The repository contains the code for "GLAS: Global-to-Local Safe Autonomy Synthesis for Multi-Robot Motion Planning with End-to-End Learning", Benjamin Riviere, Wolfgang Hoenig, Yisong Yue, and Soon-Jo Chung, to appear at IEEE Robotics and Automation Letters (RA-L), 2020. A pre-print of the paper is available at https://arxiv.org/pdf/2002.11807.pdf
The data is available in a shared Box folder: https://caltech.box.com/s/doc8hszo55m4x0oyeo1ok2yxh1m2xq3z. The dataset is split into parts with 10000 example instances each. The content of the whole dataset was shuffled before splitting it into parts, i.e., for a smaller dataset it suffices to download only a subset of the datasets.
-
Extract the data into the data subfolder. For example, if you download and extract
training/doubleintegrator/part1.7z, you should have the filedata/training/doubleintegrator/instances/map_8by8_obst06_agents004_ex000007.yaml. -
As verification step, visualize one of the examples:
python3 utils/plot_data.py data/training/doubleintegrator/instances/map_8by8_obst06_agents004_ex000007.yaml data/training/doubleintegrator/central/map_8by8_obst06_agents004_ex000007.npy
Note that the example instance files are regular yaml files, containing a list of start locations, goal locations, and obstacles. The npy files are numpy matrices containing the sampled trajectory for each of the robots (see plot_data.py to understand the matrix layout).
NOTE: This section is still work in progress!
- Install the necessary dependencies with conda (in progress, for now, you have to do install necessary packages manually):
conda env create --file environment.yml conda activate glas_env
- Train and evaluate single model examples (in ~/code):
python examples/run_singleintegrator --il python examples/run_singleintegrator
- Train, evaluate, and visualize batched model examples (in ~/results):
python singleintegrator/exp1.py --train python singleintegrator/exp1.py --sim python singleintegrator/exp1.py --plot