Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

SocialDriveGen: Generating Diverse Traffic Scenarios with Controllable Social Interactions

Jiaguo Tian, Zhengbang Zhu, Shenyu Zhang, Li Xu, Bo Zheng, Xu Liu, Weiji Peng, Shizeng Yao, Weinan Zhang

TL;DR

SocialDriveGen tackles the challenge of generating diverse, socially realistic traffic scenarios for autonomous driving validation. It couples semantic reasoning via Vision-Language Models and reward construction via Large Language Models with a two-dimensional social psychology framework (egoism and altruism) and a gradient-free, step-wise diffusion process for joint multi-agent trajectory synthesis. The approach enables controllable variation in driver personalities and interaction styles, producing scenarios from cooperative to adversarial while maintaining high fidelity. On Argoverse 2, it improves scenario diversity and policy robustness by populating rare, safety-critical interactions that standard simulators miss.

Abstract

The generation of realistic and diverse traffic scenarios in simulation is essential for developing and evaluating autonomous driving systems. However, most simulation frameworks rely on rule-based or simplified models for scene generation, which lack the fidelity and diversity needed to represent real-world driving. While recent advances in generative modeling produce more realistic and context-aware traffic interactions, they often overlook how social preferences influence driving behavior. SocialDriveGen addresses this gap through a hierarchical framework that integrates semantic reasoning and social preference modeling with generative trajectory synthesis. By modeling egoism and altruism as complementary social dimensions, our framework enables controllable diversity in driver personalities and interaction styles. Experiments on the Argoverse 2 dataset show that SocialDriveGen generates diverse, high-fidelity traffic scenarios spanning cooperative to adversarial behaviors, significantly enhancing policy robustness and generalization to rare or high-risk situations.

SocialDriveGen: Generating Diverse Traffic Scenarios with Controllable Social Interactions

TL;DR

SocialDriveGen tackles the challenge of generating diverse, socially realistic traffic scenarios for autonomous driving validation. It couples semantic reasoning via Vision-Language Models and reward construction via Large Language Models with a two-dimensional social psychology framework (egoism and altruism) and a gradient-free, step-wise diffusion process for joint multi-agent trajectory synthesis. The approach enables controllable variation in driver personalities and interaction styles, producing scenarios from cooperative to adversarial while maintaining high fidelity. On Argoverse 2, it improves scenario diversity and policy robustness by populating rare, safety-critical interactions that standard simulators miss.

Abstract

The generation of realistic and diverse traffic scenarios in simulation is essential for developing and evaluating autonomous driving systems. However, most simulation frameworks rely on rule-based or simplified models for scene generation, which lack the fidelity and diversity needed to represent real-world driving. While recent advances in generative modeling produce more realistic and context-aware traffic interactions, they often overlook how social preferences influence driving behavior. SocialDriveGen addresses this gap through a hierarchical framework that integrates semantic reasoning and social preference modeling with generative trajectory synthesis. By modeling egoism and altruism as complementary social dimensions, our framework enables controllable diversity in driver personalities and interaction styles. Experiments on the Argoverse 2 dataset show that SocialDriveGen generates diverse, high-fidelity traffic scenarios spanning cooperative to adversarial behaviors, significantly enhancing policy robustness and generalization to rare or high-risk situations.

Paper Structure

This paper contains 17 sections, 3 equations, 2 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Method
  3. Scenario Analysis and Proposal Generation with VLM
  4. Social-aware Reward Formulation
  5. Decomposition of Social Behavior
  6. The Integrated Reward Function
  7. Multi-vehicle Trajectory Generation
  8. Denoising Process with Step-wise Guidance
  9. Joint Multi-Agent Trajectory Generation
  10. Experiment
  11. Qualitative Analysis: Interactive Scenario Generation
  12. Effectiveness of the Two-Stage VLM Planner
  13. Effectiveness of Iterative Guidance
  14. Controllability via Social Preferences
  15. Impact of Social Orientation ($\phi$)
  16. ...and 2 more sections

Figures (2)

  • Figure 1: Overview of SocialDriveGen. The framework leverages VLMs and LLMs to construct social rewards that guide a multi-agent diffusion process.
  • Figure 2: Results of interactive scenarios generated by SocialDriveGen. Our framework translates high-level VLM proposals into diverse, social interactive trajectories. Red vehicles correspond to the pair selected by the VLM for interaction.