Skip to content

Calibration Bay Fixtures

Physical calibration bays turn sensor calibration from an ad hoc engineering activity into a repeatable production and maintenance station. This page covers the bay, fixtures, targets, surveyed references, evidence artifacts, and release gates. It does not replace calibration math, online drift monitoring, or fleet operations pages.

Scope

Calibration-bay fixtures answer four operational questions:

  1. Is the vehicle physically in a known reference frame?
  2. Are all targets, fixtures, lights, and radar reflectors in known positions?
  3. Did every sensor capture the evidence needed to produce or validate a calibration package?
  4. Can the resulting package be traced to the vehicle, sensor kit, serial numbers, firmware, operator, tool version, and validation run?

The bay is a controlled evidence source. The calibration solver may be Autoware/TIER IV, OpenCalib, an internal factory tool, or a targetless validation pipeline, but the bay must make the physical assumptions visible.

Production Pattern

Zoox describes a production end-of-line flow where each robotaxi enters a calibration bay before road-readiness tests. The vehicle rotates on a turntable and collects data to calibrate the full sensor suite: LiDAR, cameras, long-wave infrared, and radar. Zoox also describes the calibration bay as the station that ensures sensors are working together to produce an accurate understanding of the surroundings.

That is the correct mental model for an AV platform page: the bay is not only a camera checkerboard. It is a multi-sensor station with controlled geometry, repeatable vehicle pose, target provenance, and a signed output package that downstream readiness gates can consume.

Bay Layout

ZonePurposeFixture requirement
Vehicle datum zonePlaces base_link, wheel centers, sensor-kit datum, and ego envelope in a measured bay frameWheel stops, alignment marks, turntable index marks, floor fiducials, or surveyed parking rails
Sensor target ringGives every sensor a view of shared or modality-specific targetsRigid posts or wall rails with target IDs, surveyed coordinates, mounting height, orientation, and revision
Lighting and thermal zoneMakes visible and LWIR targets repeatableDimmable visible lighting, thermal contrast source, warm-up time, ambient temperature logging, glare control
Radar target zoneProvides radar range/angle checks, and RCS checks only when the target source supports themCalibrated radar targets or reflectors with known position, orientation, and standoff distance
Operator stationCaptures tool run, visual checks, and evidenceCalibration workstation, live topic monitor, camera/LiDAR/radar previews, tool logs, pass/fail dashboard
Evidence stationProduces release artifactsSigned calibration package, raw bag/MCAP references, target layout revision, photos, residual report, compatibility manifest

Use fixture IDs rather than informal labels. target_left_wall_03@rev_b is traceable; "the left board" is not.

Target Inventory

Target typeSensors servedWhat the fixture must record
Checkerboard / circle gridVisible cameras, sometimes LiDAR-camera correspondencesPattern size, square or circle spacing, board flatness, board pose in bay frame, lighting condition
AprilTag / ArUco boardCameras, camera-to-LiDAR PnP workflows, target identificationTag family, tag size, board ID, tag layout revision, detected pose covariance if available
Round-hole or edge targetLiDAR-camera and LiDAR-LiDAR workflowsHole diameter or edge geometry, material, reflective coating, target normal, pose in bay frame
Planar LiDAR boardLiDAR extrinsic validationPlane dimensions, normal direction, height, reflectivity, occlusion mask
Radar target / reflectorRadar range and azimuth/elevation checks; RCS checks only when explicitly sourcedTarget class, dimensions, pose, standoff distance, multipath controls, and RCS metadata if available
Heated visible targetLWIR plus visible camera registrationHeat source, temperature contrast, warm-up time, emissivity, visible pattern alignment
Floor fiducial / surveyed markVehicle and bay reference alignmentSurvey coordinate, uncertainty, datum, inspection date, damage status

OpenCalib's repository is useful as a public reference for target diversity: it lists factory calibration tools for chessboard, circle board, vertical board, AprilTag board, ArUco marker board, and round-hole board workflows, plus calibration coverage across camera, LiDAR, IMU, and radar. TIER IV's CalibrationTools similarly lists camera, LiDAR, and radar calibration tools, including camera intrinsics, camera-LiDAR, LiDAR-LiDAR, radar-LiDAR marker calibration, and tag-based PnP flows.

Surveyed Reference Frame

The bay needs its own frame tree, separate from the vehicle runtime TF tree:

bay_world
  -> turntable_center
  -> vehicle_parking_pose
  -> target_<id>
  -> floor_mark_<id>

At run time, the vehicle calibration package should connect:

bay_world
  -> vehicle_base_link
  -> sensor_kit_base_link
  -> camera_<n> / lidar_<n> / radar_<n> / thermal_<n>

Every target pose should include:

  • Coordinate frame and datum.
  • Survey method and estimated uncertainty.
  • Target revision and physical dimensions.
  • Last inspection date.
  • Damage, movement, or replacement status.
  • Environmental condition if it changes measurements, such as temperature, lighting, or floor wetness.

For airport or industrial fleets, the same principle applies to site survey control. The bay frame should not silently drift from the map, stand, depot, or maintenance-cell coordinate system used for validation evidence.

Run Workflow

StepActionEvidence
1. Vehicle intakeConfirm vehicle ID, platform, sensor-kit ID, sensor serials, firmware, active runtime version, and physical maintenance ticketIntake record and inventory scan
2. Bay setup checkConfirm target layout revision, target inspection status, lighting/thermal state, floor cleanliness, and radar reflector placementBay readiness checklist
3. Vehicle positioningPlace vehicle on turntable, rails, wheel stops, or marked pose; record parking residualVehicle pose record and photos
4. Data captureRecord synchronized images, point clouds, radar detections, thermal frames, IMU/GNSS/wheel state if requiredBag/MCAP ID, topic list, timestamp/clock state
5. Solver or validator runRun calibrated toolchain and store residuals, covariance/confidence, selected correspondences, rejected measurements, and warningsTool log and residual report
6. Cross-modal previewSave projection or registration previews for camera-LiDAR, radar-LiDAR, thermal-visible, and multi-LiDAR overlapsBefore/after visual evidence
7. Package signingWrite calibration package with frame tree hash, sensor serials, firmware, target layout revision, tool version, and operator/pipeline IDSigned calibration artifact
8. Release gateCheck package against readiness, validation, and fleet compatibility rules before autonomous usePass/fail report and manifest update

Autoware's LiDAR-camera tutorial shows the kind of implementation detail the bay workflow must preserve: the launch configuration names image, camera-info, pointcloud, frame, and calibration-output paths; the interactive calibrator requires matched image/LiDAR points; and the saved transform is written back into the sensor-kit calibration parameters. A production bay should make those inputs auditable rather than relying on an operator remembering which topics and frames were active.

Acceptance Gates

GateMinimum conditionFails when
B0 bay readinessTarget layout, survey revision, lighting state, radar reflectors, and floor marks match the approved configurationTarget moved, missing, damaged, uninspected, or replaced without a revision
B1 vehicle identityVehicle, sensor kit, serials, firmware, and maintenance ticket match the intended calibration jobPackage could be applied to the wrong vehicle or physical sensor
B2 time validitySensor timestamps, clock source, trigger mode, and bag replay policy are recordedHost receipt time or mixed clock domains are used without approval
B3 geometry observabilityEvery calibrated pair has enough shared targets, overlap, or motion excitation for the solver usedFlat walls, occlusion, weak target coverage, or insufficient correspondences dominate
B4 residual envelopePer-modality residuals and uncertainty are inside the release threshold for the platform and ODDResiduals pass globally but fail for a safety-critical pair or near-field zone
B5 cross-modal previewHuman- or tool-reviewable projections/registrations are stored for release evidenceNumeric pass has no inspectable evidence
B6 compatibilityPackage is compatible with active vehicle geometry, runtime, map, sensor firmware, and route/site manifestOld package, wrong frame tree, incompatible firmware, or unapproved target revision
B7 maintenance recoveryPost-maintenance recalibration links to the replaced part, torque/fixture checks, photos, and before/after residualsVehicle returns to service without physical root-cause evidence

The exact numeric thresholds belong to platform-specific validation. The existing multi-LiDAR calibration page uses sub-centimeter / sub-0.1-degree examples for target-based initialization, but a bay page should treat those as release targets to validate per sensor stack, range, ODD, and safety case.

Airside And Industrial Fit

DomainBay implication
Road robotaxiTurntable or target ring can cover full-surround perception and end-of-line production flow. Include LWIR/radar if installed, not only RGB/LiDAR.
Airport airsideAdd near-field ground targets, apron-marking and stand-geometry proxies, wet-apron/retroreflector checks, worker-height targets, and strong evidence capture after sensor strikes or bracket maintenance. FAA airport-marking standards cover runways, taxiways, and aprons; local apron/ramp practices still need site-specific modeling.
Industrial yard / portAdd mast, trailer, container, pallet, and dock-height targets; record dust, vibration, and washdown exposure before accepting residuals.
Indoor AMR / warehouseUse smaller surveyed cells, floor fiducials, rack/corner targets, near-field safety scanner checks, and battery/charging-dock pose validation.

Airside transfer should be explicit. A road-style target ring may be insufficient if the vehicle operates around aircraft skins, jet-bridge shadows, glycol film, wet concrete, retroreflective stand markings, and personnel at close range. Those effects do not necessarily change calibration math, but they change what the bay must validate and archive.

Fixture Failure Modes

Failure modeSymptomControl
Target movementResiduals shift across many vehicles after a maintenance eventTarget survey check, fixture tamper mark, target revision bump
Lighting driftCamera or thermal calibration looks unstable by time of dayControlled lighting state, warm-up timer, ambient logs, glare rejection
Radar multipathRadar reflector appears at wrong range/angle or with ghost returnsAbsorber/spacing policy, reflector standoff control, empty-bay background scan
Floor contaminationWheel stops or targets are covered by water, glycol, dust, or debrisBay readiness checklist and cleaning gate
Wrong frame namesCalibration package numerically fits but writes transforms under the wrong frameFrame whitelist, TF tree hash, runtime manifest compatibility check
Operator point-pick errorInteractive correspondence set produces plausible but wrong extrinsicsMinimum points, saved correspondences, second-review preview, outlier report
Stale target modelTool assumes old board dimensions or tag layoutTarget model version pinned in the package
Incomplete evidenceVehicle passes calibration but release cannot be audited laterRequire raw logs, previews, residuals, tool version, target layout, and signatures

Implementation Notes

  • Design the bay as a controlled measurement cell, not a storage area with boards.
  • Keep target geometry in source-controlled files or a calibration asset database.
  • Put QR/RFID labels on targets and fixtures so the tool can verify the physical layout revision.
  • Capture a background scan for radar and LiDAR multipath checks after fixture changes.
  • Store calibration bags separately from release packages; packages should point back to immutable raw evidence.
  • Keep thermal targets warm long enough to reach stable contrast before collecting LWIR evidence.
  • Run a known-good vehicle through the bay after target moves, lighting changes, turntable service, or floor repairs.
  • Treat temporary field fixtures as a degraded bay. They need explicit uncertainty, photos, and review before release use.
  • For airside validation, include wet or high-retroreflective marking checks when the ODD includes night rain or wet apron operations. FAA research found retroreflective pavement markers improved visual guidance under rainy, wet nighttime conditions, so the bay should not assume dry matte floor markings are the only relevant surface.

Sources

Public research notes collected from public sources.