mobmonrob / vigor

Code for the paper "Inferring Versatile Behavior from Demonstrations by Matching Geometric Descriptors" published at CoRL2022

Inferring Versatile Behavior from Demonstrations by Matching Geometric Descriptors

Install

We highly recommend using a conda environment before installing the requirements for this repository.

Requirements can be installed with pip install -r requirements.txt

Code Structure

The Code is written in pure Python. The general setup is as follows:

• "Baselines" contains state-action based environments and different imitation learning baselines for them
• "VIGOR" contains VIGOR and other trajectory based baselines.

  • All experiments use cw2 for distributing and recording purposes.

    • An example without cw2 is given in 'VIGOR/quickstart.py'.

  • The configs directory contains .yaml files that specify individual experiments. , in which individual sections describe the parameters for a given cw2 run. The default.yamlfile contains a short description of each parameter. A run can be called with given parameters using the command python main.py configs/config_file.yaml -e run_name -o

    • To run an experiment on a slurm-based cluster, simply append a -s to the above, i.e., python main.py configs/configs.yaml -e experiment -o -s --nocodecopy
    • We provide an overview of the experiments of the paper in the experiments.sh file. Note that the experiments contain multiple seeds, so they might take a while to run. You can adapt the corresponding .yaml files accordingly, e.g., to run a single seed.

  • Experiments will be automatically recorded and logged in "/VIGOR/experiments". Options for this can be specified via the parameters. Recordings include visualizations for the given tasks, model and reward parameters and graphs for different performance metrics over time. The used config as well as the logged statements are also recorded for convenience

  • Additionally, all metrics can be tracked with WandB

Requirements Python

Tested with Python 3.8.12. on Windows and Ubuntu. A list of all required packages, including version numbers can be found in req.txt and can be installed with