MARVEL

Introduction

• This is the code repository for the paper "MARVEL" that was submitted to RA-L 2024. The demo video has also been uploaded MARVEL_Demo.mp4.
• This is a joint research program of A*STAR and NUS. (© 2024 A*STAR. All rights reserved.)

Dataset

Benchmark experiments are conducted on a series of canonical transportation networks, namely Sioux Falls Network, Friedrichshain, Anaheim, and Winnipeg. Before you try to run the codes, please download the dataset from this link: https://drive.google.com/file/d/12L7PRDGWPF-S6sz-tFMFgfeBZWwQwVaV/view. Then put them into the folder Networks .

MARVEL Algorithm

Dependencies

• Python 3.8
• pandas
• random
• gensim
• NetworkX
• Numpy
• Matplotlib
• PyTorch
• PyTorch Geometric (PyG)
• wandb (optional)

Description

• We integrate our simulation for MARVEL into the training and test processing.
• Embedding.py:Implements the graph embedding process to generate node features for the robot navigation environment.
• Env.py:Constructs the environment, including robot operation and observation processes, defining different topological graph structures
• GAT.py: Implements the policy network based on Graph Attention Networks (GAT), integrating success rate changes during training to enhance adaptability and decision-making capabilities in different environments.
• main.py: Manages the training process, including all parameter settings such as the map used, training iterations, and optimizer configurations.
• Test.py: Tests the final trained MARVEL model and calculates the overall success rate for evaluation.

Usage

cd MARVEL
# training. you can config the parameters in main.py
python main.py
# testing.  use your own model by -m /path/to/your/model.pth .
python Test.py

Baseline Algorithms for Benchmark

Dependencies

• python 3.7+
• pandas
• random
• SciPy
• NetworkX 2.8.4
• Numpy
• Matplotlib
• Gurobipy (a license might be needed)

Description

The parameters of our configuration are shown in the table.

`Benchmark_SOTA`:Five benchmark algorithms OSMIP,DOT,FMA,PQL,DDRL. * `yang.py` corresponds to the OSMIP algorithm. * `prakash.py` corresponds to the DOT algorithm. * `fma.py` corresponds to the FMA algorithm. * `PQL.py` corresponds to the PQL algorithm. * `experiment_benchmark`: Includes experiments of GP3, PQL, CTD, and AC. * `experiment_benchmark2`: Includes experiments of FMA, DOT, OS-MIP, and ILP.

Benchmark_CTP:Three benchmark algorithms DDP, $\pi$($\tau$), RAO*.

• DDP_DL.py corresponds to the Dual Dynamic Programming algorithm.
• pi_tau.py corresponds to the $\pi$($\tau$) algorithm.
• RAO_Star.py corresponds to the RAO* algorithm. All the results of path will be saved in the folder Networks/{map_name}/Benchmark_Record.

Usage

# For SOTA
cd Benchmark_SOTA
python experiment_benchmark.py
python experiment_benchmark2.py
# For CTP
cd Benchmark_CTP
python DDP_DL.py
python pi_tau.py
python RAO_Star.py
# To calculate the success rate
cd Benchmark_Calculate
python calculate_prob.py