Frozen LLMs as Map-Aware Spatio-Temporal Reasoners for Vehicle Trajectory Prediction

Installation

Environment Setup

First, we'll create a conda environment to hold the dependencies.

conda env create -f env.yaml
conda activate <environment_name>

Datasets

/process_data/preprocess_first_run.py 
This script preprocesses the nuScenes dataset to prepare it for our model. 
The data will be processed into .pkl and .index files, enabling lazy loading for efficient training and testing.

Train

The model is trained on four 48GB GPUs, with each training session taking approximately 10 hours. We use sbatch to submit the job.
The training script can be found at: 
trajectory_prediction/train_and_test_result/trajectory_predict_train.py

Test

The testing script is located at: trajectory_prediction/train_and_test_result/test_metrics.py

Project Title

Introduction

This project aims to develop a robust method for vehicle trajectory prediction using frozen LLMs (Large Language Models) integrated with map-aware spatio-temporal reasoning.

Citations

Papers

• Yanjiao Liu, Jiawei Liu, Xun Gong, Zifei Nie. "Frozen LLMs as Map-Aware Spatio-Temporal Reasoners for Vehicle Trajectory Prediction." Proceedings of the 2026 IEEE Intelligent Vehicles Symposium (IV), Manuscript 216, 2026.

If you use this project or parts of it in your work, please cite the paper as follows:

@inproceedings{liu2026frozen,
  author    = {Yanjiao Liu and Jiawei Liu and Xun Gong and Zifei Nie},
  title     = {Frozen LLMs as Map-Aware Spatio-Temporal Reasoners for Vehicle Trajectory Prediction},
  booktitle = {Proceedings of the 2026 IEEE Intelligent Vehicles Symposium (IV)},
  year      = {2026},
  manuscript = {216}
}