Trajectory generative models: a survey from unconditional and conditional perspectives

Weiping Zhu, Renhe Jiang, Dongyuan Li, Xiangjie Kong, Xuan SongGeoInformatica 2025 · OpenReview

Abstract

Trajectories serve as a cornerstone of intelligent transportation systems, playing an important role in many applications such as traffic flow prediction, route planning, and urban management. However, the availability of such data is limited due to privacy issues, ethical concerns, and the high cost associated with infrastructure deployment. In recent years, rapidly developing generative models such as Variational Autoencoders (VAEs), Generative Adversarial Networks (GANs), and Diffusion Models (DMs) have demonstrated strong capabilities in learning complex data distributions and generating synthetic data, thereby alleviating the data accessibility issue. In this survey, we systematically review the existing literature on deep generative models that address the problem of trajectory generation. First, we classify the existing literature into two categories: unconditional and conditional trajectory generation. In unconditional generation, trajectories are generated without contextual constraints, whereas conditional generation incorporates several important factors such as road network topology, time of day, and user preferences to guide the trajectory generation process. Then, for each category, we further classify the literature into three methodological types, including VAEs, GANs, and DMs, and analyze how these models address key challenges under different settings. Finally, we discuss promising directions for future research and hope to inspire further advances in trajectory generation.

Introduction

Trajectory data, comprising sequences of spatial coordinates and timestamps, is vital for urban planning, traffic forecasting, and location-based services. However, the acquisition of high-quality trajectory data is often hampered by privacy regulations (e.g., GDPR), ethical concerns, and the prohibitive cost of sensor infrastructure. To address these challenges, deep generative models have emerged as a powerful tool for synthesizing realistic trajectory data. This survey systematically reviews the landscape of trajectory generation, categorizing research into unconditional and conditional paradigms.

The Challenge of Trajectory Data

Real-world trajectories are more than just a list of coordinates; they encapsulate complex human behaviors and are strictly constrained by physical environments like road networks. Traditional data collection is prone to sparsity and noise. Moreover, the sensitive nature of mobility data—revealing home and work locations—makes sharing raw datasets difficult. Generative models offer a solution by creating synthetic datasets that preserve the statistical properties of the original data without exposing individual identities.

Taxonomy of Generation Paradigms

The survey classifies existing literature into two primary categories based on the presence of contextual constraints:

Unconditional Trajectory Generation

In this paradigm, models learn the underlying distribution of trajectories without external guidance. The goal is to capture the inherent spatiotemporal patterns of movement. While useful for general mobility modeling, these models often struggle with physical realism; without constraints, they may generate paths that are statistically sound but physically impossible, such as vehicles driving through buildings or across water.

Conditional Trajectory Generation

Conditional models incorporate auxiliary information to guide the generation process, ensuring the synthetic data adheres to real-world constraints. These conditions include:

  • Road Network Topology: Ensuring vehicles stay on navigable paths.
  • Temporal Factors: Accounting for peak hours versus night-time traffic patterns.
  • User Preferences: Modeling individual habits or specific destination intents.

Methodological Frameworks

The survey analyzes three dominant deep learning architectures currently used in the field:

  1. Variational Autoencoders (VAEs): VAEs are utilized to map high-dimensional trajectories into a continuous, low-dimensional latent space. This allows for efficient reconstruction and the generation of new samples by sampling from the learned latent distribution.
  2. Generative Adversarial Networks (GANs): GANs employ a minimax game between a generator and a discriminator. A notable sub-field is Generative Adversarial Imitation Learning (GAIL), which treats trajectory generation as a reinforcement learning problem where the model learns the "policy" or strategy of human routing behavior.
  3. Diffusion Models (DMs): Representing the newest frontier, DMs generate data by reversing a gradual noise-addition process. Recent adaptations for trajectories use graph-based encoders and geographic attention mechanisms to handle the unique sequential and spatial dependencies of mobility data.

Datasets and Evaluation

The field relies on several benchmark datasets, categorized into continuous GPS trajectories (e.g., GeoLife, Porto Taxi, T-Drive, DiDi-Chengdu) and discrete check-in sequences from Location-Based Social Networks (e.g., Gowalla, Brightkite, Foursquare).

Evaluation remains a significant challenge. Since synthetic trajectories lack a direct ground truth, researchers use distribution-level metrics such as Jensen–Shannon Divergence (JSD) and Kullback–Leibler Divergence (DKL) to measure the similarity between real and generated data distributions. Successful synthesis is often validated by showing that downstream models, such as traffic predictors, perform similarly when trained on synthetic versus real data.

Insights and Future Research

The survey highlights that conditional diffusion models currently provide the highest fidelity and physical realism. However, several gaps remain for future exploration:

  • Semantic Reasoning: Leveraging Large Language Models (LLMs) to inject commonsense reasoning and dynamic "intent" into generating agents.
  • Edge Case Generation: Improving the synthesis of rare events like accidents or emergency detours.
  • Multi-modal Fusion: Developing unified frameworks that can process heterogeneous inputs (graphs, text, and coordinates) without losing resolution.
  • Generalization: Creating universal spatiotemporal representations that can adapt to various tasks via transfer learning.

By synthesizing these developments, this survey provides a roadmap for future research in creating more robust, realistic, and privacy-preserving trajectory models.

Summary updated

Abstract

Trajectories serve as a cornerstone of intelligent transportation systems, playing an important role in many applications such as traffic flow prediction, route planning, and urban management. However, the availability of such data is limited due to privacy issues, ethical concerns, and the high cost associated with infrastructure deployment. In recent years, rapidly developing generative models such as Variational Autoencoders (VAEs), Generative Adversarial Networks (GANs), and Diffusion Models (DMs) have demonstrated strong capabilities in learning complex data distributions and generating synthetic data, thereby alleviating the data accessibility issue. In this survey, we systematically review the existing literature on deep generative models that address the problem of trajectory generation. First, we classify the existing literature into two categories: unconditional and conditional trajectory generation. In unconditional generation, trajectories are generated without contextual constraints, whereas conditional generation incorporates several important factors such as road network topology, time of day, and user preferences to guide the trajectory generation process. Then, for each category, we further classify the literature into three methodological types, including VAEs, GANs, and DMs, and analyze how these models address key challenges under different settings. Finally, we discuss promising directions for future research and hope to inspire further advances in trajectory generation.
BibTeX
@inproceedings{zhu2025trajectory,
  author = {Weiping Zhu and Renhe Jiang and Dongyuan Li and Xiangjie Kong and Xuan Song},
  title = {{Trajectory generative models: a survey from unconditional and conditional perspectives}},
  booktitle = {GeoInformatica},
  year = {2025},
  url = {https://openreview.net/forum?id=SBW6xVbG3z},
}
Figure