Learning feasible transitions for efficient contact planning

Webpage

Paper

This repository is the code for the paper named "Learning feasible transitions for efficient contact planning".

---

Install

This project uses VS Code devcontainer. To run this workspace, please first install VS Code with the following extensions:

• Dev Containers (ms-vscode-remote.remote-containers)
• Remote Developement (ms-vscode-remote.vscode-remote-extensionpack)
• Docker (ms-azuretools.vscode-docker)

After that is done, clone this repository onto your developement PC.

git clone --recurse-submodules https://github.com/Atarilab/esfal.git

Then, enter the directory downloaded and open it in vscode

cd esfal && code .

After VS Code has started, there should be a prompt to start this workspace as a container. Otherwise, you can also do ctrl + shift + p then select Dev Container: Rebuild and Reopen in Container to start it manually.

The environment is set up automatically with BiConMP installed.

Usage for planning on stepping stones

• main.py runs a minimal example with a stepping stones environment.

• run_mcts.py runs MCTS (with kinematic check only or learned dynamic feasibility) on a simple stepping stones environment.

Usage to train models

• collect_data.py runs simulation and data collection process. Check the arguments. Data should be save in two train and test folders. One should be in project directory to run the script.

cd /home/atari_ws/project/

# Collect train data
python3 collect_data.py --saving_path /home/atari_ws/data/learning_jump_feasibility/train --N 10000 --cores 20 --gait jump

# Collect test data
python3 collect_data.py --saving_path /home/atari_ws/data/learning_jump_feasibility/test --N 1000 --cores 20 --gait jump

• main.py run train the models on the data. Model are saved in learning_jump_feasibility/logs

cd /home/atari_ws/project/learning_jump_feasibility

# Train classifier
python3 main.py --train supervised --cfg ./config/classifierBinary.yml

# Train regressor
python3 main.py --train supervised --cfg ./config/regressorMLP.yml

Run Experiments

• run_experiments.py runs single experiment with specified parameters. To run experiments on various environments and algorithm changes, bash script is used.

cd /home/atari_ws/project/

# run single experiment
python3 run_experiments.py --mode kin --num_remove 9 --pose_noise 0.75 --size_ratio 0.6 --id 0 --gait jump

# run multiple experiments
./multiple_experiments.sh