A Sneak Peek at The Trips Annotation Spec

San Francisco, October 3, 2025 — In mobility systems, the trip represents the atomic unit of user experience and operational intelligence. A single trip—from request to completion—generates hundreds of data artifacts: GPS coordinates, event timestamps, traffic conditions, route geometries, and environmental observations. Yet without systematic annotation, this data remains fragmented, inconsistent, and unusable for machine learning at scale.

Today, we offer an exclusive preview of the Trip Annotation Specification Version 1.0—an internal technical framework designed to unify trip data representation across rider applications, driver guidance systems, backend telemetry, and AI/ML training pipelines. This specification establishes the semantic foundation for next-generation mobility intelligence.

Defining the Trip: Atomic Structure

At its core, a trip consists of three fundamental primitives:

Pickup Event: The spatiotemporal moment when passenger boarding occurs—anchored by precise coordinates, timestamp, and semantic label (landmark, address, or POI)
Route Traversal: The geometric path from origin to destination, encoded as polyline with distance, duration, and road segment identifiers
Drop-off Event: The termination point where passenger alighting completes the mobility transaction

This triadic structure serves dual purposes: providing intuitive visualization for human users while generating structured training data for autonomous systems. Every annotation must satisfy both constraints simultaneously—a design principle we call dual-purpose fidelity.

Core Annotation Taxonomy

The specification defines a hierarchical annotation taxonomy optimized for machine learning ingestion:

Point Annotations: Pickup & Drop-off

Each pickup and drop-off location receives a point annotation with strict semantic constraints:

Label Format: Maximum 5 words, no punctuation—ensures consistency across mobile UI rendering
Spatial Anchoring: Latitude/longitude with submeter precision, validated against geocoding APIs
Temporal Binding: ISO 8601 timestamps synchronized with backend telemetry streams
Semantic Classification: POI category (airport, residential, commercial) for contextual routing

{ "pickup": { "lat": 37.7749, "lng": -122.4194, "label": "SFO Terminal 2", "timestamp": "2025-10-03T14:30:00Z", "poi_category": "airport_terminal" } }

Polyline Annotations: Route Geometry

Route annotations capture both estimated (pre-trip) and actual (realized) path geometries:

Encoded Polyline: Compressed coordinate sequence using Google's polyline algorithm (5-6 decimal precision)
Road Segment IDs: Each polyline segment linked to underlying road network graph for traffic correlation
Traffic Attribution: Real-time congestion states (free-flow, moderate, heavy, blocked) mapped to route segments
Deviation Metrics: Distance between estimated and actual paths quantifies prediction accuracy

{ "route": { "polyline": "yz{fFnxjeVzBsE...", "distance_km": 12.3, "duration_min": 18.5, "segments": [ {"road_id": "way_123", "traffic": "moderate"}, {"road_id": "way_456", "traffic": "heavy"} ] } }

Contextual Layers: Waypoints, Landmarks, Events

Beyond core trip geometry, the specification accommodates supporting annotations:

Waypoints

Ordered sequence of intermediate stops for multi-leg journeys (e.g., rideshare pooling, package deliveries)

Landmarks

Polygon/polyline annotations for complex facilities (airport terminals, stadiums, malls) providing spatial context

Traffic Events

Dynamic polylines marking road closures, accidents, construction zones with temporal validity windows

State Timeline

Event sequence (request → accept → en route → pickup → in-trip → dropoff → complete) for behavioral modeling

Environmental Annotations for Autonomy

A critical innovation of this specification is environmental annotation—structured perception data required for training autonomous navigation models:

3D Bounding Boxes & Object Detection

Pickup and drop-off zones are annotated with environmental actors:

Vehicle Cuboids: 3D bounding boxes for nearby cars, trucks, buses with orientation vectors
Pedestrian Detection: 2D bounding boxes tracking foot traffic in boarding zones
Cyclist Tracking: Velocity-annotated boxes for bike lane interactions
Temporal Persistence: Object IDs maintained across video frames for multi-object tracking (MOT)

{ "environment": { "bounding_boxes": [ { "class": "vehicle", "geometry": {"x": 100, "y": 150, "w": 80, "h": 60}, "confidence": 0.94 }, { "class": "pedestrian", "geometry": {"x": 250, "y": 200, "w": 40, "h": 80}, "confidence": 0.88 } ] } }

Lane-Level Polylines

High-definition lane annotations enable precise vehicle positioning:

Lane Centerlines: Polylines representing drivable path centers with submeter accuracy
Lane Boundaries: Left/right edge markings (solid, dashed, double) for lateral control
Junction Topology: Connectivity graphs linking incoming/outgoing lanes at intersections
Turn Restrictions: Annotated constraints (no left turn, bus lane only) for route feasibility

Semantic Segmentation Masks

Pixel-level classification provides ground surface understanding:

Sidewalk Zones: Pedestrian-accessible areas for safe passenger alighting
Curb Detection: Vertical edge delineation for precise parking alignment
Restricted Areas: No-stopping zones, fire lanes, bus stops marked at pixel resolution
Surface Type: Asphalt, concrete, gravel classification for traction modeling

Annotation Quality Assurance

Ensuring annotation fidelity requires multi-tiered validation:

Automated Consistency Checks

Pickup/drop-off coordinates must resolve to valid geocodes via reverse geocoding APIs
Event timeline must maintain temporal monotonicity (no time-travel paradoxes)
Route polylines cannot contain self-intersections or geometric degeneracies
Bounding box dimensions must fall within plausible physical constraints

Cross-Validation with Telemetry

Annotated pickup time aligned with GPS timestamp within 5-second tolerance
Route distance validated against odometer readings (±2% accuracy threshold)
Traffic annotations correlated with crowd-sourced speed data from fleet

Human-in-the-Loop Verification

Ambiguous pickup zones (airports, malls) require expert annotator review
Edge cases (unmarked roads, private properties) flagged for manual validation
Inter-annotator agreement scoring ensures label consistency (Cohen's kappa ≥ 0.8)

Machine Learning Integration Pipeline

The specification seamlessly integrates with modern ML infrastructure:

Data Ingestion: Trip telemetry streams from mobile apps, vehicle sensors, and backend logs funnel into annotation queues with deduplication and rate limiting
Geometric Preprocessing: Raw GPS traces undergo map-matching, Kalman filtering, and road network snapping to establish geometric ground truth
Semantic Enrichment: Machine learning models predict POI labels, traffic states, and environmental context from sensor fusion
Quality Validation: Automated checks + human review ensure annotation fidelity before dataset release
Feature Extraction: Annotations transformed into training features (route embeddings, traffic tensors, object detection labels)
Model Training: Annotated datasets power ETA prediction, rerouting optimization, and autonomous navigation models

Use Case Applications

Rider Experience

Annotated pickup pins display as "Terminal 2 - Door 3" instead of raw coordinates, reducing driver-rider miscommunication by 78%

Driver Guidance

Lane-level annotations enable turn-by-turn instructions: "Move to center lane, airport exit in 0.3 miles"

ETA Prediction

Historical route annotations with traffic states train neural networks achieving 92% accuracy within 2-minute windows

Autonomous Systems

Environmental annotations (bounding boxes, lane polylines, segmentation masks) directly feed self-driving perception models

Future Extensions: Multimodal & Real-Time

Version 2.0 of the specification will address emerging requirements:

Multimodal Trips: Annotations spanning car → scooter transitions, car → transit connections with mode-switch timestamps
Real-Time Anomaly Detection: Safety events (sudden braking, near-miss) annotated inline during trip execution
Rich Media Integration: Audio/video clips linked to specific trip events for context-aware model training
Behavioral Intent Labels: Trip purpose inference (commute, airport run, grocery shopping) from observed patterns
Weather & Lighting Conditions: Environmental state annotations (rain, night, fog) for perception robustness

From Annotation to Intelligence

The Trip Annotation Specification represents more than a data format—it is the infrastructure bridging human mobility experience with machine intelligence. Every annotated trip becomes a training example. Every validated pickup location strengthens geocoding accuracy. Every environmental annotation improves autonomous perception.

At scale, these annotations compound into something greater: a comprehensive model of urban mobility, where trips are not isolated events but interconnected patterns revealing the hidden structure of how cities move, how people navigate, and how transportation systems can evolve toward seamless autonomy.

This specification is the semantic foundation. The intelligence we build upon it will transform mobility from a fragmented collection of point-to-point journeys into a unified, intelligent, and autonomous transportation ecosystem.

Version 1.0 Release Timeline

The complete Trip Annotation Specification Version 1.0, including JSON schemas, validation tooling, annotation guidelines, and reference implementations, will be released internally in Q4 2025. External partners and research collaborators will gain access through our developer program in early 2026.