Skip to content

Latest commit

 

History

History
51 lines (36 loc) · 2.32 KB

README.md

File metadata and controls

51 lines (36 loc) · 2.32 KB
Raw

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

Zhipeng Zhao, Bowen Li, Yi Du, Taimeng Fu, and Chen Wang*

This repository provides the official implementation of our paper, "PhysORD: A Neuro-Symbolic Approach for Physics-infused Motion Prediction in Off-road Driving" [PDF].

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

If you find our research helpful for your work, please consider starring this repo and citing our paper.

License

This project is available under the BSD 3-Clause License - see the LICENSE file for details.