Language-Guided Traffic Simulation via Scene-Level Diffusion

  • 2023-06-10 06:20:30
  • Ziyuan Zhong, Davis Rempe, Yuxiao Chen, Boris Ivanovic, Yulong Cao, Danfei Xu, Marco Pavone, Baishakhi Ray
  • 0

Abstract

Realistic and controllable traffic simulation is a core capability that isnecessary to accelerate autonomous vehicle (AV) development. However, currentapproaches for controlling learning-based traffic models require significantdomain expertise and are difficult for practitioners to use. To remedy this, wepresent CTG++, a scene-level conditional diffusion model that can be guided bylanguage instructions. Developing this requires tackling two challenges: theneed for a realistic and controllable traffic model backbone, and an effectivemethod to interface with a traffic model using language. To address thesechallenges, we first propose a scene-level diffusion model equipped with aspatio-temporal transformer backbone, which generates realistic andcontrollable traffic. We then harness a large language model (LLM) to convert auser's query into a loss function, guiding the diffusion model towardsquery-compliant generation. Through comprehensive evaluation, we demonstratethe effectiveness of our proposed method in generating realistic,query-compliant traffic simulations.

 

Quick Read (beta)

loading the full paper ...