Skip to content

AV Data Evaluation Fundamentals

AV Data Evaluation Fundamentals curated visual

Visual: AV evaluation split firewall showing data sources, split manifests, scenario/ODD coverage, leakage gates, open-loop/pseudo/closed-loop benchmark ladder, and release evidence.

Why It Matters

AV data evaluation is the contract between a claim and the evidence used to support it. A dataset split, benchmark score, scenario label, or closed-loop metric only means something when it is tied to the model, task, ODD, protocol, metric, uncertainty, and release decision being made.

The failure mode is not just overfitting. A system can improve an offline benchmark while exploiting near-duplicate frames, map-version leakage, route familiarity, simulator artifacts, repeated submissions, weak scenario tags, or an open-loop metric that does not represent closed-loop behavior. AV evaluation therefore has to bind data lineage, split hygiene, scenario and ODD coverage, benchmark semantics, and release artifacts into one auditable package.

This page is narrower than the general ML leakage and calibration note. It focuses on AV-specific evidence design: logged driving data, scenario libraries, simulator and pseudo-simulator benchmarks, public leaderboard interpretation, cross-domain transfer, and the release bundle expected before a learned autonomy component is promoted.

Evaluation Claim Contract

Start by writing the claim before selecting the metric:

text
claim =
  system or component
  + task and interface
  + ODD and scenario family
  + data source and split rule
  + benchmark or protocol version
  + metric and uncertainty
  + release decision

Examples:

  • A planner has better route progress in real-log pseudo-simulation without increasing collisions on held-out rainy urban scenes.
  • A perception stack has higher pedestrian recall at fixed false positives per hour on unseen night sites.
  • A VLA policy has lower intervention rate in closed-loop simulator routes that include held-out construction and cut-in scenarios.
  • An airside tug planner completes stand-entry routes in a digital twin while respecting aircraft, personnel, and hold-line gates.

A score without the claim is not release evidence. It may still be a useful regression signal, but it cannot justify deployment scope.

Split And Leakage Firewall

AV data often contains correlated frames, repeated routes, repeated actors, reused maps, and shared simulator generators. Random frame splits are usually too weak for deployment claims.

Useful split units include:

Generalization claimSplit unit
Same site, new timeDate, shift, season, lighting, and weather block
New route or operational zoneRoute ID, lane graph region, stand, dock, yard block, or mine segment
New geographyCity, airport, warehouse, port, mine, or farm site
New sensor stateVehicle ID, sensor rig, calibration ID, firmware, cleaning state, or mounting revision
New map stateMap version, lane graph version, semantic layer version, construction closure, or temporary geofence
New scenario familyScenario tag, hazard family, actor interaction, failure trigger, or safety requirement
New simulator behaviorGenerator seed, simulator version, agent policy, scenario template, map, and asset pack

The split manifest is a release artifact, not a convenience file. It should record:

text
dataset version
source logs and lineage
split IDs and split rationale
scenario and ODD tags
map and calibration versions
label, pseudo-label, and annotation-tool versions
preprocessing and feature-extraction commits
benchmark version and metric configuration
model checkpoint, threshold, temperature, and policy hash
submission count and tuning history
known exclusions and blocked-access gaps

Leakage gates should explicitly check near-duplicate frames, clips from the same route traversal, repeated rare events, copied pseudo-labels, shared human-review batches, train-time access to test maps or scenario labels, leaderboard-driven threshold tuning, and simulator assets reused across training and held-out tests.

Scenario And ODD Coverage

Scenario coverage is the bridge between logged data and safety meaning. ISO 34502 defines a scenario-based ADS safety evaluation framework for road vehicles. ASAM OpenODD provides a way to describe the operational design domain. ASAM OpenSCENARIO DSL provides an executable scenario-description layer for abstract, logical, and concrete scenarios.

For AV data evaluation, use three levels:

LevelWhat it capturesEvaluation use
Abstract scenarioNatural-language family such as "vehicle cuts in" or "baggage tractor merges at stand entry"Backlog, safety case, coverage target
Logical scenarioParameter ranges for actors, speeds, weather, lighting, topology, and marginsCoverage analysis and test generation
Concrete scenarioFixed initial state, actors, map, timing, and oracleReplay, simulator test, regression, release gate

ODD coverage should be reported by scenario family, environment, infrastructure, sensor state, traffic or actor density, map state, and authority boundary. Aggregates are not enough: a high mean score can hide a total miss on a small but safety-critical slice.

Evaluation Ladder

Use the cheapest protocol that can support the claim, but do not let cheaper protocols stand in for interaction evidence.

LayerWhat it testsMain useMain limitation
Offline/open-loopPredictions against logged labels or futuresFast training and regressionLogged future is only one possible outcome
Non-reactive replayEgo trajectory scored against real scene without reactive agentsCheap planning sanity check on real logsOther actors do not respond to ego behavior
Pseudo-simulationReal-data benchmark with synthetic observations or approximated future statesScalable planner and E2E comparisonStill approximates interaction and reconstruction
Closed-loop simulatorEgo actions affect future observations and agentsCompounding-error, recovery, progress, and interaction evidenceSim-to-real and behavior-model realism gaps
Sim-agent realismGenerated agents or futures evaluated for realismScenario generation and behavior-model validationNot the same as proving ego stack safety
Controlled site and shadow modeReal sensors, maps, operators, and proceduresRelease evidence and ODD-specific risk burn-downExpensive, harder to repeat, limited rare-event rate

Open-loop evidence is necessary for ML development. Closed-loop and controlled-site evidence are necessary for claims about autonomy behavior under interaction, recovery, fallback, and rule compliance.

Public Benchmark Interpretation

Public benchmarks are useful when their protocol matches the claim. They are weak when treated as generic proof of readiness.

Benchmark or sourceWhat it is good forInterpretation caveat
NAVSIMReal-data non-reactive and pseudo-simulation evaluation for autonomous driving planners and E2E systemsGood planning evidence, but not full interaction or domain transfer proof
Bench2DriveClosed-loop CARLA benchmark with multi-ability routes and rich annotationsGood interaction stress test, but simulator and road-domain assumptions remain
CARLA LeaderboardClosed-loop route completion, infraction, and driving-score style evaluationUseful for full-stack route behavior, but CARLA-specific artifacts and route rules matter
Waymo Open DatasetLarge public AV perception, motion, E2E, and scenario-generation benchmark ecosystemStrong road-data reference, but leaderboard metrics and yearly tracks must be read by task
Waymo Open Sim AgentsChallenge for realistic future behavior of all agents in scenariosMeasures behavior realism, not direct ego-stack deployment safety
nuPlanOpen-loop and closed-loop planner benchmark on real-world driving dataUseful planner benchmark, but still road-domain and protocol-bound

Public road benchmarks transfer as methodology to warehouses, yards, ports, mines, construction sites, farms, campuses, delivery robots, and airside autonomy. They do not transfer their ODD, actor ontology, rules, or safety oracles.

Domain Fit

DomainData-evaluation focus
Road AVCity, route, weather, traffic-density, map-version, and vulnerable-road-user splits; public benchmarks are most directly aligned here.
AirsideAircraft, GSE, personnel, stand, service-road, jet-blast, FOD, marshalling, clearance, and airport-rule coverage; public road benchmarks provide evaluation patterns, not sufficient evidence.
Warehouse, yard, and port autonomyDock, aisle, trailer, container, pedestrian, forklift, gate, and mixed-manual-traffic coverage; route/task progress and near-field safety gates matter more than public-road rules.
Mining, construction, and agricultureSite phase, haul road, implement, terrain, dust, slope, GNSS degradation, and exclusion-zone coverage; rare physical hazards need controlled-site evidence.
Campus and delivery robotsSidewalk, curb, doorway, pedestrian, accessibility, weather, and municipal-rule coverage; human-interaction slices should not be hidden in mean route scores.

Failure Modes

Failure modeSymptomControl
Random-frame leakageValidation looks strong while new routes or new days failSplit by route, time, site, and clip family
Map-version leakagePlanner appears robust because train and test share map artifactsBind data to map version and hold out map changes
Scenario undercoverageMean score improves while a hazard class regressesReport metrics by scenario family and ODD tag
Open-loop optimismLow trajectory error but poor recovery or progress in simulatorRequire pseudo-sim, closed-loop, or controlled-site evidence for behavior claims
Leaderboard tuningPrivate test improves through repeated submissions rather than generalizationTrack submissions, freeze thresholds, preserve untouched holdouts
Simulator exploitationPolicy learns renderer, agent, or route artifactsCross-simulator checks, real-log pseudo-sim, and controlled real tests
Weak uncertaintyMetric difference is smaller than run-to-run varianceConfidence intervals, paired tests, and seed/run manifests
Domain overclaimRoad benchmark score is used as airside, yard, or mining readinessState domain transfer limits and build domain-specific oracles

Implementation Checklist

  • Write the release claim before choosing metrics.
  • Define split units that match the generalization claim.
  • Store a split manifest with data, map, calibration, label, benchmark, model, and threshold versions.
  • Report scenario and ODD coverage before aggregate scores.
  • Separate open-loop, non-reactive, pseudo-sim, closed-loop, sim-agent, shadow, and controlled-site results.
  • Preserve per-scenario traces, safety-gate outcomes, planner/controller states, and intervention labels.
  • Freeze benchmark and metric versions for release comparisons.
  • Track public leaderboard submission count and tuning decisions.
  • Require statistical uncertainty for comparisons that affect release decisions.
  • Record blocked source or dataset access as a gap instead of filling it with inferred claims.

Release Artifact Bundle

A reviewable AV evaluation package should include:

text
claim statement
ODD and scenario taxonomy
dataset and split manifest
data lineage and consent/retention constraints
map, calibration, and sensor-health versions
annotation and pseudo-label provenance
benchmark and simulator versions
metric implementation and threshold config
model checkpoint and runtime configuration
per-scenario score tables
uncertainty intervals and regression analysis
known exclusions and blocked gaps
release decision and owner sign-off

The bundle should make it possible to answer a concrete question: "What exactly did this evidence prove, for which autonomy component, in which ODD, under which split and metric assumptions?"

Sources

Public research notes collected from public sources.