-08-10.