← ML Research Wiki / 2506.17213

Long-term Traffic Simulation with Interleaved Autoregressive Motion and Scenario Generation

(2025)

Paper Information
arXiv ID

Abstract

Ego agentInitially placed agents Dynamically generated agents Figure 1.Long-term traffic simulation with InfGen and prior SOTA [31].InfGen keeps scene layout realistic while [31] becomes empty.

Summary

This paper presents InfGen, a long-term traffic simulator that integrates autoregressive motion and scenario generation in a unified framework. InfGen is designed to address the limitations of existing simulators that fail to maintain realism in long-term traffic simulations by dynamically generating agents based on the scene layout as the ego vehicle navigates. Using a tokenizer, the authors convert real-world driving logs into a sequence of discrete tokens that capture both motion and scenario generation, enabling stable long-term rollouts. Through extensive experimentation, the authors demonstrate that InfGen significantly outperforms state-of-the-art models in both motion and scenario realism, particularly in a 30-second rollout scenario, while also proving adaptable to shorter simulation tasks. The paper discusses various methodologies related to motion simulation and scenario generation, as well as the tokenization process involved in preparing data for the model. Finally, the authors acknowledge the limitations of their study, including challenges related to real trip-level evaluation and potential overfitting due to supervised training on historical data, and propose future directions for improving traffic simulation realism.

Methods

This paper employs the following methods:

  • Interleaved autoregressive token prediction
  • Closed-loop motion simulation
  • Scene generation

Models Used

  • InfGen

Datasets

The following datasets were used in this research:

  • Waymo Open Motion Dataset (WOMD)

Evaluation Metrics

  • Negative Log-Likelihood (NLL)
  • Agent Count Error (ACE)
  • WOSAC

Results

  • InfGen outperforms prior state-of-the-art models in long-term traffic simulation
  • Achieves significant improvements in both motion and scenario realism
  • Competitive performance in short-term simulation settings

Technical Requirements

  • Number of GPUs: 8
  • GPU Type: NVIDIA A5000
  • Compute Requirements: batch size of 8

Papers Using Similar Methods

External Resources