# Language-Guided Traffic Simulation via Scene-Level Diffusion

**Authors:** Ziyuan Zhong, Davis Rempe, Yuxiao Chen, Boris Ivanovic, Yulong Cao, Danfei Xu, Marco Pavone, Baishakhi Ray  
**arXiv:** [2306.06344](https://arxiv.org/abs/2306.06344)  
**Date:** 2023-06-10  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2306.06344v1/

## Key points

- Realistic and controllable traffic simulation is crucial for the development of autonomous vehicles (AVs)
- The authors propose CTG++, a scene-level conditional diffusion model guided by language instructions
- CTG++ addresses challenges in controlling learning-based traffic models
- It generates realistic and controllable traffic simulations using a spatio-temporal transformer backbone and a large language model (LLM)
- User queries are converted into loss functions to generate query-compliant simulations
- Comprehensive evaluation demonstrates the effectiveness of CTG++ in generating query-compliant simulations
- CTG++ provides an accessible tool for practitioners to simulate various scenarios and test autonomous systems' responses

## Summary

Realistic and controllable traffic simulation is crucial for the development of autonomous vehicles (AVs). The authors propose CTG++, a scene-level conditional diffusion model guided by language instructions to address challenges in controlling learning-based traffic models. This approach generates realistic and controllable traffic simulations using a spatio-temporal transformer backbone and a large language model (LLM) to convert user queries into loss functions. Through comprehensive evaluation, the authors demonstrate its effectiveness in generating query-compliant simulations. CTG++ offers a valuable contribution to AV development by providing an accessible tool for practitioners to simulate various scenarios and test autonomous systems' responses.