Multi-Vehicle Trajectories Generation for Vehicle-to-Vehicle Encounters

Wenhao Ding, Wenshuo Wang, Ding Zhao

Abstract

Generating multi-vehicle trajectories from existing limited data can provide rich resources for autonomous vehicle development and testing. This paper introduces a multi-vehicle trajectory generator (MTG) that can encode multi-vehicle interaction scenarios (called driving encounters) into an interpretable representation from which new driving encounter scenarios are generated by sampling. The MTG consists of a bi-directional encoder and a multi-branch decoder. A new disentanglement metric is then developed for model analyses and comparisons in terms of model robustness and the independence of the latent codes. Comparison of our proposed MTG with $\beta$-VAE and InfoGAN demonstrates that the MTG has stronger capability to purposely generate rational vehicle-to-vehicle encounters through operating the disentangled latent codes. Thus the MTG could provide more data for engineers and researchers to develop testing and evaluation scenarios for autonomous vehicles.

Introduction

The development and testing of autonomous vehicles (AVs) require a vast array of complex driving scenarios, particularly multi-vehicle interactions or "encounters." Real-world datasets often lack the necessary diversity and volume of these edge cases due to the high cost and logistical difficulty of data collection. While deep generative models offer a potential solution for creating synthetic data, standard approaches often struggle with multi-agent time-series data, frequently producing physically unrealistic trajectories or "entangled" latent spaces that lack controllability. This paper introduces the Multi-vehicle Trajectory Generator (MTG), a framework designed to encode multi-vehicle interaction scenarios into an interpretable representation for generating realistic synthetic encounters.

Methodology

The MTG frames trajectory generation as a sequence-to-sequence task. It utilizes a Bi-directional GRU Encoder to process multi-vehicle trajectories simultaneously, capturing both forward and backward temporal dependencies. The Multi-Branch Decoder is a key innovation; it splits into separate branches for each vehicle. These branches share hidden states to maintain contextual awareness of the interaction while generating independent coordinates, preventing interference between agent sequences.

The MTG architecture featuring a bi-directional GRU encoder and a multi-branch decoder.

Furthermore, the authors propose a novel Disentanglement Metric. Disentanglement ensures that individual variables in the latent space control distinct features (e.g., speed vs. approach angle). Unlike previous metrics that relied on sensitive linear classifiers, this new metric measures the variance of output latent codes when specific input dimensions are perturbed, providing a more stable evaluation of variable independence.

Results and Discussion

The model was trained on the UMTRI Dataset, comprising GPS data from approximately 3,500 vehicles. Evaluation focused on qualitative realism, traffic rationality, and disentanglement.

Comparison of real trajectories and generated results from InfoGAN, original VAE, and the proposed MTG.

Key findings include:

  • Qualitative Realism: MTG generated smooth, physically plausible trajectories, outperforming $\beta$-VAE, which sometimes produced impossible sharp turns.
  • Traffic Rationality: MTG's outputs closely matched real-world distributions for vehicle distance and speed variation. In contrast, InfoGAN suffered from significant mode collapse, producing very low diversity.
  • Superior Disentanglement: The proposed metric confirmed that MTG achieves a highly disentangled latent space with minimal dependency among latent codes. This allows engineers to precisely manipulate specific encounter features, such as the meeting point or approach direction, by adjusting single latent variables.

Conclusion

MTG provides a robust framework for generating diverse, realistic vehicle-to-vehicle encounters. While current limitations include relatively short sequence lengths and imperfect encoder-decoder variance ratios, future work will focus on incorporating road profiles and map data. By cascading short trajectories, the researchers aim to generate sustained, complex traffic scenarios to further enhance AV training and evaluation.

Summary updated

Abstract

Generating multi-vehicle trajectories from existing limited data can provide rich resources for autonomous vehicle development and testing. This paper introduces a multi-vehicle trajectory generator (MTG) that can encode multi-vehicle interaction scenarios (called driving encounters) into an interpretable representation from which new driving encounter scenarios are generated by sampling. The MTG consists of a bi-directional encoder and a multi-branch decoder. A new disentanglement metric is then developed for model analyses and comparisons in terms of model robustness and the independence of the latent codes. Comparison of our proposed MTG with $\beta$-VAE and InfoGAN demonstrates that the MTG has stronger capability to purposely generate rational vehicle-to-vehicle encounters through operating the disentangled latent codes. Thus the MTG could provide more data for engineers and researchers to develop testing and evaluation scenarios for autonomous vehicles.
BibTeX
@inproceedings{ding2019multi,
  author = {Wenhao Ding and Wenshuo Wang and Ding Zhao},
  title = {{Multi-Vehicle Trajectories Generation for Vehicle-to-Vehicle Encounters}},
  booktitle = {IEEE International Conference on Robotics and Automation},
  year = {2019},
  url = {https://openreview.net/forum?id=I1itg2_rBcm},
}
Figure