PhysORD: A Neuro-Symbolic Approach for Physics-infused Motion Prediction in Off-road Driving

This repository provides the official implementation for paper:

Zhipeng Zhao, Bowen Li, Yi Du, Taimeng Fu, and Chen Wang, "PhysORD: A Neuro-Symbolic Approach for Physics-infused Motion Prediction in Off-road Driving." IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2024

Prerequisites

• Python 3.10.13
• PyTorch 2.0.1
• PyPose 0.6.7

Dataset

• This project utilizes the TartanDrive dataset. Follow the instructions in its repository to create the train, test-easy and test-hard sets.
• test-easy is used for validation during training, and test-hard for model evaluation.
• We also provide pre-processed data with 20-step and 5-step sequences for quick reproduction. You can download them into the data folder.

# 20-step
wget -P data/ https://github.com/sair-lab/PhysORD/releases/download/data/train_val_easy_507_step20.pt

# 5-step
wget -P data/ https://github.com/sair-lab/PhysORD/releases/download/data/train_val_easy_507_step5.pt

Reproduce Guide

To reproduce our result in the paper, you can follow the the steps below.

Train

• You need to set the size of the training data with --train_data_size (from 1 to 507), and the number of training steps with --timesteps.
• You can specify the prepared data directory by --preprocessed_data_dir, and the the directory for saving the model by --save_dir.

python train.py

Evalution

• Specify the path to the evaluation data with --eval_data_fp, and the test timesteps with --timesteps.
• You can also set the sample intervals of the data with --test_sample_interval.
• We provide a 20-step and a 5-step pretrained models for quick evaluation - see the folder pretrained for both models.

python test.py

Citation

@inproceedings{zhao2024physord,
  title = {{PhysORD}: A Neuro-Symbolic Approach for Physics-infused Motion Prediction in Off-road Driving},
  author = {Zhao, Zhipeng and Li, Bowen and Du, Yi and Fu, Taimeng and Wang, Chen},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
  year = {2024},
  pages = {11670--11677}
}