Open-Travel: Open Mobility Research

• Open-Travel is an ecosystem of standardized open datasets, algorithms, and evaluation protocols designed to advance research in travel behavior and multimodal routing.
• Benchmarking studies using Open-Travel demonstrate robust performance in travel mode detection, achieving F1-scores up to 0.96 with feature-augmented models.
• Its extensible open-source infrastructure supports reproducible methodologies for demand modeling, adaptive itinerary generation, and real-time crisis-response routing.

Open-Travel denotes the combined set of open methodologies, datasets, algorithmic frameworks, and evaluation protocols that underpin data-driven research and public innovation across travel behavior, mobility planning, multi-modal journey routing, personalized itinerary synthesis, agent-based demand generation, and collaborative spatiotemporal urban modeling. The term is used in recent literature both to describe specific open datasets—such as labeled GPS trajectory corpora for travel mode detection—and, more broadly, to encompass fully open-source pipelines for benchmarking, demand modeling, real-time planning, and participatory urban experience. Open-Travel systems and resources are designed to be extensible, reproducible, and interoperable, enabling academic and applied researchers to standardize, compare, and advance algorithms for travel demand understanding, multimodal routing, behavioral simulation, and adaptive planning.

1. Foundational Open Datasets and Benchmarks

A core motivation for Open-Travel initiatives is the absence of standardized, labeled, and publicly accessible datasets to support rigorous benchmarking of travel-behavior algorithms. The "Open-Travel" GPS trajectory dataset (Chen et al., 2021) was established in direct response to the widespread reliance on idiosyncratic, researcher-collected mobility traces, which impede direct comparison of detection algorithms and fair assessment of generalizability. This corpus comprises:

• Volunteer-based Trajectory Collection: Data from 7 independent volunteers in the Greater Tokyo and Chiba area, recorded via Android smartphones at 1 Hz for an entire month.
• Mode-Labeled Segments: Every trajectory annotated with ground-truth travel mode (walking, bicycle, bus, railway), amounting to 475 trips and ~4,500 km across mixed urban and rural settings.
• Rich Metadata and Full Access: Trips segmented, labeled, organized in standardized CSVs, with a master metadata index and GeoJSON overlays for geospatial tool compatibility; available under Creative Commons Attribution 4.0 International.

This dataset supports algorithmic benchmarking for travel mode detection by supplying a diverse, mode-balanced set of ground-truth trajectories, with observed variations in congestion, time-of-day, and repeated routes.

2. Algorithmic Approaches and Feature Engineering

Open-Travel benchmarking foregrounds transparent algorithmic pipelines and explicit, mathematically formulated feature extraction protocols. In the travel mode detection task (Chen et al., 2021), trajectory segments are transformed into feature vectors comprising:

• Point-wise Dynamics:
  • Instantaneous speed:

  $v_i = \frac{d(p_i, p_{i-1})}{t_i - t_{i-1}}$ - Acceleration:

  $a_i = \frac{v_i - v_{i-1}}{t_i - t_{i-1}}$ - Heading-change rate:

  $$\omega_i = \frac{|\theta_i - \theta_{i-1}|}{t_i - t_{i-1}}$$

• Segment-level Aggregates:

  • Total distance, duration, number of points
  • Velocity change rate (average and max), maximum acceleration
  • Time- and point-averaged speeds, extrema

Reference benchmarks (Random Forest, 100 trees, stratified 80/20 splits, 10-fold cross-validation) yield per-mode F1-scores of 0.92–0.96 (walking, biking, bus, railway), with overall accuracy from 89.3% to 100%. Main confusions occur between bus and railway on short-haul trips.

A specific challenge is addressed in distinguishing between walking and biking under sparse (5-min subsampled) data, where classic thresholding fails; the feature-augmented Random Forest recovers ~95% accuracy.

3. Extensible Open-Source Infrastructures and Reproducibility

Open-Travel encompasses not only data provision but full-stack, modular benchmarking platforms (Caicedo et al., 2023), exemplified by collaborative infrastructure for transit demand prediction. Key features include:

• Unified Data Ingestion: Smart-card ridership logs, calendar/weather/event features; raw→processed pipelines.
• Multi-Model Evaluation: ARIMA, SARIMA, ARIMA–GARCH, Gradient Boosted Decision Trees, and LSTM (sequence-to-one, rolling retrain) standardized under a common training/testing API.
• Stable/Dynamic Scenarios: Explicit train/test splits enable systematic evaluation under both stationary periods and shocks (e.g., protest, COVID-19).
• Performance Metrics:
  • MAAPE ($$\frac{1}{N}\sum_{i=1}^N \arctan \left| \frac{\hat y_i - y_i}{y_i} \right|$$), minimizing issues of infinite percentage error.
  • LSTM exhibits rapid error stabilization (MAAPE ~0.12) in dynamic windows where classical models lag.
• Reproducibility Protocols:
  • Locked environments (conda, Docker)
  • Data versioning (raw/processed separation)
  • CI/CD for standardized ingest/training/benchmarking
  • Open contribution guides, pull requests, and model/dataset plug-in patterns

This infrastructure allows rapid scaling to new demand domains (bike-share, ride-hail, intercity) by modularizing schema, covariate, and target definitions.

4. Open-Travel Planning Agents and Personalization

Recent Open-Travel research extends beyond passive analysis, providing end-to-end, adaptive itinerary generation systems and language-powered agent coordination:

• Vaiage System (Liu et al., 16 May 2025):
  • Graph-Structured Multi-Agent Architecture: Agents (Chat, Information, Recommender, Planner, Strategy, Communication) interact on a formal directed graph, consume structured JSON context, and coordinate via state transitions.
  • LLM-Based Reasoning: LLMs infer user intent, map natural language to actionable planning context, and effect sequential resource/tool orchestration.
  • Tool Integration: Real-time calls to mapping, weather, POI, and transport APIs trigger dynamic replanning on external event changes (e.g., rain shifting outdoor activities).
  • Itinerary Optimization: Combinatorial assignment of POIs per day, constrained by temporal, cost, and diversity requirements, solved via greedy insertion, local refinement, and dynamic repair, with LLM-based explainable traces (“reasoning” nodes in the graph).
  • Evaluation and Extensibility: Human-in-the-loop (GPT-4) rubric scoring (relevance, feasibility, personalization, satisfaction), modular plug-in agents (events, dining, budgets), and distributed deployment for scalability.
• ITINERA System (Tang et al., 2024):
  • Hybrid LLM + Spatial Optimization:
  • LLMs decompose unconstrained user requests, retrieve/score POIs, encode soft/hard constraints, and ultimately synthesize legible itineraries.
  • Cluster-aware spatial optimization: Proximity graph clustering, hierarchical TSP on clusters and subclusters, explicit ordering for minimal travel.
  • Comparative Performance: Significant gains on recall, spatial compactness, and request-alignment metrics over GPT-4 Chain-of-Thought and black-box neural baselines.
  • Open-domain, dynamic POI curation and group itineraries: Personalization accommodates niche or emergent requirements.

5. Urban, National, and Synthetic Open Mobility Models

Open-Travel further incorporates wide-area synthetic demand generation and open urban/city-scale modeling, integrating statistical, survey-based, or OSM-sourced data:

• OMOD (Strobel et al., 2023):
  • Agent-based Activity Models: Fully disaggregated schedules at the individual-building scale, using OSM for footprints, POIs, and road networks.
  • Synthetic Population and Schedule Generation: Socio-demographics sampled from real or default distributions; activity-chains and dwell-times from survey-fitted Gaussian mixtures.
  • Destination Assignment: Multinomial logit with attraction, deterrence (distance) terms, stratified by empirical accessibility distributions.
  • Validated at Urban Scale: Trip attraction R² up to 0.95 (5 km), OD matrix R² up to 0.96, mean daily distance error <5% using only open data.
• Pseudo-PFLOW (Kashiyama et al., 2022):
  • Nationwide Synthetic Flow: 130 million agents, activities sequenced by role/time Markov chain, allocations using census ODs, open geospatial data, and modal splits.
  • Validation Metrics: Population distributions, trip counts and purpose-matched volumes compared to mobile phone and survey data, with R² ≥0.5 at sub-kilometer resolution.
  • Comprehensive Schema: Origin/destination at building/grid scales, per-minute trajectories, aggregated OD-matrices.
  • Research Access: The data is made available under a controlled, academic-friendly license.

6. Open-Source Multimodal Routing and Crisis Applications

OpenTripPlanner (OTP) and associated open standards (OSM, GTFS) (Narboneta et al., 2016) exemplify open-source platforms for real-time, multi-modal trip planning, with extensibility for disaster preparedness and recovery scenarios:

• Data Integration:
  • OSM for detailed street networks
  • GTFS for transit schedules, frequencies, agency metadata
• Routing Algorithms:
  • Time-dependent Dijkstra/A* on multimodal graph (streets, transit, walk/bike connectors)
  • Cost functions blending in-vehicle time, walk distance, number of transfers, and mode penalties
  • Pareto-optimal multivariate label-setting for multi-criteria optimization
• Disaster-Use Modifications:
  • Dynamic graph edits for road closures/obstructions
  • Capacity constraints on edges for relief vehicles
  • Batch shelter accessibility analysis and real-time evacuation routing
• Deployment Considerations:
  • Nightly ingestion of OSM and GTFS deltas
  • Clustered deployments, precomputed route caching, geospatial pruning

7. Future Directions and Open Challenges

Ongoing development and extension of Open-Travel approaches emphasize:

• Expansion of Labeled Data: Inclusion of additional transport modes (private car, e-scooter), and broader regional coverage (Chen et al., 2021).
• Sensor Fusion and Data Modality Integration: Augmentation with inertial-sensor data for enhanced robustness (e.g., in low-GPS scenarios), linking to anonymized telecom or crowd-sourced sources.
• Advanced Personalization and Constraints: Handling of multi-user requests, temporal dependencies (opening hours, events), and compositional user requirements, e.g., via DSLs as pioneered in ChinaTravel (Shao et al., 2024).
• Neuro-symbolic Agent Frameworks: Combining LLM-based free-form reasoning with explicit logical constraint grounding, measured against compositional and user-specified benchmarks.
• Community-driven Model and Tool Development: Standardized contribution protocols, robust documentation, CI, and open governance for research and operational deployments (Caicedo et al., 2023).

A plausible implication is that Open-Travel, as a research and innovation paradigm, is best characterized by its focus on verifiable, extensible, and reproducible infrastructure, which is necessary for robust mobility analysis, benchmarking, and algorithmic advancement in heterogeneous, real-world settings. This approach ensures open, scalable, and trustworthy development of mobility data science, supporting both academic inquiry and applied deployment.