Experimental code fragments for "Fail-Safe Adversarial Generative Imitation Learning" paper

This repository contains experimental code fragments for the following scientific paper:

Geiger, Philipp, and Christoph-Nikolas Straehle. "Fail-Safe Adversarial Generative Imitation Learning." Transactions on Machine Learning Research.

Purpose

The code in this repo has the form of partial, preliminary experimental code fragments. The repository does not contain polished, complete, runnable, ready-to-use code. The pupose is to give additional details about the algorithms and experiments in the paper, similar to pseudocode, for those interested. The main reason why only incomplete code could be published is that part of the code is third-party code without permission to be published.

This software is a research prototype, solely developed for and published as part of the aforementioned paper. It will neither be maintained nor monitored in any way. It is experimental code.

Outline

The ril folder contains ...

• data: data-processing-related code
• envs: the simulation envs, in particular for highway
• robustness_checker: the safe set inference module described in the paper, for both, Lipschitz-based and extremality-based safety argumentation
• sac_gail: side parts for the SAC-GAIL algorithm (main parts see below)
• scripts: Hydra configs as well as train scripts for the main methods studied in the paper experiments
• utilities: the name says it all

The il_rl folder contains ...

• sac_gail: the SAC-based GAIL code, and additionally behavior cloning
• further code for integrating the parts

Further remarks

• To train/evaluate on the highD dataset, the dataset itself has to be obtained, and the source code from the highD repository has to be placed in the directory data/highDorig.

License

This repo is open-sourced under the AGPL-3.0 license. See the LICENSE file for details.

For a list of other open source components included in this repo, see the file 3rd-party-licenses.txt.