Skip to content

Sensor-to-Algorithm Readiness Contract

This contract defines the minimum conditions sensor-derived data must satisfy before perception, sensor fusion, SLAM, localization, tracking, occupancy, mapping, runtime assurance, or planning-facing modules consume it.

It does not replace the detailed calibration, timing, signal-processing, runtime, or validation pages. It is the bridge that makes their handoff explicit: sensor data is acceptable only when its acquisition time, calibration package, frame tree, preprocessing state, health state, and provenance are valid for the consuming algorithm and current release manifest.

Readiness Stack

LayerConsumesProducesTypical failureAffected consumers
Physical sensorOptics, antenna, scanner, MEMS, IMU, GNSS, wheel encoder, thermal/event imagerRaw measurements and hardware statusContamination, saturation, vibration, thermal drift, degraded GNSS, wheel slipEvery downstream module using that modality
Acquisition timestampSensor clock, trigger, PTP/PPS/GNSS/vehicle clock, driver receive pathSource timestamp, receive timestamp, clock state, latency metadataHost-receive fallback, future stamps, clock-domain mix, dropped or reordered framesFusion, tracking, deskew, SLAM, replay evidence
Calibration packageIntrinsics, extrinsics, time offsets, vehicle geometry, sensor serials, firmware, tool provenanceVersioned calibration and TF treeWrong package, stale transform, weak observability, untraceable serial or firmwareProjection, fusion, occupancy, localization, map building
Frame treeMap, odom, base, sensor, image, radar, IMU, antenna, and vehicle-body framesTime-valid transform lookupMissing transform, inverted transform, stale/future TF, unapproved frame aliasProjection, registration, tracking, planning
PreprocessingRaw or feature streams plus calibration and timing stateRectified images, deskewed clouds, compensated radar, filtered features, health metadataInvisible filtering, inconsistent deskew policy, wrong projection, missing source sensor setPerception, occupancy, mapping, safety monitors
Health statePer-sensor diagnostics, cross-sensor consistency, environmental state, runtime statusGreen/yellow/red/unknown health and degradation reasonSilent confidence under soiling, rain, glare, multipath, clock drift, or packet lossRuntime assurance, degraded mode, release gates
Algorithm inputPreprocessed data, covariance/confidence, frame/time/provenance/health metadataAccepted or rejected algorithm inputConsuming stale, frame-invalid, health-unknown, or untraceable dataAll autonomy decisions and evidence claims

Calibration Gates

GateRequired evidenceBlocks or degrades when
IntrinsicsCamera model and distortion, LiDAR beam model, radar mounting model, thermal/event model where installedMissing package, stale package, wrong sensor serial, wrong firmware, high residual, unvalidated model family
ExtrinsicsSensor-to-sensor, sensor-to-IMU, sensor-to-base, sensor-kit-to-base, antenna lever arm, camera-radar and camera-LiDAR transformsInvalid TF, unexplained transform delta, weak observability, residual drift, package applied to wrong rig
Temporal calibrationSensor offset, trigger skew, LiDAR-IMU alignment, radar integration window, camera exposure midpoint, replay time policyOffset outside validation envelope, unknown clock source, mixed timestamp domains, replay wall-time/sim-time mismatch
Vehicle geometryBase frame, ego box, wheelbase, antenna lever arm, sensor occlusion mask, body-to-kit transformWrong collision envelope, wrong projection mask, map mismatch, planner clearance mismatch
ProvenanceCalibration ID, tool version, operator or pipeline, sensor serials, firmware, route/session evidence, signaturesPackage cannot be joined to vehicle, sensor kit, active runtime manifest, or validation evidence

Modality Checks

ModalityMust be explicit before algorithm usePrimary downstream risk
CameraIntrinsics, distortion model, trigger mode, exposure timestamp, rolling/global shutter assumption, ISP or RAW contract, rectification state, image frame, camera healthProjected boxes or lifted BEV features align visually but are geometrically wrong
LiDARPer-point or per-column time, scan start/end semantics, beam model, return policy, deskew reference time, intensity/ring availability, multi-LiDAR overlap alignmentAggregated clouds smear, duplicate obstacles, or corrupt scan-to-map residuals
RadarFrame timestamp, chirp/integration window, Doppler sign convention, ego-velocity compensation, radar-camera or radar-LiDAR association residual, covariance modelVelocity and range evidence is fused at the wrong time or with the wrong sign
IMU/GNSS/RTK/wheel odometryClock source, IMU axis convention, antenna phase center, lever arms, covariance/protection-level semantics, outage/holdover state, wheel scale, slip healthPose propagation appears stable while biases or lever arms corrupt map or planner coordinates
Thermal cameraTimestamp semantics, lens/window material, NUC/dead-pixel health, radiometry or contrast assumption, extrinsics to visible/LiDAR framesNight or jet-blast cues are trusted outside their calibration and health envelope
Event cameraEvent timestamp resolution, contrast threshold, polarity convention, hot-pixel filtering, extrinsics, clock sourceHigh-rate events are fused with frame sensors under inconsistent time and contrast assumptions

Preprocessing Contract

Preprocessing is a monitored and versioned contract, not invisible cleanup.

Preprocessing stepRequired metadataReject or degrade when
Image undistortion and rectificationIntrinsic ID, distortion model, rectification map version, output frame IDUnknown distortion model, stale intrinsics, untraceable rectification map
Camera exposure and rolling-shutter handlingExposure start/end or midpoint, shutter model, trigger mode, motion model if correctedTimestamp means receive time, rolling shutter ignored under high ego motion
LiDAR deskew and ego-motion compensationPer-point or per-column time, reference time, pose interpolation source, deskew versionPer-point time unavailable but downstream assumes deskewed geometry
Multi-LiDAR mergeSource sensor IDs, pairwise extrinsics, overlap health, merge frame, duplicate policyAny consumed pair is calibration-red or frame-invalid
Radar Doppler compensationRadar time model, ego velocity source, Doppler sign convention, covariance modelEgo compensation source is stale, sign convention is undocumented, integration window is unknown
Point-cloud filteringFilter version, weather-artifact policy, removed-point class or reason, source topicFiltered points cannot be distinguished from unobserved space
Projection and liftingSource/target frames, calibration IDs, timestamp used for transform lookup, projection covarianceTF lookup fails, calibration package mismatches, or target frame is ambiguous
Fused outputsSource sensor set, per-modality health, covariance/confidence semantics, provenance IDsFused output drops the ability to trace which sensors and transforms produced it

Algorithm Handoff Table

ConsumerRequired before consumption
2D/3D perceptionValid frames, source timestamps, intrinsics/extrinsics, preprocessing version, sensor health, source sensor IDs, and ODD validity
Sensor fusionCross-modal time alignment, transform validity, covariance/confidence semantics, modality health, and source provenance
SLAM/localizationDeskewed or consistently raw scans, IMU timing, extrinsics, map frame, pose covariance/protection level, residual health, and map/calibration compatibility
TrackingMeasurement timestamp, source frame, object covariance, latency budget, association confidence, and dropout/jitter state
Occupancy/free-spaceSource sensor set, blind-spot policy, unknown/free semantics, projection validity, map-frame validity, and health-aware confidence
MappingCalibration package, pose source, traversal provenance, dynamic/static filtering state, raw-log references, and map datum/frame compatibility
Runtime assurance/planningFreshness, health state, unknown policy, degraded-mode action, safety monitor state, and fail-closed behavior for invalid inputs

Reject And Degrade Rules

RuleReject or degrade conditionRequired response
Time validityStale source timestamp, future timestamp, mixed clock domain, host-receive fallback without approved degraded modeReject the message or enter a validated degraded mode
TF validityMissing transform, stale transform, future transform, unapproved frame alias, unapproved TF tree hashReject the message and raise frame/TF diagnostics
Calibration validityRed or unknown calibration for a consumed sensor pair, wrong calibration package, unexplained transform deltaRemove affected modality or stop according to the safety case
Provenance joinInput cannot join to active build, model, map, calibration, config, vehicle, and sensor-kit IDsExclude from release evidence and raise runtime/fleet diagnostics
Health stateSensor health missing, stale, unknown, or red for a safety-relevant inputTreat confidence as invalid and trigger runtime assurance policy
Replay consistencyReplay mixes wall time and simulation time or lacks bag/MCAP IDs and active manifest IDsReject replay as release evidence
Free-space conservatismUnknown or low-observation cells are promoted to traversable free space in protected zonesReject the free-space output and trigger safety monitor action

Evidence Artifacts

ArtifactContents
Calibration packageIntrinsics, extrinsics, time offsets, frame tree, vehicle geometry, sensor serials, firmware, tool version, signatures
Timing validation reportClock source, offset envelope, skew, jitter, dropout, timestamp-shift sweep, latency-jitter stress result
TF tree evidenceFrame names, transform authorities, static/dynamic split, TF tree hash, lookup failure counts
Sensor health logPer-sensor diagnostics, cross-sensor consistency, environmental state, green/yellow/red/unknown transitions
Preprocessing manifestRectification maps, deskew policy, filter versions, radar compensation settings, source topic list
Projection or registration previewBefore/after visual or numeric evidence for calibration and drift events
Replay acceptance reportBag/MCAP IDs, sim-time policy, active manifest IDs, deterministic replay settings, pass/fail gates
Runtime manifestBuild, model, map, calibration, config, vehicle, route, site, sensor-kit, and evidence IDs

Foundations:

Platform sensors:

Runtime, fleet, and validation:

Public research notes collected from public sources.