Help & Reference

My-Curator · Scenario DNA v0.1 Reference Guide

Platform Overview

What My-Curator does and how it fits into the AV data pipeline.

My-Curator is a Verify-by-Exception (VBE) curation platform for autonomous-vehicle driving clips. A NVIDIA DeepStream 9.0 pipeline ingests raw video, a Scout VLM (Cosmos-Reason2-8B FP8) generates structured Scenario DNA for each clip, and reviewers curate only the exceptions — clips flagged as elevated or critical risk.

Each clip carries a 4-layer DNA descriptor: ODD (environment), Topology (road infrastructure), Actor Dynamics (road users), and Planner Logic (ego intent + risk). DNA is stored in PostgreSQL (JSONB + GIN index) and in Milvus (768-dim Cosmos-Embed1-336p embeddings) for hybrid search.

Verify-by-Exception4-Layer DNAHybrid Vector + Filter SearchISO 21448 SOTIFASAM OSI / OpenDRIVE Taxonomy

How to Use

The four-step curation workflow from raw video to curated corpus.

1Ingest

DS pipeline captures frames, Scout generates DNA, clip is stored in PG + Milvus with review_status = pending.

2Search & Curate

Use the hybrid search page to query clips by natural language or DNA filters. Results are ranked by cosine similarity.

3Review Queue

Work through pending clips. Click any card to open the full detail view with video playback, DNA accordion, and similar clips.

4Approve / Reject

Approve clips to add them to the curated corpus. Reject duplicates, low-quality, or mislabeled clips.

Layer 1

ODD — Operational Design Domain

Environmental conditions under which the clip was captured.

Source: ASAM OpenSCENARIO v1.0 · ASAM OSI v3.x · ISO 22736

weatherenum · required

Precipitation and atmospheric state at clip time.

↳ ASAM OpenSCENARIO v1.0 CloudState + ASAM OSI PrecipitationIntensity

clearNo precipitation, clear sky

overcastOvercast sky, no precipitation

light_rainLight rainfall

heavy_rainHeavy rainfall

snowModerate snowfall

heavy_snowHeavy snowfall

fogFog — low horizontal visibility

mistMist — reduced visibility, less severe than fog

sleetSleet — mixed rain, snow, and ice

lightingenum · required

Ambient illumination category. Dawn and dusk are operationally distinct from day: low-angle sun and long shadows create sensor challenges absent in full daylight. overcast_day (~100 lx) differs from day (~10,000 lx) by reduced contrast and absence of shadows.

↳ ASAM OSI AmbientIllumination lux-level buckets

dayFull daylight (~10,000 lx)

dawnSunrise transition (~1–100 lx) — low-angle sun, long shadows

duskSunset transition (~1–100 lx) — low-angle sun, long shadows

nightDarkness — artificial or no illumination (<1 lx)

tunnelEnclosed road with artificial lighting — rapid illumination transition on entry/exit

overcast_dayDaylight under overcast cloud cover (~100 lx) — reduced contrast, no shadows

sensor_fidelityenum list

Active degradation modes affecting the camera sensor output. Zero or more values; duplicates not allowed. An empty array means no degradation detected.

↳ Domain-specific (Korean urban road coverage)

cleanNo degradation detected

lens_flareBright light source causing lens flare artifacts

droplets_on_lensWater droplets or condensation on the camera lens

motion_blurTemporal blur from camera or ego-vehicle motion

low_contrastReduced image contrast

overexposedOverexposure — blown-out highlight regions

Layer 2

Topology — Road Infrastructure

Road classification, lane configuration, and intersection geometry.

Source: OpenDRIVE v1.5M XSD

road_typeenum · required

Road classification by purpose and traffic characteristics. Based on OpenDRIVE e_roadType. 'highway' was renamed to 'motorway' for ASAM standard alignment; 'urban' was split into 'primary' and 'secondary' for finer granularity.

↳ OpenDRIVE v1.5M XSD e_roadType

motorwayControlled-access highway — limited entry/exit points (formerly 'highway')

trunkMajor national or inter-city road (high-speed, non-motorway)

primaryPrimary urban or inter-city road — controlled traffic flow

secondarySecondary urban road — moderate traffic density (formerly part of 'urban')

residentialResidential street — low speed, mixed pedestrian and vehicle use

serviceService road — parking lots, back alleys, delivery access roads

ruralRural or unclassified road outside urban areas

parkingParking lot or structured parking area

walkwayPedestrian walkway — no motor vehicles permitted

cyclingBicycle-dedicated path

lane_eventenum · required

Lane configuration or temporary alteration at the clip location.

↳ Domain-derived from traffic management conventions

normalStandard lane markings, no temporary changes

construction_divertConstruction zone with active lane diversion

lane_closedOne or more lanes closed to traffic

mergeLanes merging — two lanes converge into one

splitLanes splitting — one lane diverges into two

unmarkedLanes physically present but with no visible markings

intersection_typeenum · required

'direct_connection' specifically refers to highway on/off-ramp merge zones (OpenDRIVE direct junction) — not a generic road connection. 'crosswalk' marks a clip centred on a designated pedestrian crossing.

↳ OpenDRIVE v1.5M junction types

noneNo intersection — straight road section

signalizedTraffic-signal controlled intersection

unsignalizedUncontrolled intersection (yield or stop signs only)

roundaboutCircular intersection with yield-on-entry rule

t_junctionT-shaped three-way intersection

crosswalkDesignated pedestrian crossing (zebra crossing)

direct_connectionHighway on/off-ramp — direct merge zone (OpenDRIVE direct junction)

Layer 3

Actor Dynamics — Dynamic Actors

Per-actor classification, behavioral state, and proximity for each detected road user. Each clip may contain zero or more actor entries.

Source: ASAM OSI VehicleClassification · PedestrianCategory · MiscObjectCategory

actor_classenum · per actor

Object classification by type and role. Korean urban-specific classes (e_bike_rider, delivery_motorcycle, standup_scooter_rider) are retained for local road coverage. 'vehicle_sedan' / 'vehicle_suv' were consolidated into 'vehicle_car' per ASAM OSI TYPE_CAR.

↳ ASAM OSI VehicleClassification + PedestrianCategory + MiscObjectCategory

pedestrianFoot-walking person

cyclistBicycle rider

motorcyclistMotorcycle or motorbike rider

standup_scooter_riderE-scooter rider (standing position)

e_bike_riderElectric bicycle rider

delivery_motorcycleDelivery motorcycle or scooter (food/parcel courier)

wheelchair_userWheelchair occupant

vehicle_carPassenger car — sedan, SUV, hatchback (ASAM OSI TYPE_CAR)

vehicle_vanVan or minivan (passenger or light cargo)

vehicle_truckTruck — pickup truck or heavy goods vehicle

vehicle_busBus — public transit or coach

vehicle_emergencyEmergency vehicle — ambulance, police car, fire truck

vehicle_constructionConstruction machine — excavator, loader, road roller

animalAnimal on or near the roadway (dog, cat, livestock)

debrisFallen debris or foreign objects on the road

construction_objectConstruction equipment, barriers, or site signage

obstacleGeneric obstacle — rock, fallen branch, pothole

stateenum · per actor

Behavioral state or motion intent of the detected actor at clip time.

↳ Domain-derived from driving scenario analysis

crossingActively crossing the ego's path or travel lane

hesitatingSlowed or paused — uncertain or interrupted motion

jaywalkingCrossing the road without a designated crosswalk or signal

cutinLateral insertion into ego's lane from the side

cutoutLateral departure from ego's lane to the side

stoppedStopped — may resume motion (e.g. waiting at a signal)

emergingAppearing into the scene from occlusion or off-screen

tailingFollowing ego vehicle at close distance

oncomingApproaching ego head-on in the opposite travel direction

parkedParked — not moving, occupying a stationary position

staticImmobile object or permanently stationary actor

distance_bucketenum · per actor

Spatial proximity to the ego vehicle, binned into three tiers for threat prioritisation.

↳ Domain-derived (threat-zone bucketing)

near0 – 10 m — immediate threat zone

mid10 – 50 m — reaction window

far> 50 m — background context

Layer 4

Planner Logic — Ego Intent & Risk

Ego vehicle maneuver classification and ISO 21448 SOTIF risk level.

Source: WOD-E2E arXiv:2510.26125 · PEGASUS HAD-F · ISO 21448:2022 SOTIF

ego_maneuverenum · required

Primary driving maneuver executed or intended by the ego vehicle. 'emergency_brake' = AEB/panic stop (maximum deceleration, automated trigger) — distinct from 'brake_hard' (intentional high-g deceleration). 'swerve' = lateral avoidance that does not complete a full lane change.

↳ WOD-E2E arXiv:2510.26125 + PEGASUS HAD-F maneuver taxonomy

cruiseConstant speed, no active maneuver

accelerateIntentional speed increase

brake_softGentle, gradual deceleration

brake_hardIntentional hard deceleration (high g)

emergency_brakeAEB / panic stop — maximum deceleration, automated trigger

nudge_leftSmall lateral adjustment to the left without changing lanes

nudge_rightSmall lateral adjustment to the right without changing lanes

lane_change_leftComplete lane change to the left

lane_change_rightComplete lane change to the right

yieldSlowing or pausing to give priority to another actor

stopFull stop — vehicle at rest with no motion

reverseBackward motion

swerveLateral avoidance maneuver — does not complete a full lane change

risk_levelenum · required

Scenario risk classification per ISO 21448 SOTIF. Drives the Review Queue priority — 'critical' clips surface first. The DNA pass rate metric counts approved / (approved + rejected), excluding pending and schema_invalid.

↳ ISO 21448:2022 SOTIF

nominalNo safety concern — absence of unreasonable risk (normal operation)

elevatedTolerable risk — hazard present but mitigation is in place

criticalUnreasonable risk (SOTIF trigger) — intervention or override required

Review States

The four lifecycle states a clip occupies in the curation workflow.

pending

Awaiting human review. All newly ingested clips start in this state.

approved

Accepted into the curated corpus. DNA payload has been verified by a reviewer.

rejected

Manually rejected by a human reviewer. Excluded from the training corpus.

rejected_schema_invalid

Automatically rejected at ingestion time — the DNA payload failed JSON Schema validation. Displayed as 'Schema Invalid' in the UI.

Glossary

Key terms and concepts used throughout the platform.

Verify-by-Exception (VBE)

A curation strategy where clips pass through automatically unless flagged. Reviewers focus effort only on uncertain or high-risk cases, dramatically reducing manual load on nominal clips.

Scenario DNA

A 4-layer structured descriptor (ODD + Topology + Actor Dynamics + Planner Logic) attached to every clip. Stored as JSONB in PostgreSQL and indexed in Milvus. Schema version: v0.1.0 (frozen).

ODD (Operational Design Domain)

The specific conditions under which an AV system is designed to operate safely (ISO 22736). In My-Curator, ODD covers weather, lighting, and sensor fidelity.

Scout

The VLM that generates Scenario DNA from video frames. Current model: Cosmos-Reason2-8B FP8. Multiple Scout samples per clip are aggregated by BestOfN Aggregator using a symbolic reward signal.

Gold Set

A manually verified subset of clips (is_gold = true) used as ground truth for Recall@5 benchmark evaluation. Current: 14 clips, Recall@5 = 0.929.

Hybrid Search

Retrieval combining Milvus ANN vector search (Cosmos-Embed1-336p, 768-dim, cosine / inner product on L2-normalised vectors) with PostgreSQL JSONB GIN filter on DNA fields. ANN top-1000 candidates are re-ranked by exact filter matching.

DNA Pass Rate

Approved / (Approved + Rejected). Excludes pending and schema_invalid states. Shown on the Dashboard as a percentage.

SOTIF (ISO 21448)

Safety of the Intended Functionality — ISO standard defining risk categories for AV systems. My-Curator's risk_level enum maps directly: nominal → no unreasonable risk, elevated → tolerable risk, critical → unreasonable risk trigger.

dna_version

Schema version lock ('0.1.0'). Any schema change bumps this value and triggers a full prompt_regression + schema test run.

causal_trigger_actor_index

Index into actor_dynamics[] identifying which actor caused the ego maneuver. Reserved for post-v0.1 Judge model (Qwen2.5-14B-AWQ); null in current v0.1.0 single-Scout deployments.

grounded_by_yolo26

Boolean per actor. True if YOLO26 object detection independently confirmed the actor's presence, reducing hallucination risk for that actor entry.

hallucination_flags

Array of field-name strings in the confidence block flagging fields where the Scout may have fabricated values. Used to surface low-confidence DNA regions for reviewer attention.