PyTorch implementation for Graph Neural Networks for Decentralized Multi-Robot Submodular Action Selection

Code accompanying the paper Graph Neural Networks for Decentralized Multi-Robot Submodular Action Selection

from Lifeng Zhou¹, Vishnu D. Sharma², Qingbiao Li³, Amanda Prorok³, Alejandro Ribeiro¹, Vijay Kumar¹ at the (1) University of Pennsylvania, (2) University of Maryland, and (3) University of Cambridge.

The code here is based on the PyTorch Project for Graph Neural Network based MAPF by Qingbiao Li.

Requirements:

easydict>=1.7
matplotlib>=3.1.2
numpy>=1.14.5
Pillow>=5.2.0
scikit-image>=0.14.0
scikit-learn>=0.19.1
scipy>=1.1.0
tensorboardX>=1.2
torch>=1.1.0
torchvision>=0.3.0

How to use this repo:

Generating the data

1. Update the constants.py with the required number of robots, FoV, step size, communication range, etc.
2. Create a folder (say ./robot20_data) to save the data

python -u gnn_setup.py --batch_size 2000 --save_path ./robot20_data/ --mode rect

This command generates 100 files each containing batch_size examples each. First 60 files are used for training, next 20 for validation, and the rest for testing.

You could also use the already generated data by clicking on this link.

Train a new network

1. Create/update the configuration file to reflect the correct path to the training data and parameters (data_root, save_data, save_tb_data, max_epoch)

python main.py configs/coverageTask_20rob_FIX_6FOV_20STEP_10COMM_3layer_2filtertap_GNN2layer_32_128.json --mode train --tgt_feat 20 --rbt_feat 10 --nGraphFilterTaps 2 --num_agents 20 --trained_num_agents 20

The command above trains the model with 2 filter taps i.e. 1-hop.

The pre-trained network can be downloaded by clicking here. The best way to use it is to put the model in the experiment checckpoint directory. Else, you can change the model path in Line 98 of test_model_loader.py.

Test data generation

For testing the reward coverage performance, generate more datailed data with random configuration:

python test_data_generation.py --batch_size 1000 --save_path ./test_data_20robot --mode rect

This script also depends on constants.py and the number of robots can be changed in it to generate data with difference number of robots.

Testing the model

The results can be obtained over this data by running the following command (requires config file path, time ID (from output log) and timeid/log_time_trained):

python test_coverage.py configs/coverageTask_20rob_FIX_6FOV_20STEP_10COMM_3layer_2filtertap_GNN2layer_32_128.json --log_time_trained 1630851013 --nGraphFilterTaps 2 --data_path ./test_data_20robot

Testing model accuracy

Test accuracy indicates if the GNN action matches with the Centralized Greedy action. Since this metric considers a sub-optimal action as incorrect action, it may not truely reflect the coverage performance.

python get_results.py configs/coverageTask_20rob_FIX_6FOV_20STEP_10COMM_3layer_2filtertap_GNN2layer_32_128.json --mode test --num_agents 20  --nGraphFilterTaps 2 --timeid 1661540583

License:

The project of graph mapf is licensed under MIT License - see the LICENSE file for details

Citation:

If you use this paper in an academic work, please cite:

@article{zhou2021graph,
  title={Graph Neural Networks for Decentralized Multi-Robot Submodular Action Selection},
  author={Zhou, Lifeng and Sharma, Vishnu D and Li, Qingbiao and Prorok, Amanda and Ribeiro, Alejandro and Kumar, Vijay},
  journal={arXiv preprint arXiv:2105.08601},
  year={2021}
}