Skip to content

This repository contains a deep learning-based approach for improving A* search efficiency on grid graphs. By learning instance-dependent heuristic proxies like correction factors and path probability, our method significantly reduces computational effort in obstacle-rich environments.

License

Notifications You must be signed in to change notification settings

AIRI-Institute/TransPath

Repository files navigation

TransPath: Learning Heuristics For Grid-Based Pathfinding via Transformers

This is the code repository for the following paper accepted at AAAI 2023:

Daniil Kirilenko, Anton Andreychuk, Aleksandr Panov, Konstantin Yakovlev, "TransPath: Learning Heuristics For Grid-Based Pathfinding via Transformers", AAAI, 2023.

Visual abstract

Data

Grids

Train, validation, and test maps with pre-computed values mentioned in our paper are available here. One can download and exctract it manually or just run download.py.

DEM

DEM data with paired imagery used in our work are available here. Use get_dem_focals.py to generate gt-focal values.

Pretrained models

Directory ./weights contains parameters for some of the pre-trained models from the paper.

Use train.py to train a model from scratch. Argument --mode defines the type of the model: cf and f are the models for grid-based pathfinding that predict correction factor and focal values respectively, dem is the model for DEM data.

Use eval.py and eval_dem.py to evaluate a model on the test set.

Examples

Check example.ipynb for some examples of predictions and search results of our models. There are a few examples of train and out-of-distribution maps in the directory ./maps.

About

This repository contains a deep learning-based approach for improving A* search efficiency on grid graphs. By learning instance-dependent heuristic proxies like correction factors and path probability, our method significantly reduces computational effort in obstacle-rich environments.

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published